From predictive AI to autonomous food bioprocessing: a critical review of real-time quality monitoring, energy optimization, digital twins, and intelligent process control

1 Introduction

Food bioprocessing is undergoing a transition from conventional process control toward connected and increasingly intelligent manufacturing systems. Traditional operations such as drying, fermentation, thermal processing, extraction, separation, and preservation commonly rely on predetermined operating conditions, mechanistic models, periodic measurements, and operator experience. Although these approaches remain fundamental to food engineering, increasing demands for consistent product quality, food safety, energy efficiency, traceability, and sustainable production have accelerated the adoption of technologies associated with Industry 4.0. Smart sensors, the Internet of Things (IoT), advanced data analytics, machine learning, and digital twins are increasingly being incorporated into food-processing environments to improve process monitoring and decision-making (Hassoun et al., 2023). The need for more responsive processing systems is particularly important because foods are heterogeneous biological materials. Differences in composition, maturity, moisture content, geometry, storage history, and physicochemical properties can alter their response to otherwise identical processing conditions. Conventional control systems typically regulate variables such as temperature, pressure, airflow, or processing time, while important quality characteristics including moisture content, texture, colour, nutrient retention, and microbial status may be determined intermittently or after processing. Consequently, fixed operating conditions may not adequately accommodate variations in raw materials or changes occurring during processing. Digital representations of food processes must therefore consider both equipment behaviour and the changing physical, chemical, and biological characteristics of the product (Krupitzer et al., 2022).

Data-driven modelling offers new possibilities for addressing this complexity. Machine-learning methods can identify nonlinear relationships among raw-material characteristics, process conditions, sensor measurements, energy use, and product outcomes. When combined with real-time sensing, these models can estimate variables that are difficult or time-consuming to measure continuously and can support earlier detection of undesirable process changes. Their demonstrated uses extend from quality prediction and process-state estimation to energy management and operating-condition optimization, primarily at laboratory and pilot scale. However, high predictive accuracy alone does not constitute intelligent process control. A model that predicts moisture content, product quality, energy consumption, or processing time remains a predictive tool unless its outputs influence subsequent process decisions. This distinction is important because monitoring, prediction, automation, intelligent control, and autonomy represent different, empirically distinguishable levels of process capability rather than synonyms for “digitalisation.” Conventional automation generally executes predefined actions, whereas intelligent control uses process information to determine appropriate adjustments. Autonomous bioprocessing requires further integration, allowing a system to assess its current state, anticipate process outcomes, select an intervention, implement that decision, evaluate the resulting response, and adapt subsequent actions with limited human involvement. The technological progression can therefore be viewed as monitoring, prediction, optimization, decision-making, closed-loop control, adaptation, autonomy, and this review deliberately reports, at each stage, whether the cited evidence reflects demonstrated industrial operation, pilot-scale trials, simulation, or conceptual proposal.

Digital twins provide a potential link between these stages. A digital twin connects a physical process with a dynamic virtual representation that can be updated using process data and employed for monitoring, simulation, prediction, optimization, and control (Purlis, 2024). In food processing, combining mechanistic knowledge with operational data is particularly valuable because heat and mass transfer, reaction kinetics, fluid flow, microbial behaviour, and product transformation remain governed by physical and biological principles. Digital food twins can therefore integrate established food-engineering models with data-driven methods to improve representation of complex processing systems (Krupitzer et al., 2022). Nevertheless, industrial implementation remains relatively limited, with persistent challenges involving data availability, model validation, interoperability, computational requirements, cybersecurity, sensor reliability, and integration with existing control infrastructure (Abdurrahman and Ferrari, 2025; Li et al., 2026). A further challenge is the need to balance product quality, energy consumption, and process efficiency. These outcomes are often investigated separately despite their strong interdependence. More intensive processing may reduce processing time while increasing energy demand or accelerating quality deterioration, whereas energy-saving conditions may reduce throughput or compromise required product characteristics. Future intelligent processing systems must therefore address these competing objectives simultaneously rather than optimizing individual outcomes in isolation.

Despite growing research on smart sensing, machine learning, optimization, and digital twins, an important gap remains between predicting food-process behaviour and achieving reliable autonomous process control. Existing studies frequently address individual technological components, while less attention has been given to how sensing, prediction, optimization, digital representations, decision-making, actuation, and feedback must operate as an integrated system. Furthermore, promising laboratory models do not necessarily demonstrate the robustness, transferability, safety, and real-time performance required for industrial closed-loop operation. This narrative review critically examines the transition from predictive approaches toward autonomous food bioprocessing, focusing on real-time quality monitoring, energy optimization, digital twins, and intelligent process control. Unlike broader reviews that catalogue individual digital technologies, the present review (i) organizes current evidence according to the progression from sensing and prediction to optimization, closed-loop control, and autonomy; (ii) explicitly separates demonstrated, pilot-scale, simulation-based, and conceptual technologies at each stage; (iii) situates the proposed autonomy framework against established industrial reference models (ISA-95, RAMI 4.0, SAE J3016) and Technology Readiness Levels; and (iv) considers quality, energy, safety, and economic performance as interconnected engineering objectives rather than isolated outcomes. It further identifies the technological and validation gaps that currently limit industrial implementation and provides a basis for understanding how predictive capabilities can progress toward reliable, adaptive, and resource-efficient autonomous food bioprocessing systems.

2 Review methodology

This narrative review adopted a structured and transparent approach to identify, organize, and critically interpret literature concerning the progression from predictive computational methods to autonomous food bioprocessing. The methodological approach was informed by the Scale for the Assessment of Narrative Review Articles (SANRA), which emphasizes a clearly stated review purpose, description of the literature search, appropriate referencing, and critical presentation of relevant evidence (Baethge et al., 2019). The narrative design was selected, rather than a systematic review or meta-analysis, because the subject encompasses heterogeneous evidence from food engineering, process control, computer science, automation, and digital manufacturing, including experimental studies, modelling investigations, technology demonstrations, reviews, and conceptual frameworks that are not amenable to quantitative pooling (Table 1).

TABLE 1. Search strategy and methodological framework for the narrative review.

2.1 Review design and information sources

Relevant literature was identified from Scopus, Web of Science, ScienceDirect, IEEE Xplore, and Google Scholar, with supplementary backward and forward citation tracking used to identify influential studies that might not have been retrieved through the primary searches. The literature search was conducted from June to August 2026, with the final search updated on 17 August 2026; literature published up to that date was therefore eligible for consideration. A second, targeted search cycle was subsequently conducted to strengthen coverage of process analytical technology and vibrational/optical sensing (near-infrared and Raman spectroscopy), electronic-nose and electronic-tongue systems, convolutional-network-based spectral and image analysis, and digital-twin applications in baking, thermal sterilization, and cold-chain operations, in response to peer review. Particular attention was given to peer-reviewed research addressing food and bioprocess engineering, real-time monitoring, predictive modelling, process optimization, digital twins, and intelligent or autonomous process control. The search prioritized recent literature while retaining earlier publications where they provided important conceptual, methodological, or technological foundations.

2.2 Search strategy and keywords

Search terms were developed around the principal technological and food-processing concepts covered by the review. Combinations of keywords included “food processing,” “food bioprocessing,” “smart food processing,” “artificial intelligence,” “machine learning,” “deep learning,” “convolutional neural network,” “real-time monitoring,” “smart sensors,” “process analytical technology,” “near-infrared spectroscopy,” “electronic nose,” “electronic tongue,” “Internet of Things,” “quality prediction,” “energy optimization,” “process optimization,” “digital twin,” “intelligent control,” “closed-loop control,” “model predictive control,” “reinforcement learning,” “autonomous processing,” “technology readiness level,” and “Industry 4.0.” These terms were combined using Boolean operators (AND OR) and adapted to the syntax of individual databases (for example, in Scopus: TITLE-ABS-KEY ((“food processing” OR “food bioprocessing”) AND (“digital twin” OR “artificial intelligence” OR “machine learning”) AND (“closed-loop” OR “autonomous” OR “intelligent control”))). Process-specific terms such as drying, fermentation, thermal processing, extraction, refrigeration, and non-thermal processing were additionally used to capture applications across major food-processing operations.

2.3 Literature selection and eligibility

Peer-reviewed journal articles, relevant review articles, and high-quality conference publications were considered where they directly addressed the technological scope of the review. Priority was given to studies that reported applications of sensing, predictive modelling, optimization, digital twins, or process-control technologies within food and bioprocessing systems. Studies addressing related manufacturing sectors (for example, additive manufacturing or general process industries) were considered only where they provided digital-twin, control, or autonomy-framework concepts with clear methodological relevance to food bioprocess engineering, and are identified explicitly as cross-domain evidence rather than food-specific demonstrations. Literature was excluded when it lacked a clear relationship to food or bioprocess engineering, focused exclusively on agricultural production without a postharvest or processing component, provided insufficient methodological information, or discussed computational technologies without meaningful application to process monitoring, optimization, or control. Duplicate publications and sources lacking adequate scientific support were also excluded. Each retained study was additionally coded according to its demonstrated technology readiness conceptual/proposal, simulation-only, laboratory-/bench-scale demonstration, pilot-scale demonstration, or reported industrial deployment so that claims in the narrative could be qualified accordingly.

2.4 Evidence extraction and critical narrative synthesis

Relevant information was extracted on the food process investigated, sensing or data-acquisition technology, computational approach, predicted or optimized variables, quality indicators, energy or efficiency outcomes, validation strategy, control architecture, scale of application, demonstrated technology readiness, and reported limitations. Particular attention was given to whether technologies were demonstrated through offline modelling, real-time monitoring, decision support, closed-loop control, or physical process actuation. Evidence was synthesized according to a technological maturity pathway: Monitoring, Prediction, Optimization, Digital Twin, Intelligent Control and Autonomy.

Monitoring included technologies that acquire and interpret process or product information without predicting future states. Prediction covered models estimating quality, energy consumption, process states, or future outcomes. Optimization included approaches used to identify processing conditions that balance quality, energy, productivity, and related objectives. Digital twin referred to dynamically connected representations integrating physical-process information with computational models, and was distinguished explicitly from static digital models or one-way digital shadows (Section 6.1). Intelligent control encompassed systems in which model outputs influence process operation through feedback and actuation. Autonomy represented the most advanced stage, characterized by integrated sensing, prediction, decision-making, actuation, feedback, and adaptive response with limited human intervention. The synthesis emphasized technological maturity, engineering relevance, validation quality, scalability, and the distinction between claimed and demonstrated capabilities. Particular attention was given to identifying gaps between high predictive performance under controlled conditions and evidence of reliable operation in real-time or industrial environments; where a claim in the underlying literature was expressed only as a possibility (for example, using “could,” “may,” or “potentially”), that qualification is preserved in this review rather than presented as an established capability.

2.5 Methodological limitations

The narrative approach enables integration of evidence from diverse technological and engineering disciplines but does not provide the exhaustive study identification, PRISMA-style flow accounting, or quantitative pooling associated with systematic reviews and meta-analyses. Differences in terminology, process conditions, datasets, validation procedures, performance metrics, and technological maturity also limit direct comparisons among studies. Publication bias toward successful applications and inconsistent use of terms such as smart, digital twin, intelligent, and autonomous may further influence interpretation. A further limitation, common to narrative reviews of fast-moving cross-disciplinary fields, is that the absolute number of citable, food-specific primary studies at the digital-twin and closed-loop-control end of the maturity spectrum remains genuinely small; this review reports that scarcity explicitly (Section 6.8, Section 8.8, and Section 9.9) rather than substituting citation volume for demonstrated evidence. These limitations were addressed through transparent reporting of the search approach, explicit technology-readiness coding of cited studies, critical comparison of reported technological capabilities, and explicit differentiation between predictive modelling, optimization, digital representation, closed-loop control, and demonstrated process autonomy.

3 Smart sensing and real-time food quality monitoring

Real-time monitoring represents the foundation of intelligent food bioprocessing because reliable process decisions depend on timely information about both the processing environment and the evolving state of the product. Conventional quality assessment frequently relies on destructive sampling and laboratory analyses performed at discrete intervals, creating a delay between changes occurring within a process and their detection. Smart sensing seeks to reduce this information gap through continuous or near-real-time measurement of physical, chemical, biological, and process variables. Advances in sensor technologies, process analytical technologies (PAT), imaging, spectroscopy, and connected data-acquisition systems are therefore shifting food-quality management from end-point inspection toward continuous process observation and quality assurance (O’Donnell et al., 2014) (Table 2).

TABLE 2. Smart sensing, predictive modelling, and optimization technologies in food bioprocessing.

3.1 Smart food bioprocessing architecture

A smart food-bioprocessing architecture connects the physical process with sensing, communication, data-processing, and decision-support functions (Figure 1). At the physical level, sensors capture variables such as temperature, relative humidity, pressure, airflow, pH, dissolved oxygen, moisture, mass change, conductivity, and energy consumption. Product-oriented sensing can provide additional information on colour, texture, composition, surface defects, volatile compounds, and microbial changes. These measurements are transferred through wired or wireless communication networks to local, edge, or cloud-based computational environments where they can be stored, analysed, and interpreted. The resulting architecture can be represented as physical process → sensing → data acquisition → communication → data processing → quality-state estimation → decision support (Figure 2). This arrangement differs from conventional instrumentation because sensor measurements are not treated as isolated observations; instead, multiple data streams can be synchronized to describe the evolving condition of the process and product. Recent developments in food-sensor research similarly position real-time sensing and adaptive control as central components of Industry 4.0-enabled food systems. Such integration is particularly important in biological processing systems, where changes in raw-material composition and environmental conditions can alter process behaviour. A fermentation system, for example, may require simultaneous interpretation of temperature, pH, dissolved oxygen, substrate concentration, and metabolic activity, whereas drying requires information concerning temperature, humidity, airflow, moisture migration, and energy demand. Smart sensing therefore provides the observational layer upon which subsequent prediction, optimization, digital-twin development, and intelligent control depend.

3.2 Sensors, IoT, and process analytical technologies

Sensor development has expanded the range of variables that can be measured during food processing. Temperature, humidity, pressure, flow, pH, conductivity, gas composition, biosensing, acoustic, and optical sensors can provide information at different stages of processing. Their integration with the Internet of Things (IoT) enables measurements from distributed equipment and processing locations to be collected and communicated continuously. This connectivity is important for processes involving several unit operations because quality changes can be tracked across production rather than at a single processing stage. Process Analytical Technology provides a complementary framework in which process measurements and analytical tools are used to improve understanding and control of manufacturing operations. Its relevance to food processing lies in moving quality assurance away from reliance solely on final-product testing toward obtaining information while processing is occurring. PAT approaches encompass spectroscopy, imaging, chemometrics, sensors, process models, information management, and feedback-control strategies (O’Donnell et al., 2014). Near-infrared (NIR) spectroscopy has become one of the most widely implemented in-line PAT tools, with documented industrial-scale application in dairy processing (composition monitoring of milk and milk powder blending, renneting and coagulation control) and increasing use across grains, meat, and oils (Pu et al., 2020; Grassi and Casiraghi, 2022; Nadimi and Paliwal, 2024). Reported limitations of in-line NIR implementation include calibration-transfer difficulty across instruments and product batches, sensitivity to probe fouling and temperature, and the comparatively small proportion of published NIR studies that report true in-line, as opposed to at-line or offline, deployment (Pu et al., 2020). The combination of PAT with IoT infrastructure strengthens this approach because analytical measurements can become part of a connected monitoring system. However, collecting large volumes of sensor data does not itself create a smart process. Measurements must be sufficiently accurate, representative, synchronized, and interpretable to support reliable assessments of product and process conditions.

3.3 Machine vision, spectroscopy, and hyperspectral imaging

Optical technologies have become particularly important because they permit rapid and often non-destructive assessment of food characteristics. Conventional machine vision primarily captures spatial information associated with appearance, including colour, shape, size, surface defects, and morphological characteristics. It is consequently well suited to automated grading and inspection but provides limited information about internal chemical composition. Spectroscopic techniques complement imaging by exploiting interactions between electromagnetic radiation and food constituents. Near-infrared and related spectroscopic approaches have been applied to estimate moisture, protein, fat, soluble solids, and other compositional attributes. Hyperspectral imaging (HSI) extends these capabilities by combining spectroscopy with imaging, producing spatial and spectral information for individual regions or pixels within a sample. This enables both detection and localization of compositional or quality differences. HSI has consequently been investigated for moisture determination, contaminant detection, defect identification, compositional analysis, authenticity assessment, and food-safety inspection (Huang et al., 2014). Recent developments have further strengthened the role of HSI as a non-destructive quality-assessment technology across fruits, vegetables, meat, grains, and other foods. Its capacity to simultaneously capture spectral and spatial information provides an important advantage where quality is heterogeneous across a product. However, high-dimensional datasets, equipment costs, computational requirements, sensitivity to environmental conditions, and limited transferability of laboratory-developed models continue to restrict widespread online industrial deployment.

3.4 Artificial intelligence for real-time quality prediction

The increasing volume and complexity of sensor and imaging data have created a need for computational methods capable of identifying patterns that are not easily captured by conventional calibration approaches. Convolutional neural networks (CNNs) in particular have been applied directly to spectral and hyperspectral data for food-quality evaluation, learning discriminative features end-to-end rather than relying solely on manually engineered spectral pre-processing, with documented applications spanning grading, contaminant detection, and authenticity assessment across fruit, meat, grain, and dairy matrices (Luo et al., 2024). Electronic-nose and electronic-tongue systems provide a complementary sensing modality, using arrays of partially selective gas or electrochemical sensors combined with pattern-recognition algorithms to characterize volatile-compound and taste profiles associated with freshness, spoilage, and authenticity; reported classification accuracies in food-quality studies commonly fall in the 80%–96% range under the specific conditions tested, though sensor drift, cross-sensitivity, and limited long-term stability remain persistent, reported constraints on industrial deployment (Liu et al., 2019; Sanislav et al., 2025). Increasing integration of electronic-nose sensing with IoT infrastructure is beginning to extend these technologies from laboratory bench instruments toward continuous, networked monitoring, for example, in meat-spoilage detection (Sanislav et al., 2025).

3.5 Nutritional, physicochemical, and microbiological quality prediction

Beyond appearance-based grading, machine-learning and spectroscopic approaches have increasingly been directed at nutritional and safety-relevant attributes, including moisture, protein, fat, and soluble-solids content, texture-related parameters, and indirect indicators of microbial load. These applications are of particular engineering interest because the underlying reference measurements (proximate composition, microbial counts) are typically slow, destructive, or both, making real-time surrogate prediction valuable where it has been adequately validated against reference methods across representative sample sets. Reported model performance varies substantially with product matrix, spectral range, and calibration-set diversity, and few studies report independent, multi-batch or multi-season validation; this remains an important gap rather than a demonstrated capability (Section 8.8).

3.6 Shelf-life and spoilage prediction

Predictive shelf-life modelling combines time–temperature history, compositional information, and, increasingly, sensor or electronic-nose signals associated with spoilage-related volatile compounds to estimate remaining shelf life dynamically rather than relying on a fixed expiry date established under nominal storage conditions. Electronic-nose-based spoilage detection has been demonstrated for meat, seafood, and dairy products, generally in controlled storage trials rather than full commercial cold-chain deployment (Sanislav et al., 2025). The practical value of such systems depends on their ability to generalize across batches, packaging formats, and realistic temperature-abuse scenarios, which remains comparatively under-reported.

3.7 Challenges in real-time monitoring

Despite substantial technological progress, several challenges limit the routine industrial deployment of real-time monitoring systems. Sensor drift, fouling, and calibration transfer across instruments, batches, and product matrices reduce the reliability of continuous measurements and require ongoing maintenance and recalibration (Section 8.7). High-dimensional spectral and imaging data require robust chemometric or deep-learning pipelines that are resilient to environmental variation, and models trained under narrow experimental conditions frequently show reduced accuracy when transferred to new products, seasons, or production lines. Data quality, synchronization across multiple sensing modalities, and integration with existing plant infrastructure present further practical barriers. These challenges are discussed collectively in Sections 8.6–8.8 rather than repeated separately for each sensing technology.

4 AI-driven prediction and optimization of energy and process efficiency

4.1 Energy demand in food bioprocessing

Food processing is energy-intensive across thermal, mechanical, and refrigeration-based operations. Drying, evaporation, sterilization, and pasteurization involve substantial thermal energy input, while refrigeration, freezing, and cold storage require continuous electrical energy to maintain product temperature. Mechanical operations such as size reduction, mixing, and separation additionally contribute to overall energy demand. Improving energy efficiency is therefore an important objective not only for cost reduction but also for reducing the environmental footprint of food manufacturing (Faraldo and Byrne, 2024; Grossmann et al., 2024).

4.2 Machine and deep learning for energy prediction

Machine-learning models can estimate specific energy consumption, thermal efficiency, or energy-related process outcomes from operating conditions, raw-material characteristics, and process measurements. Such models can complement mechanistic energy-balance approaches by learning relationships that are difficult to specify analytically, for example, when heat losses, equipment condition, or material variability influence energy demand in ways not fully captured by design equations. Reported applications concentrate heavily on drying processes, reflecting both the energy intensity of drying and the relative maturity of sensing and modelling approaches in this operation (Wang et al., 2025; Yudhistira and Hsieh, 2024).

4.3 AI-based process performance prediction

Beyond energy consumption specifically, machine-learning models have been applied to predict broader process performance indicators, including processing time, throughput, product yield, and quality retention, from process conditions and raw-material characteristics. These models support process design and optimization by enabling rapid evaluation of alternative operating scenarios without requiring extensive physical experimentation for every combination of conditions, although their reliability outside the sampled operating range remains an important caveat (Section 5.1).

4.4 Multi-objective optimization

Because food-processing objectives frequently conflict, single-objective optimization is often inadequate. Multi-objective optimization approaches identify Pareto-optimal solutions that represent different trade-offs among competing objectives such as quality retention, energy consumption, processing time, and productivity, allowing decision-makers to select an operating point that reflects their priorities rather than a single “optimal” condition. Such approaches have been investigated most extensively in fruit and vegetable drying (Zhong et al., 2025), where quality–energy–time trade-offs are well characterized, with comparatively less evidence in other unit operations.

4.5 Quality–energy–productivity trade-offs

Optimization objectives in food bioprocessing rarely align in a single direction. Increasing processing intensity may reduce processing time and improve throughput while increasing energy demand or accelerating quality deterioration through excessive heat or mechanical stress. Conversely, milder processing conditions may better preserve quality attributes while reducing throughput or requiring longer processing times. Explicit recognition of these trades-offs, rather than optimization of a single objective in isolation, is required to identify operating conditions that are genuinely favourable from an overall process-engineering perspective.

4.6 AI for sustainable and low-carbon processing

AI-based prediction and optimization have been proposed as tools for reducing the carbon footprint of food processing by identifying operating conditions that minimize energy use without compromising product quality (Yudhistira and Hsieh, 2024). The evidentiary basis for such claims, however, is frequently limited to specific-energy-consumption comparisons within a single study rather than full life-cycle assessment; broader carbon and resource accounting is addressed separately in Section 8.10 (Di Bonito et al., 2024).

4.7 From predictive to prescriptive bioprocess engineering

A predictive model that estimates moisture content, energy consumption, or product quality provides decision-relevant information, but the transition to prescriptive engineering—where that information directly informs a recommended or implemented action—requires additional architecture. In many reported systems, a human operator remains the link between prediction and action; this configuration corresponds to Level 3 of the autonomy framework proposed in Section 7.8 and requires operator intervention whenever a recommendation is not communicated directly to the process controller. Closed-loop operation begins when optimized decisions are implemented through actuators and their effects are measured through subsequent sensor feedback. Current research therefore suggests a progression from models used primarily for offline prediction toward systems capable of dynamic optimization and intelligent control. Reviews of smart food drying similarly identify the integration of sensors, computational modelling, optimization, IoT connectivity, and real-time control as an important future direction for energy-efficient processing (Wang et al., 2025; Zhong et al., 2025; Guo et al., 2026). Guo et al. (2026), reviewing intelligent monitoring, prediction, and control technology for food drying, report that most published systems remain at the monitoring–prediction stage, with adaptive, closed-loop control demonstrated in comparatively fewer studies, mirroring the pattern identified throughout this review. The next stage of this progression requires models that retain physical credibility and provide interpretable, reliable predictions under changing conditions. Hybrid physics–AI and explainable modelling approaches consequently form an important bridge between data-driven prediction and trustworthy process control.

5 Hybrid physics–AI and explainable intelligence

5.1 Limitations of purely data-driven AI

Data-driven models have demonstrated considerable capacity to represent nonlinear relationships between processing conditions and food-quality outcomes. Their performance, however, depends strongly on the quality, diversity, and representativeness of the data used for model development. Food-processing datasets are particularly susceptible to variation arising from cultivar, composition, maturity, season, equipment, production scale, storage history, and operating conditions. Models trained within narrowly controlled experimental conditions may consequently lose accuracy when applied to different products, processing lines, or production environments. A further limitation is that purely data-driven models do not inherently obey physical principles. A statistically accurate model may generate physically implausible predictions when applied beyond its training domain. This limitation becomes important in food bioprocessing, where heat and mass transfer, fluid flow, reaction kinetics, phase transitions, and microbial behaviour constrain process responses. Complex models may also provide limited insight into why a particular prediction was generated, reducing their usefulness where decisions influence food safety, quality, energy consumption, or process control. Explainability remains comparatively underused in food engineering despite growing reliance on increasingly complex predictive models (Arrighi et al., 2026).

5.2 Hybrid mechanistic–machine learning models

Hybrid modelling provides a means of combining established process knowledge with the flexibility of machine learning. Mechanistic models describe relationships using conservation laws, transport phenomena, thermodynamics, kinetics, and other established principles, whereas data-driven models learn relationships directly from observations. Their integration can compensate for weaknesses associated with either approach when used independently. In food drying, for example, mechanistic equations can describe heat transfer, moisture diffusion, equilibrium moisture content, and drying kinetics, while machine-learning models can capture nonlinear effects or residual behaviour that is difficult to specify analytically. Similar opportunities exist in fermentation, extraction, thermal processing, refrigeration, and membrane operations. A hybrid model can be represented conceptually as:

Process knowledge + sensor/experimental data + machine learning → physically informed process prediction.

This approach is particularly relevant to digital twins because reliable virtual representations require both responsiveness to real-time data and consistency with the physical behaviour of the underlying process. Hybrid models may also reduce the amount of experimental data required compared with unconstrained data-driven approaches, although their effectiveness depends on the quality of both the mechanistic assumptions and the observational data.

5.3 Physics-informed artificial intelligence

Physics-informed modelling extends hybridization by embedding physical knowledge directly into model development. Physical constraints can be incorporated through model architecture, engineered variables, governing equations, boundary conditions, conservation requirements, or additional terms within the learning objective. The purpose is not simply to improve prediction accuracy but to restrict predictions to physically meaningful behaviour. This principle has considerable relevance to food bioprocessing. Conservation of mass and energy, Fourier heat conduction, Fickian moisture diffusion, fluid-flow equations, reaction kinetics, and thermodynamic relationships can provide constraints for computational models. A physics-informed drying model, for example, could use temperature, humidity, airflow, and product properties while ensuring that predicted moisture and thermal behaviour remain compatible with heat- and mass-transfer principles.

Physics-informed approaches may be particularly useful where experimental datasets are limited or where predictions must extend beyond narrowly sampled conditions. Nevertheless, incorporation of physical equations does not automatically guarantee model validity. Simplified or incorrect physical assumptions can introduce systematic errors, and model validation against independent experimental observations remains essential. Kannapinn et al. (2022) demonstrate one of the more concrete implementations of this principle for food processing: a physics-based, data-driven reduced-order model of a convection oven, trained from a single high-fidelity computational-fluid-dynamics dataset and validated against real-time thermal measurements, providing faster-than-real-time predictions suitable for closed-loop control an explicit, published example of the hybrid principle described above being carried through to a working control-oriented digital twin rather than remaining a conceptual proposal.

5.4 CFD–AI integration

Computational fluid dynamics (CFD) has long been used in food engineering to investigate airflow, temperature distribution, turbulence, heat transfer, and mass transfer within processing equipment. Applications include dryers, ovens, cold rooms, heat exchangers, refrigeration systems, spray dryers, and thermal-processing equipment. CFD provides spatial information that is difficult to obtain experimentally, making it particularly useful for identifying non-uniform temperature, velocity, or moisture conditions. Its principal limitation for real-time applications is computational cost. High-fidelity CFD simulations can require substantial computational resources and execution times that are incompatible with rapid process decisions. Machine-learning surrogate models provide a possible solution. Data generated from validated CFD simulations can be used to train computationally efficient models that approximate relationships between operating conditions and spatial or process responses. A CFD–AI workflow can therefore take the form: Experimental measurements validated CFD model, simulation database, surrogate model, rapid prediction, and optimization/control.

Such integration could allow computational knowledge developed through CFD to support near-real-time optimization and digital-twin applications. The reliability of the surrogate, however, remains dependent on the quality of the CFD model, mesh independence, boundary conditions, material properties, experimental validation, and coverage of the simulated operating domain. A fast surrogate of an inadequately validated CFD model merely reproduces the uncertainty of the original simulation.

5.5 Explainable artificial intelligence

Interpretability becomes increasingly important as computational models move from research tools toward systems that influence processing decisions. Complex models can provide high predictive performance without revealing which variables are responsible for a particular prediction. This creates difficulties in determining whether the model has learned scientifically meaningful process relationships or spurious associations within the training data. Explainable artificial intelligence (XAI) provides methods for examining model behaviour and identifying influential variables. Techniques such as SHAP, partial dependence analysis, feature importance, lime, and Grad-CAM can provide different forms of explanation depending on the model and data type. In food-quality applications, SHAP can indicate which spectral wavelengths or process variables contribute most strongly to a prediction, while Grad-CAM can identify image regions influencing deep-learning classifications. Recent evidence indicates that XAI remains underutilized in food engineering despite its potential contribution to transparency, model verification, and quality-control decision-making (Arrighi et al., 2026).

Explainability also has engineering value beyond increasing user confidence. If a model predicts excessive energy consumption and attributes the prediction primarily to airflow and processing temperature, the explanation can be compared with established engineering knowledge. Agreement may strengthen confidence in the model, while implausible explanations can reveal data leakage, confounding, or inappropriate model behaviour. Explanations should nevertheless not be interpreted as proof of causality. XAI techniques describe how models use available variables rather than necessarily identifying the physical causes of an observed process response. Their outputs may also vary according to the selected explanation method and modelling assumptions.

5.6 Uncertainty quantification and model reliability

A prediction without an associated measure of confidence provides limited support for autonomous decision-making, because a controller acting on an unreliable estimate cannot distinguish a well-supported recommendation from an extrapolation beyond the model’s validated domain. Uncertainty quantification methods including ensemble prediction, Bayesian and quantile-regression approaches, and conformal-prediction techniques provide a statistical basis for expressing this confidence and for triggering conservative action, human review, or fallback control when uncertainty exceeds a defined threshold. Reported use of formal uncertainty quantification specifically within food-bioprocess control studies remains limited relative to its use in general process-control and machine-learning literature, representing a concrete methodological gap rather than a solved problem (Section 9.9).

5.7 Generalizability, transferability, and trustworthy AI

Models validated within a single facility, product, or season provide limited assurance of performance elsewhere. Transferability requires evaluation across independent batches, raw-material sources, equipment, and, ideally, production sites, together with monitoring for model drift as conditions change over time. Trustworthy AI in this context therefore combines accuracy with explainability, calibrated uncertainty, and demonstrated rather than assumed transferability; these four properties are treated as complementary requirements throughout the remainder of this review rather than as substitutable indicators of model quality.

5.8 Comparing data-driven, mechanistic, hybrid, and physics-informed approaches

Because reviewers and practitioners frequently need to select among modelling paradigms for a specific application, Table 3 summarizes the comparative characteristics of purely data-driven, purely mechanistic, hybrid mechanistic–machine-learning, and physics-informed approaches across the criteria most relevant to industrial deployment: data requirements, extrapolation capability, computational cost, interpretability, physical consistency, and reported industrial reliability. No single paradigm dominates across all criteria; the appropriate choice depends on data availability, the criticality of extrapolation beyond sampled conditions, and the computational budget available for real-time operation.

TABLE 3. Comparison of modelling paradigms for autonomous food bioprocessing.

6 Digital twins in food bioprocess engineering

Digital twins represent an important stage in the transition from predictive modelling toward intelligent and autonomous food bioprocessing. Unlike conventional computational models that operate independently of the physical process, a digital twin maintains a dynamic relationship with its physical counterpart through process data, computational models, communication infrastructure, and, at advanced levels, control actions. This relationship allows the virtual representation to evolve as processing conditions change and creates opportunities for real-time monitoring, simulation, prediction, optimization, and decision support. Food processing is particularly suited to this approach because product properties change continuously during operations such as drying, fermentation, heating, cooling, storage, and transportation (Krupitzer et al., 2022; Purlis, 2024; Li et al., 2026; Khan et al., 2022).

6.1 Digital models, digital shadows, and digital twins

Clear terminology is necessary because a computational representation of a food process is not automatically a digital twin. Three levels can be distinguished according to the degree and direction of information exchange between the physical and virtual systems. A digital model is an offline representation in which information exchange with the physical process is largely manual. A digital shadow receives data automatically from the physical system, allowing the virtual representation to follow changes in the process, but information flow is predominantly unidirectional. A digital twin establishes a stronger connection between physical and virtual systems, supporting dynamic updating and, at advanced levels, information exchange that can influence operation of the physical process.

This distinction has important implications for food-engineering research. A CFD simulation of airflow in a dryer, an offline fermentation model, or a dashboard displaying historical production data should not be described as a complete digital twin when it lacks dynamic process connectivity. Similarly, automated sensor transmission to a virtual model may represent a digital shadow rather than a fully integrated twin. Recent food-processing literature increasingly recognizes this distinction and emphasizes connectivity between models, sensors, actuators, and communication systems as an essential characteristic of digital-twin methodology (Abdurrahman and Ferrari, 2025; Purlis, 2024; Li et al., 2026).

The distinction also provides a useful technological maturity pathway: Digital model, Digital shadow, Digital twin, intelligent digital twin and Autonomous cyber-physical system. This progression is directly relevant to autonomous food bioprocessing because increasing connectivity allows digital representations to move from passive simulation toward active participation in process decisions. Only a small number of published food-processing studies satisfy the bidirectional-connectivity criterion required for a true digital twin under this definition; the remainder the majority of the published “digital twin” literature in food processing are more accurately described as digital shadows or digital models, a distinction this review maintains explicitly in Sections 6.8 and 8.8 rather than treating the terms as interchangeable.

6.2 Architecture of food bioprocess digital twins

A food-bioprocess digital twin requires several interacting layers rather than a single computational model (Figure 3). The physical layer contains the food material, processing equipment, sensors, actuators, and environmental conditions. A data-acquisition and communication layer transfers measurements from the physical process to the computational environment. The modelling layer represents the process using mechanistic, statistical, data-driven, or hybrid models. A decision layer interprets model outputs and evaluates alternative operating conditions, while an interface or control layer communicates information to operators or process controllers. A generalized architecture can therefore be represented as (Figure 1).

Feedback closes the loop by returning new sensor measurements to the digital representation after an intervention. This architecture distinguishes a functional digital twin from an isolated simulation because the digital representation changes with the physical process rather than remaining static. Food-specific twins must also represent the product itself. The chemical, physical, and microbiological characteristics of foods can change independently and in response to processing conditions. Krupitzer et al. (2022) therefore proposed a digital food-twin concept in which conventional food-process models are integrated with data-driven approaches to account for variability in raw materials and evolving product characteristics.

6.3 Integration of sensors, IoT, mechanistic models, and AI

The effectiveness of a digital twin depends on the quality of information connecting its physical and virtual components. Sensors provide measurements of process variables such as temperature, humidity, pressure, airflow, pH, dissolved oxygen, moisture, mass, energy consumption, and equipment condition. Spectroscopy, machine vision, hyperspectral imaging, biosensors, and other analytical technologies can extend this information to product characteristics that are difficult to measure continuously. IoT infrastructure provides the communication layer through which distributed measurements can be synchronized and transferred to computational models. Mechanistic models then describe known process relationships involving heat and mass transfer, fluid flow, reaction kinetics, microbial behaviour, or thermodynamics. Data-driven models can complement these relationships where process behaviour is too complex, uncertain, or variable to describe entirely through first-principles equations. Hybrid integration is particularly relevant to food processing because biological variability limits the usefulness of purely deterministic representations, while purely data-driven models may violate physical constraints or fail outside their training domain. Combining mechanistic knowledge with operational data can therefore produce virtual representations that retain physical meaning while adapting to changing processing conditions (Krupitzer et al., 2022).

6.4 Real-time state and quality estimation

Many variables that determine food-process performance cannot be measured continuously using physical sensors. Internal moisture distribution, microbial state, nutrient degradation, remaining processing time, internal temperature distribution, reaction progress, and product shelf life may require destructive analysis or substantial laboratory time. Digital twins can address this limitation through state estimation, in which available sensor measurements are combined with process models to estimate variables that are not directly observable. In drying, measurements of air temperature, relative humidity, airflow, product mass, and surface temperature could be combined with heat- and mass-transfer models to estimate internal moisture distribution and drying progress. In fermentation, measurements of pH, temperature, dissolved oxygen, and gas composition may support estimation of biomass, substrate utilization, or metabolite formation. Thermal-processing and baking twins can similarly estimate internal temperature histories and quality changes that cannot be monitored directly at every location within a food product: Cabeza-Gil et al. (2023) demonstrate this directly with a deep-neural-network-based digital twin developed for a baked crêpe product, and Kannapinn et al. (2022), Kannapinn et al. (2026) with a physics-based reduced-order digital twin of a convection oven used for closed-loop, model-predictive control of thermal food processing. State estimation transforms the digital twin from a process-visualization tool into a continuously updated representation of product and process conditions. Purlis (2024) identifies the ability to generate actionable process metrics that cannot be obtained directly through physical sensors as an important advantage of digital-twin methodology in food processing.

6.5 Virtual simulation and what-if analysis

One of the principal advantages of a digital twin is the capacity to evaluate alternative operating conditions in the virtual environment before modifying the physical process. Such what-if analysis can examine questions such as how changes in temperature, airflow, pressure, residence time, product loading, or cooling rate would influence quality, energy demand, processing time, or safety. This capability can reduce dependence on repeated physical experimentation and provide a safer environment for evaluating operating conditions that could otherwise produce unacceptable products or equipment states. It also allows several alternative interventions to be compared before selecting the most appropriate action. Virtual experimentation has particular value when combined with optimization. Instead of asking only what will happen under the current operating conditions? the digital twin can evaluate what would happen if the operating conditions were changed? This distinction moves digital-twin functionality beyond state monitoring toward decision support and represents an important intermediate step between prediction and intelligent control.

6.6 Digital twins for quality and energy optimization

Food-process optimization requires balancing several objectives that may conflict. Increasing processing intensity can improve throughput or shorten processing time while increasing energy consumption or accelerating quality deterioration. A digital twin provides a computational environment in which these interactions can be evaluated continuously rather than only during initial process design. Real-time process information can update predictions of product quality, energy demand, and processing efficiency, while optimization algorithms explore alternative operating conditions. The selected solution may therefore change as raw-material characteristics or process conditions evolve. Such dynamic optimization differs from conventional offline optimization because the process model is repeatedly updated using information from the physical system. Potential objectives include minimizing specific energy consumption, processing time, carbon intensity, product losses, and quality deterioration, while maximizing nutrient retention, yield, throughput, safety, and process efficiency. Recent reviews identify optimization of resource use, sustainability, product quality, and processing efficiency among the principal potential benefits of digital-twin adoption in food manufacturing (Abdurrahman and Ferrari, 2025; Li et al., 2026). Analogous digital-twin-enabled model-predictive-control architectures have been demonstrated outside the food sector, for example, in additive manufacturing (Kolate et al., 2026) illustrating that the underlying control methodology is technically mature even where food-specific industrial evidence remains comparatively limited.

6.7 Predictive maintenance

Digital twins can also represent the condition of processing equipment rather than focusing exclusively on the food product. Pumps, motors, compressors, fans, refrigeration equipment, conveyors, heat exchangers, and other components generate operational information through vibration, temperature, pressure, electrical current, flow, and performance measurements. Changes in these signals can indicate deterioration before complete equipment failure occurs. Predictive maintenance uses such information to estimate equipment condition and identify abnormal behaviour or emerging faults. Within a digital-twin framework, observed equipment performance can be compared with expected behaviour, allowing maintenance decisions to be based on condition rather than solely on fixed schedules. The relevance to food processing extends beyond maintenance cost. Equipment deterioration can affect temperature control, airflow distribution, refrigeration capacity, processing uniformity, hygiene, and product quality. Predictive maintenance can therefore contribute indirectly to food safety and process consistency. However, much of the digital-twin literature in food manufacturing remains focused on monitoring and process optimization, and robust evidence of industrial-scale predictive-maintenance twins specifically developed for food-processing equipment remains comparatively limited.

6.8 Current limitations of food bioprocess digital twins

Despite considerable conceptual promise, digital-twin implementation in food processing remains at an early stage compared with more mature industrial sectors. A recent systematic review identifies documented applications concentrated in three scenarios thermal processing and sterilization, fermentation, and supply-chain/cold-chain integration using a four-layer perception–model–simulation–application architecture, but finds that routine industrial deployment remains limited relative to laboratory and pilot-scale demonstrations across all three (Li et al., 2026; Abdurrahman and Ferrari, 2025; Purlis, 2024). Within thermal processing specifically, published examples now extend from cooking and baking (Cabeza-Gil et al., 2023; Kannapinn et al., 2022; 2026) to sterilization and cold-chain/storage applications (Monteiro and Barata, 2025), addressing what was previously a comparatively narrow evidence base concentrated on drying and refrigerated transport alone.

Several barriers explain the continued gap between conceptual promise and industrial deployment. Data quality and availability remain fundamental because incomplete, noisy, poorly synchronized, or non-representative sensor data can reduce the accuracy of the virtual representation. Raw-material variability further complicates modelling because biological products can change between cultivars, batches, seasons, suppliers, and storage conditions. A twin developed under one set of conditions may therefore require recalibration before being transferred to another process.

Model fidelity and computational speed present another trade-off. Detailed CFD, heat-transfer, mass-transfer, or kinetic models may accurately represent process behaviour but be too computationally expensive for real-time operation. Reduced-order models and machine-learning surrogates such as the non-intrusive reduced-order model reported by Kannapinn et al. (2022), trained from a single simulation dataset to achieve faster-than-real-time inference can improve computational speed, but simplification may introduce additional uncertainty. The appropriate digital twin must therefore balance physical detail against the time constraints of process decision-making. Interoperability also remains problematic. Sensors, processing equipment, industrial control systems, databases, and modelling platforms may use different communication standards and data structures. Legacy equipment may provide limited connectivity, making digital-twin implementation technically difficult or expensive. Cybersecurity becomes increasingly important as physical food-processing equipment is connected to digital networks and remote computational environments. Validation represents an equally important challenge. Demonstrating agreement between a virtual model and experimental observations under controlled conditions does not establish that the twin will remain reliable during prolonged industrial operation. Digital twins intended for process control require continued evaluation of model drift, sensor degradation, uncertainty, abnormal operating conditions, and failure behaviour.

A final concern is terminological inflation. The increasing popularity of digital twins creates a risk that offline simulations, dashboards, digital models, and one-way monitoring systems are labelled as digital twins without demonstrating the required degree of physical–digital integration (Section 6.1). Clear reporting of data flow, updating frequency, model structure, validation, decision functionality, and actuator connectivity is therefore essential when evaluating technological maturity. Digital twins should consequently be regarded as an enabling architecture rather than evidence of autonomy in themselves. A twin may monitor, simulate, and optimize a process without directly controlling it. Progress toward autonomous food bioprocessing requires the digital representation to interact with a validated control system capable of converting predictions and optimization outputs into safe physical actions. This transition establishes the basis for intelligent process control and closed-loop bioprocessing, where the virtual and physical systems become components of an integrated decision-and-action cycle.

7 Intelligent process control and the transition to autonomous bioprocessing

7.1 From prediction to closed-loop decision-making

The transition from predictive modelling to intelligent process control represents a critical step toward autonomous food bioprocessing. Predictive systems estimate future process states or product characteristics but do not necessarily modify the process. Closed-loop control extends prediction into action by continuously comparing measured or estimated process states with desired targets and using the resulting information to adjust operating conditions. This distinction is particularly important in food processing, where raw-material variability and nonlinear process dynamics can make fixed operating schedules inadequate. Recent developments in intelligent food drying demonstrate a similar progression from real-time monitoring and prediction toward adaptive control strategies capable of modifying process conditions according to changes in material state (Guo et al., 2026). A closed-loop architecture therefore requires more than an accurate predictive model. Sensor information must be acquired and interpreted, desired outcomes must be defined, an appropriate control action must be calculated, and actuators must implement that decision. The effect of the intervention is subsequently measured and returned to the controller. The process can be represented as: Process state → Sensing → State estimation/prediction → Decision → Controller → Actuator → Process response → Feedback. This feedback mechanism differentiates intelligent control from offline optimization. The objective is not merely to determine theoretically favourable processing conditions but to adjust those conditions as the food material and processing environment change.

7.2 Model predictive and adaptive control

Model predictive control (MPC) is particularly relevant to food and biological processes because it uses a process model to anticipate future behaviour over a defined prediction horizon. The controller evaluates possible future control actions, selects those that best satisfy specified objectives and constraints, implements an initial action, and repeats the optimization when new process measurements become available. MPC can therefore accommodate interactions among several process variables while explicitly considering operational constraints. This capability is valuable where quality, safety, energy consumption, and productivity must be considered simultaneously. A drying controller, for example, could adjust temperature and airflow according to predicted moisture evolution while maintaining upper temperature limits and minimizing energy demand. In fermentation, MPC can potentially regulate feeding, temperature, dissolved oxygen, and related variables according to predicted biological responses. Applications in broader bioprocessing demonstrate the capacity of MPC to combine real-time monitoring with optimization while accounting for process constraints (Eslami, 2024). Kannapinn et al. (2026) provide a food-specific example, reporting a digital-twin-based MPC framework for a convection oven that uses neural-ordinary-differential-equation reduced-order models to achieve faster-than-real-time prediction for closed-loop temperature control one of the few published food-processing studies to combine an explicit digital twin with a working MPC loop rather than either component in isolation. Comparable digital-twin–MPC integration has also been demonstrated in adjacent manufacturing domains such as additive manufacturing (Kolate et al., 2026), reinforcing that the control methodology itself is not the limiting factor for food-sector adoption; food-specific challenges instead concentrate in sensing of biological state variables and validation across raw-material variability. Adaptive control addresses another important characteristic of food bioprocessing: process behaviour may change with time. Raw-material properties, microbial activity, equipment condition, environmental factors, and product characteristics can vary during production. Adaptive controllers modify model parameters or control policies as new information becomes available, potentially maintaining performance when the underlying process differs from its original calibration.

7.3 Fuzzy and AI-based controllers

Precise mathematical representation of food and biological processes is not always possible because of nonlinearities, uncertainties, incomplete measurements, and qualitative process knowledge. Fuzzy logic provides an alternative in which control decisions can be expressed through linguistic relationships such as if product moisture remains high and drying rate is decreasing, then increase airflow moderately. Such rules can incorporate operator knowledge without requiring an exact mathematical model.

Fuzzy control has been investigated extensively in drying and food and beverage processing because these systems often contain nonlinear, time-varying, and uncertain relationships. Fuzzy systems have been applied to modelling, classification, optimization, and process control, while hybrid architectures combine fuzzy reasoning with neural networks or conventional controllers (Birle et al., 2013; Hosseinpour and Martynenko, 2022). Neural-network and neuro-fuzzy controllers provide additional flexibility because they can approximate complex process relationships from data. Their use, however, introduces concerns regarding interpretability, generalization, and behaviour outside the training domain. These issues become more consequential when computational outputs directly modify physical processing conditions.

7.4 Reinforcement learning for process control

Reinforcement learning (RL) offers a different approach to process control. Rather than learning only the relationship between inputs and predefined outputs, an RL agent learns a control policy through interactions with an environment. Actions generate consequences, and a reward function evaluates whether those consequences advance the desired objective. Repeated interaction allows the controller to learn actions that maximize cumulative reward. For food bioprocessing, a reward function could simultaneously penalize excessive energy consumption, quality deterioration, unsafe operating conditions, and long processing times while rewarding productivity and achievement of desired product specifications. Published applications of RL specifically to food-bioprocess control remain concentrated in simulated or laboratory-scale environments rather than industrial deployment; reviews of RL in process industries more broadly identify applications spanning low-level control, optimization, scheduling, and fault management, but caution that direct trial-and-error learning on physical equipment raises safety and economic concerns that are not yet resolved for food-safety-critical operations (Nian et al., 2020). An RL agent cannot be permitted to explore unsafe temperatures, pressures, microbial conditions, or other operating states simply to determine their consequences. Training within validated simulations or digital twins provides a safer alternative, followed by carefully constrained transfer to the physical process. Hybrid approaches combining process models and RL may also improve data efficiency where experimental observations are expensive or limited.

7.5 Sensor–AI–controller–actuator feedback loops

Autonomous operation requires integration of sensing, computation, decision-making, and physical actuation. A useful closed-loop architecture can be expressed as (Figure 3).

Sensors provide information about the product, equipment, and processing environment. Computational models convert these measurements into estimates of current and future process states. Optimization or control algorithms determine an appropriate intervention, while actuators modify variables such as temperature, airflow, pressure, feed rate, mixing speed, cooling intensity, or processing time. Subsequent measurements determine whether the intervention produced the intended response. Recent food-drying literature emphasizes multimodal sensing, data fusion, predictive modelling, and intelligent control as complementary components of closed-loop processing rather than independent technologies (Guo et al., 2026). Such integration is essential because a sophisticated model without reliable sensors cannot determine the actual process state, while accurate predictions without actuators cannot alter process behaviour. Safety mechanisms must remain independent of optimization objectives. Hard constraints, emergency shutdown systems, equipment limits, and food-safety requirements should not depend exclusively on learned control policies. Autonomous decision-making therefore requires a supervisory safety layer capable of overriding computational recommendations when operating limits are approached; Section 9.10 develops the regulatory and technical requirements of this safety layer in more detail, in direct response to peer review.

7.6 Human-in-the-loop versus autonomous control

Increasing automation does not necessarily require immediate removal of human decision-making. A human-in-the-loop architecture allows computational systems to generate predictions, detect abnormalities, evaluate alternatives, or recommend control actions while an operator retains responsibility for approval or intervention. This approach may be appropriate where uncertainty is high, safety consequences are substantial, or regulatory requirements demand human oversight. More advanced systems can move toward human-on-the-loop supervision, where routine control actions are implemented automatically while operators monitor system behaviour and intervene when necessary. Fully autonomous operation represents a further stage in which routine sensing, prediction, optimization, actuation, and adaptation occur without continuous human approval. The appropriate degree of autonomy should therefore depend on process risk, model reliability, sensor performance, regulatory requirements, and the consequences of incorrect decisions. Automated food-processing adoption remains comparatively limited, and future progress will require attention to human–machine interaction as well as technical performance (Zhang et al., 2025).

7.7 Defining autonomous food bioprocessing

For the purposes of this review, autonomous food bioprocessing is defined as the capability of an integrated processing system to sense its physical and product state, interpret process information, predict relevant outcomes, select and implement appropriate control actions, evaluate their effects, and adapt subsequent decisions within defined safety and quality constraints with limited routine human intervention. This definition distinguishes autonomy from automation. An automated dryer that follows a predetermined temperature schedule may require little operator involvement but does not independently determine whether the schedule remains appropriate. Similarly, a machine-learning model that predicts moisture content does not constitute autonomy if its predictions are displayed only to an operator. Autonomous functionality emerges when perception, prediction, decision-making, physical intervention, feedback, and adaptation are integrated. The distinction is particularly important because current food-bioprocessing research remains substantially more mature in prediction and optimization than in demonstrated autonomous industrial control.

7.8 Levels of food bioprocess autonomy

To provide a consistent basis for evaluating technological maturity, this review proposes a six-level Food Bioprocess Autonomy Framework (FBAF). The framework is conceptually informed by autonomy-level approaches used in industrial process engineering, where six levels ranging from no autonomy to full autonomy have previously been proposed for the “autonomous industrial plant” (Gamer et al., 2020). The present framework adapts this logic specifically to food bioprocessing and distinguishes systems according to their capabilities in sensing, prediction, decision-making, actuation, feedback, and learning (Figure 4). Its specific novelty relative to Gamer et al. (2020) and the basis on which it can be judged independently of that source is developed in Section 7.9, together with an explicit comparison against the ISA-95 automation hierarchy, the RAMI 4.0 reference architecture, and the SAE J3016 vehicle-autonomy scale, and with measurable, per-level classification criteria that were not previously specified.

Level 0 — Manual Processing.

Process observation, interpretation, decision-making, and adjustment are performed predominantly by human operators. Instrumentation may be minimal, and product quality is commonly evaluated through periodic or end-point measurements. Classification criterion: no continuous digital sensing of process variables; no automated set-point control.

Level 1 — Instrumented Processing.

Sensors continuously or periodically measure important process variables, but human operators interpret the information and determine control actions. Digital monitoring improves process visibility without independently modifying processing conditions. Classification criterion: ≥1 continuously logged process variable; control actions remain manual.

Level 2 — Automated Processing.

Controllers automatically maintain predetermined set points or execute programmed processing schedules. Feedback control may be present, but the system does not independently predict product outcomes or determine new operating objectives. Classification criterion: closed-loop PID/on-off control to fixed set points; no predictive model informs the set point itself.

Level 3 — AI-Assisted Processing.

Predictive models, soft sensors, machine vision, or optimization algorithms estimate product quality, process states, energy consumption, or desirable operating conditions. Computational outputs support human decisions, but substantial responsibility for approving or implementing interventions remains with the operator. Classification criterion: a validated predictive or optimization model exists and is used operationally, but its output requires human approval before implementation.

Level 4 — AI-Optimized Closed-Loop Processing.

Real-time sensing, predictive modelling, optimization, and actuation are integrated. The system can modify processing conditions automatically according to changing process or product states while operating within predefined constraints. Human supervision remains available for exceptional conditions, model uncertainty, or safety-related interventions. Classification criterion: model output is implemented by the controller without routine human approval; a documented safety envelope constrains permissible actions. Kannapinn et al.’s (2026) digital-twin MPC oven-control study is, on this criterion, an illustrative Level 4 demonstration at laboratory scale.

Level 5 — Self-Learning Autonomous Bioprocessing.

The system integrates sensing, state estimation, prediction, optimization, control, actuation, feedback, uncertainty assessment, and adaptive learning. It can modify its control strategy in response to changing raw materials, equipment behaviour, or processing conditions while respecting validated safety, quality, and operational constraints. Human involvement primarily concerns strategic objectives, governance, maintenance, and exceptional situations rather than routine process control. Classification criterion: the control policy itself updates from operational experience (online learning, retraining, or policy adaptation) under a validated governance process, not merely the process set point. No food-bioprocessing system reviewed here meets this criterion at industrial scale; the closest published analogues are found in adjacent domains, such as self-driving laboratories for materials discovery (MacLeod et al., 2020), which satisfy the criterion for a different application. The progression can therefore be summarized as: Level 0: Human operates → Level 1: System senses → Level 2: System executes → Level 3: System predicts and advises → Level 4: System predicts, decides, and controls → Level 5: System learns, adapts, and controls autonomously.

The proposed framework should not be interpreted as evidence that Level 5 autonomous food-processing systems are already established industrial practice. Rather, it provides a conceptual scale for distinguishing current technological capabilities and preventing systems with fundamentally different degrees of autonomy from being grouped under the same terminology. Existing evidence suggests that many applications remain concentrated around instrumentation, automation, predictive modelling, and decision support, while fully integrated closed-loop and self-learning systems remain considerably less mature. Progress toward higher autonomy levels will therefore depend not only on more sophisticated algorithms but also on reliable sensing, physical-process understanding, uncertainty quantification, external validation, fail-safe control, cybersecurity, interoperable equipment, and regulatory acceptance. Autonomy should consequently be evaluated according to demonstrated closed-loop capability and validated adaptive behaviour, rather than the presence of artificial intelligence alone.

7.9 FBAF compared with existing industrial autonomy and maturity frameworks

Several established reference architectures already describe automation and digitalization maturity in general manufacturing: the ISA-95 (IEC 62264) functional hierarchy, which stratifies enterprise, planning, supervisory, control, and field levels of an automation pyramid; the Reference Architectural Model Industry 4.0 (RAMI 4.0), a three-dimensional model combining hierarchy levels, life-cycle stages, and functional layers; and the SAE J3016 taxonomy of six driving-automation levels (0–5), which Gamer et al. (2020) already drew upon conceptually in proposing autonomy levels for the process industry. Reviewer comment correctly identifies that the FBAF, as originally presented, did not make its relationship to these frameworks explicit, nor did it specify testable, food-domain-specific classification criteria beyond a qualitative description of each level. Table 4 addresses both points directly.

TABLE 4. FBAF compared with ISA-95, RAMI 4.0, and SAE J3016-Derived industrial autonomy scales.

TRL, bands are illustrative classifications assigned by the present authors using the standard nine-point Technology Readiness Level scale (TRL, 1: basic principles observed; TRL, 9: proven in operational environment), based on the demonstrated scope of the cited study rather than vendor claims. They are offered as a starting point for author- or reader-verified reassessment rather than as an authoritative rating, since no cited source itself reports a TRL.

The FBAF’s contribution relative to these existing frameworks is threefold. First, ISA-95 and RAMI 4.0 describe where functions sit within a plant’s information hierarchy and lifecycle but do not, by themselves, specify how autonomous the decision-making performed at a given level is; a Level-3 ISA-95 supervisory system can correspond to anywhere from FBAF Level 1 to Level 4 depending on whether its logic is fixed, advisory, or closed-loop. Second, SAE J3016 and Gamer et al.’s (2020) adaptation describe autonomy generically for, respectively, road vehicles and generic industrial plants; neither incorporates food-specific classification anchors biological raw-material variability, food-safety critical control points, and destructive/indirect quality measurement that determine what “sensing” and “prediction” concretely mean in this domain, and that this review’s per-level criteria make explicit. Third, the FBAF is presented together with an illustrative classification of specific, citable published systems (Table 4, final column), which none of the frameworks compared here provide for the food-bioprocessing domain; this satisfies the operational-testability requirement previously absent from the framework.

This mapping is deliberately approximate rather than a one-to-one crosswalk: ISA-95 and RAMI 4.0 describe organizational and informational hierarchy, whereas the FBAF describes decision-making autonomy, and a system can occupy a high ISA-95/RAMI 4.0 tier while remaining at a low FBAF level if its logic is fixed rather than adaptive. Applying Table 4 to the applications summarized in Table 5 (Section 8) indicates that the great majority of currently published food-bioprocessing systems occupy FBAF Levels 1–3, with a small but growing number of Level 4 demonstrations concentrated in thermal processing and drying, and no verified food-specific Level 5 system identified in the literature reviewed.

TABLE 5. Applications, technological maturity, benefits, and challenges of intelligent food bioprocessing.

Maturity/FBAF, ratings are the authors’ classification of the cited evidence base against the criteria in Section 7.8/Table 4, not a claim by the cited authors themselves.

8 Applications, challenges, and sustainability implications

The practical value of intelligent food bioprocessing depends on its ability to improve product quality, process efficiency, resource use, and operational reliability under realistic production conditions. Applications now extend across drying, fermentation, thermal processing, extraction, refrigeration, and food supply chains, although technological maturity varies considerably among operations, as summarized against the FBAF (Table 4) and against demonstrated readiness in Table 5. Digital-twin evidence, for example, remains concentrated in selected food categories, with substantially fewer documented industrial applications than academic demonstrations (Abdurrahman and Ferrari, 2025; Li et al., 2026). The transition toward autonomous processing therefore requires consideration not only of technical performance but also of data quality, sensor reliability, scalability, cybersecurity, food safety, and economic and environmental consequences (Table 5).

8.1 Smart drying and dehydration

Drying is among the most promising applications for intelligent food bioprocessing because it involves strongly coupled heat and mass transfer and substantial trade-offs among moisture removal, product quality, processing time, and energy consumption. Product characteristics also change continuously during drying, making fixed temperature–time schedules potentially inefficient when raw-material properties vary.

Smart drying combines sensors with computational models to estimate moisture content, drying rate, temperature distribution, quality changes, and energy requirements. Machine vision, hyperspectral imaging, electronic noses, temperature and humidity sensors, and mass measurements can provide complementary information on the evolving product state. Recent reviews indicate increasing integration of these sensing technologies with predictive models for online monitoring and optimization of food drying, alongside a documented evolution of control strategies from PID and fuzzy logic toward model-predictive and adaptive control (Wang et al., 2025; Guo et al., 2026). The principal opportunity lies in progressing from end-point prediction toward dynamic control. Processing temperature, airflow, humidity, or energy input could be modified according to estimated moisture content and quality rather than following an invariant schedule, and closed-loop drying-control demonstrations are now reported at pilot scale (Guo et al., 2026). However, reliable online moisture and quality estimation, sensor robustness, and validation across different products and dryer configurations remain important requirements.

8.2 Intelligent fermentation

Fermentation presents a different control problem because microbial and enzymatic activities create dynamic biological systems whose behaviour can change substantially during processing. Temperature, pH, dissolved oxygen, agitation, substrate concentration, gas composition, biomass, and metabolite concentrations may interact nonlinearly. Conventional control based on a limited number of variables may therefore provide an incomplete description of fermentation state. Intelligent fermentation systems can combine online sensors, soft sensors, kinetic models, and data-driven prediction to estimate variables that are difficult to measure continuously. Predictive models can support identification of fermentation endpoints, abnormal behaviour, substrate requirements, and changing metabolic states. Digital twins provide a further opportunity to integrate biological models with real-time measurements and evaluate alternative feeding, temperature, aeration, or agitation strategies before they are implemented (Li et al., 2026). The challenge lies in biological variability. Microbial behaviour may change with strain, substrate composition, inoculum characteristics, contamination, and environmental conditions. Models developed for one fermentation system may therefore have limited transferability. Reliable autonomous fermentation requires models that can accommodate biological uncertainty rather than assuming stable input–output relationships.

8.3 Thermal and non-thermal processing

Thermal processing requires careful balancing of microbial safety against undesirable quality changes. Pasteurization, sterilization, cooking, baking, blanching, and related operations must deliver sufficient treatment to achieve safety objectives while minimizing nutrient degradation, texture changes, colour deterioration, and unnecessary energy consumption. Dynamic models can integrate temperature histories with kinetic relationships to estimate both microbial inactivation and quality loss. Digital twins have been investigated in thermal-processing, baking, and cooking applications, including reduced-order and deep-learning-based twins developed specifically for oven and convective heating processes, with the potential to support simulation, optimization, and improved process control (Abdurrahman and Ferrari, 2025; Kannapinn et al., 2022; 2026; Cabeza-Gil et al., 2023; Li et al., 2026). Real-time estimation of internal product temperature and accumulated lethality is particularly relevant where direct measurement is difficult. Non-thermal technologies such as high-pressure processing, pulsed electric fields, ultrasound, cold plasma, and related emerging processes introduce additional optimization challenges. Their effects depend on interacting treatment intensity, exposure time, product composition, geometry, and equipment characteristics. Predictive modelling can assist in identifying treatment combinations that achieve safety and quality requirements without unnecessary processing intensity. Industrial adoption nevertheless requires robust validation because microbial safety cannot depend solely on predictions from models developed under limited experimental conditions.

8.4 Extraction and separation processes

Extraction and separation are central to the production of oils, proteins, flavours, pigments, bioactive compounds, and functional ingredients. Conventional optimization often involves extensive experimentation because extraction yield and selectivity depend on temperature, solvent composition, pressure, particle size, solid-to-liquid ratio, residence time, and other interacting variables. Predictive models can reduce the experimental search space by estimating extraction yield, composition, or product functionality under different processing conditions. Multi-objective optimization can further balance recovery yield against energy demand, solvent consumption, processing time, and compound degradation. Computational methods are also increasingly being investigated for identifying and screening food-derived bioactive compounds, potentially reducing the time and cost associated with conventional screening approaches. For industrial application, however, optimization should consider downstream separation and solvent recovery rather than extraction yield alone. Conditions producing maximum laboratory-scale recovery may be unattractive when energy consumption, solvent requirements, purification difficulty, environmental impact, or scale-up constraints are included.

8.5 Refrigeration and cold-chain applications

Refrigeration is essential for maintaining food quality and safety but represents a major source of energy demand across food processing, storage, retail, and transportation. Temperature fluctuations during distribution can accelerate microbial growth, biochemical deterioration, moisture loss, and shelf-life reduction. Fixed shelf-life estimates may consequently fail to represent the actual condition of products exposed to variable temperature histories. Connected temperature sensors, IoT systems, predictive shelf-life models, and digital twins provide opportunities for condition-based cold-chain management. Product histories can be used to estimate remaining shelf life and identify deviations before substantial deterioration occurs. Digital-twin approaches have also been proposed for improving traceability, resource efficiency, resilience, and optimization across food supply chains and cold chains specifically (Monteiro and Barata, 2025; Li et al., 2026). Energy optimization is equally important. Refrigeration technologies must maintain appropriate product temperatures while minimizing electricity consumption and environmental impact. Recent assessments emphasize the importance of more efficient refrigeration technologies, improved system design, thermal storage, alternative refrigerants, and intelligent energy management for sustainable cold chains (Faraldo and Byrne, 2024).

8.6 Data availability and quality challenges

Reliable intelligent processing depends fundamentally on reliable data. Food-processing datasets are often relatively small, fragmented, proprietary, or collected under narrowly defined experimental conditions. Differences in raw materials, equipment, sensor configurations, sampling rates, processing protocols, and laboratory reference methods further complicate aggregation across studies. Data quantity alone is not sufficient. Missing observations, sensor noise, incorrect labels, temporal misalignment, class imbalance, and inadequate reference measurements can introduce systematic errors into predictive models. Data leakage between training and evaluation sets can also produce apparently strong performance that does not persist during deployment. The absence of standardized, representative datasets limits comparison among computational methods and contributes to reproducibility problems. Greater attention is therefore required to data provenance, metadata, calibration procedures, preprocessing, independent test datasets, and transparent reporting.

8.7 Sensor reliability and interoperability

Sensors constitute the interface between the physical food process and its computational representation. Their reliability directly affects every subsequent stage of prediction, optimization, and control. Fouling, drift, vibration, moisture, cleaning chemicals, high temperatures, variable lighting, electromagnetic interference, and mechanical damage can degrade sensor performance under industrial conditions a constraint already introduced in Section 3.7 and applied here specifically to interoperability. Redundancy and sensor fusion can improve robustness by allowing several measurements to contribute to process-state estimation. However, additional sensors increase system complexity, data volume, calibration requirements, and cost. Interoperability presents a further, distinct barrier: processing equipment, laboratory systems, sensors, industrial controllers, databases, and digital-twin platforms may use incompatible communication protocols and data structures, an issue that ISA-95 and RAMI 4.0 were designed to address at the enterprise level but that legacy food-processing equipment frequently does not implement. Recent digital-twin reviews identify interoperability and data integration among the principal challenges to wider implementation in food processing and food supply chains (Abdurrahman and Ferrari, 2025; Monteiro and Barata, 2025; Li et al., 2026).

8.8 Model validation, standardized metrics, and industrial scalability

Strong performance under laboratory conditions does not demonstrate industrial readiness. Many computational models are developed using limited datasets collected from a single food material, processing system, or operating range. Industrial environments introduce greater variability in raw materials, throughput, environmental conditions, equipment behaviour, and sensor performance. External validation is therefore essential. Models intended for industrial use should be evaluated across independent batches, seasons, products, processing equipment, and, where possible, manufacturing sites. Prospective validation under actual operating conditions provides stronger evidence than retrospective evaluation of previously collected datasets.

Reported model performance in the food-bioprocessing literature is overwhelmingly expressed through in-sample or single-split accuracy metrics R ², RMSE, or classification accuracy evaluated on data drawn from the same product, batch range, and instrument as the training set. This narrow reporting convention limits the conclusions that can be drawn about deployment readiness, because none of these metrics characterize behaviour under distribution shift, none quantify predictive confidence, and none capture computational feasibility under real-time constraints. Reliable evaluation of a model intended for autonomous or closed-loop use should additionally report: (i) robustness, evaluated as performance degradation under batch-to-batch, seasonal, raw-material, and equipment variability, ideally via leave-one-batch-out or prospective validation rather than random splits; (ii) calibration, the agreement between predicted uncertainty and observed error frequency (Section 5.6); (iii) computational latency, measured against the time constant of the physical process being controlled, since a model too slow for the relevant control interval cannot support real-time actuation regardless of its offline accuracy; and (iv) model degradation over time, tracked through scheduled or continuous monitoring of live-deployment error relative to the validation baseline, to detect drift before it affects product quality or safety. This four-part reporting standard is proposed here as a benchmark against which future food-bioprocessing AI studies, and this review’s own cited evidence base, can be assessed; very few of the primary studies cited in this review report all four elements, which is itself informative about the field’s current state of validation practice (Section 9.8).

Computational speed also becomes important when predictions influence real-time decisions. Detailed mechanistic or CFD models may provide valuable physical information but operate too slowly for online control. Reduced-order models and computational surrogates can improve execution speed, although simplification must not undermine reliability. Evidence from digital-twin research demonstrates that industrial adoption remains substantially less extensive than academic development, highlighting the persistent gap between proof-of-concept studies and scalable deployment (Abdurrahman and Ferrari, 2025; Li et al., 2026).

8.9 Cybersecurity, regulation, and food safety

Greater connectivity creates new forms of operational risk. Networked sensors, cloud platforms, digital twins, and remotely accessible controllers can expose food-processing systems to unauthorized access, manipulation, data corruption, or disruption. Cybersecurity is therefore not simply an information-technology concern when digital systems can influence physical processing conditions. Data integrity is especially important. Incorrect or manipulated sensor information could result in inappropriate processing decisions affecting temperature, processing duration, formulation, refrigeration, or other safety-critical variables. Recent digital-twin literature identifies cybersecurity, data integration, and standardization among the principal barriers to more extensive food-sector deployment (Abdurrahman and Ferrari, 2025; Li et al., 2026). Autonomous systems must also operate within established food-safety requirements. Computational recommendations cannot replace validated critical limits, hazard-control procedures (for example, HACCP critical control points), or regulatory obligations. As requested by peer review, the technical architecture for enforcing this separation is specified concretely rather than left generic: critical control points, temperature/time lethality limits, and other validated safety parameters should be encoded as hard, non-learned constraints in a supervisory layer architecturally separate from the optimization or learning components described in Sections 7.2–7.4 (see also Figure 4; Section 9.10), so that a change in model weights, a data-drift event, or an erroneous sensor reading cannot by itself move the process outside a pre-validated safe operating envelope; any action outside that envelope should require the fail-safe response (hold, revert to last validated set point, or shutdown) rather than a model-generated recommendation. Safety constraints, fail-safe mechanisms, audit trails, model-version control, and human override should therefore be incorporated into control architectures as described. Increasing autonomy strengthens rather than eliminates the need for validation and accountability.

8.10 Energy, carbon, water, and food-waste reduction

Sustainability represents one of the strongest potential benefits of intelligent food bioprocessing, but environmental improvements must be demonstrated rather than assumed. Real-time monitoring and optimization can identify unnecessary heating, cooling, airflow, pumping, water use, and processing time. Digital twins may further support resource management by allowing alternative operating strategies to be evaluated before physical implementation. Recent food-industry evidence identifies energy and water management, resource efficiency, and waste reduction as important sustainability opportunities associated with digital-twin technologies (Abdurrahman and Ferrari, 2025). Food-waste reduction may arise through several pathways. Earlier detection of quality deterioration can prevent unnecessary processing of unsuitable materials; improved control can reduce overprocessing and rejected batches; predictive maintenance can prevent losses associated with equipment failure; and condition-based cold-chain management can improve shelf-life decisions. Digital-twin approaches across food supply chains have similarly been associated with opportunities for improved resource efficiency, reduced waste, and greater resilience (Monteiro and Barata, 2025). Energy savings, however, should not be considered equivalent to environmental sustainability. Computational infrastructure, sensor manufacture, cloud computing, equipment upgrades, water use, refrigerants, raw-material losses, and the carbon intensity of electricity can influence total environmental performance. Life-cycle assessment, carbon-footprint analysis, water-footprint assessment, and multi-objective techno-economic evaluation (Section 8.11) should therefore complement conventional indicators such as specific energy consumption; despite repeated calls for such integration in the reviewed literature, food-bioprocessing studies that report full life-cycle or Pareto-front analysis spanning quality, safety, energy, water, waste, and cost simultaneously remain uncommon, and this scarcity is reported here as an identified gap (Section 9.12) rather than implied to be resolved. The sustainability objective can consequently be expressed as a multi-dimensional problem: Product quality, Food safety, Productivity, Energy efficiency, Water efficiency, Waste reduction, Carbon reduction and Economic feasibility.

8.11 Economic feasibility and techno-economic considerations

Peer review correctly identifies that economic justification was previously absent from this review despite being decisive for industrial adoption. A complete techno-economic assessment of autonomous food bioprocessing would need to weigh capital expenditure (sensors, digital-twin software and integration, edge/cloud computing infrastructure, control-system upgrades) against operating expenditure (energy and water savings, reduced product loss and rework, reduced unplanned downtime through predictive maintenance, sensor replacement and recalibration, and cybersecurity management), together with the labour and validation cost of maintaining a system whose behaviour must remain auditable to food-safety regulators. At the level of general digital-manufacturing practice, capital projects of this kind are commonly justified through payback-period and total-cost-of-ownership analysis; however, the authors were unable to identify peer-reviewed, food-bioprocessing-specific techno-economic or Pareto-front studies reporting verified capital cost, operating savings, or payback data for autonomous or near-autonomous systems at pilot or industrial scale, as distinct from vendor-reported figures for digital-twin adoption in manufacturing generally, which are not sufficiently rigorous or food-specific to cite as scientific evidence here. This absence is itself a substantive finding: without independently verified techno-economic evidence, claims that autonomous food bioprocessing is close to being “economically preferable to conventional advanced process control” cannot currently be supported or refuted, and this review reports that as an open question rather than resolving it either way. Establishing standardized techno-economic reporting capital and operating cost, energy/resource savings, payback period, and the sensitivity of that payback to plant scale and product value for autonomous food-processing pilots is identified as a specific research priority in Section 9.12.

9 Future roadmap for autonomous food bioprocessing

The transition from predictive modelling to autonomous food bioprocessing will require more than improvements in individual algorithms. Future systems must integrate reliable sensing, physical process knowledge, real-time computation, uncertainty-aware prediction, optimization, control, and continuous validation within a coordinated architecture. Current evidence indicates substantial progress in monitoring, prediction, and optimization, but fully integrated autonomous food-processing systems remain comparatively immature. A realistic roadmap should therefore prioritize technological integration, industrial validation, safety, interoperability, and measurable engineering benefits rather than autonomy as an objective in itself (Abdurrahman and Ferrari, 2025; Krupitzer et al., 2022).

9.1 Proposed autonomous food bioprocessing framework

This review proposes an Autonomous Food Bioprocessing Framework (AFBF) in which information moves continuously between the physical process and its computational representation (Figure 5). Smart sensors characterize the food material, equipment, and processing environment; predictive models estimate current and future states; a physics-informed digital twin evaluates possible process trajectories; optimization determines suitable operating actions; and a controller implements selected decisions through physical actuators. Subsequent measurements provide feedback for state updating and model evaluation. The proposed architecture is in Figure 5.

Safety constraints should surround rather than merely form another step within this architecture. Temperature limits, microbial-safety requirements, equipment operating envelopes, quality specifications, emergency shutdown conditions, and other critical restrictions must remain enforceable regardless of the recommendations generated by computational models. Autonomous operation should therefore be understood as constrained autonomy, where adaptation occurs within scientifically validated and regulatory-compliant boundaries.

9.2 Multimodal and physics-informed AI

Future process monitoring is likely to rely increasingly on combinations of sensor modalities rather than individual measurements. Temperature, humidity, pressure, images, spectra, acoustic signals, volatile compounds, electrical properties, and equipment telemetry describe different aspects of the same process. Multimodal models can potentially integrate these signals to provide a more complete representation of product and process state. Physical knowledge should remain central to this development. Food processing is governed by heat and mass transfer, fluid mechanics, thermodynamics, reaction kinetics, phase transitions, and microbial behaviour. Incorporating such constraints can reduce dependence on large datasets and limit physically implausible predictions. Digital food-twin research already supports the integration of food-science models with data-driven approaches as a means of addressing biological variability and incomplete process knowledge (Krupitzer et al., 2022). Future work should therefore move beyond comparisons of isolated machine-learning algorithms toward evaluating whether multimodal, hybrid, and physics-informed models improve robustness, extrapolation, and control performance under changing production conditions.

9.3 Self-updating digital twins

Most current digital twins depend on models developed before deployment. Long-term industrial use, however, exposes these models to changes in raw materials, sensor behaviour, equipment condition, seasons, formulations, and production schedules. A model that remains static may progressively become less representative of the physical process. Self-updating twins offer a potential solution through continuous state and parameter estimation. New observations could be used to recalibrate selected parameters, identify model drift, and update predictions without rebuilding the complete system. Such adaptation must be controlled carefully. Unrestricted online learning could introduce instability or allow erroneous sensor measurements to modify a previously validated model. Future digital twins should therefore incorporate model-drift detection, uncertainty thresholds, version control, validation gates, and rollback mechanisms. Recent food-sector assessments similarly identify model accuracy, real-time data integration, interoperability, and validation as major requirements for wider digital-twin implementation (Abdurrahman and Ferrari, 2025; Li et al., 2026).

9.4 Edge AI and real-time computing

Autonomous control requires decisions within the time scale of the physical process. Transmitting every measurement to remote cloud infrastructure may introduce latency, connectivity dependence, cybersecurity exposure, and data-governance concerns. Edge computing provides an alternative in which selected computations are performed near the processing equipment. Edge deployment could support rapid image inspection, anomaly detection, soft sensing, quality estimation, and control decisions while cloud systems remain responsible for computationally intensive model development, historical analysis, and fleet-level learning. This creates a distributed architecture in which edge systems manage time-critical decisions and cloud infrastructure supports longer-term intelligence. Model compression, pruning, quantization, reduced-order modelling, and computational surrogates will be important where high-fidelity models cannot execute within industrial control intervals; the reduced-order digital-twin approach reported by Kannapinn et al. (2022), Kannapinn et al. (2026), which achieves faster-than-real-time inference from a physics-based model trained on a single simulation dataset, illustrates one concrete route to this requirement. The objective should not be maximum model complexity but sufficient predictive reliability at the computational speed required for safe process intervention, consistent with the latency criterion introduced in Section 8.8.

9.5 Reinforcement learning for autonomous control

Reinforcement learning is potentially valuable for processes requiring sequences of decisions rather than a single optimal set point. An RL controller could learn how changes in temperature, airflow, feeding rate, mixing intensity, pressure, or cooling influence subsequent quality, energy consumption, and productivity. The reward function could represent several objectives simultaneously. Direct exploration on industrial food equipment is unlikely to be acceptable when inappropriate actions could compromise safety, product quality, or equipment integrity. Digital twins provide a potential training environment in which alternative control actions can be evaluated virtually. This suggests a future pathway of: Validated process model/digital twin → simulated control learning → safety-constrained policy → pilot-scale testing → supervised industrial deployment. Research in process industries already identifies RL as a promising approach for control and operational optimization, while also emphasizing safety, data efficiency, interpretability, and deployment as important challenges (Nian et al., 2020). Food-specific studies should therefore focus less on demonstrating that RL can optimize a simulated process and more on establishing safe transfer from simulation to physical processing.

9.6 Generative AI and foundation models in food engineering

Foundation models and generative methods represent a more speculative but potentially important research direction. Large models trained across heterogeneous scientific and industrial datasets could support literature-informed process design, troubleshooting, multimodal interpretation, equipment diagnostics, or generation of candidate operating strategies. Generative models may also assist in exploring complex formulation or process-design spaces. Their role in autonomous control should nevertheless be approached cautiously. General-purpose foundation models do not inherently understand physical constraints, food-safety requirements, or the consequences of incorrect process interventions. Outputs may also be difficult to reproduce or validate. Direct control of safety-critical food operations by unconstrained generative systems would therefore be inappropriate without deterministic safety layers and validated process models. A more credible near-term role is decision support above the control layer. Foundation models could assist operators in interpreting alarms, retrieving process knowledge, summarizing historical events, or proposing candidate interventions, while validated controllers and physical models retain responsibility for process execution. Research should distinguish clearly between such decision-support applications and claims of autonomous control.

9.7 Autonomous laboratories and self-optimizing processes

Autonomous laboratories provide a potential route for accelerating food-process development before autonomous control is transferred to manufacturing. Robotic experimentation can be combined with automated measurements and optimization algorithms to select subsequent experiments according to results obtained from earlier trials. Rather than executing a predetermined experimental design, the system progressively explores the process space. Such approaches could be valuable for fermentation optimization, extraction, formulation, drying protocols, alternative proteins, enzyme processing, and other applications involving large experimental spaces. Related developments in chemistry and materials science have demonstrated the feasibility of combining robotics, computational decision-making, and iterative experimentation within self-driving laboratories, achieving the FBAF Level-5 criterion of Section 7.8 within that adjacent domain (MacLeod et al., 2020). Food systems introduce additional complications, including biological variability, perishability, cleaning requirements, sensory properties, and food-safety constraints. Autonomous laboratories should therefore be viewed as an emerging research opportunity rather than established food-industry practice.

9.8 Standardized datasets and benchmarking

Progress in intelligent food processing is currently hindered by the absence of widely accepted benchmark datasets. Studies frequently use proprietary or laboratory-specific data collected under different operating conditions, making direct comparisons among models difficult. High reported accuracy may consequently reflect dataset characteristics rather than genuine superiority of a modelling approach. Future datasets should include raw sensor measurements, processing conditions, reference quality measurements, equipment characteristics, sampling information, metadata, and clearly defined training and testing partitions. Where commercial confidentiality prevents full data sharing, standardized benchmark experiments and federated approaches may provide alternatives. Benchmarking should also extend beyond predictive accuracy to the four-part standard proposed in Section 8.8: accuracy, calibration, robustness, computational latency, energy requirements, uncertainty, transferability, and closed-loop control performance. Such standards would shift evaluation from “Which algorithm gives the highest R ²?” toward “Which system remains reliable enough to support physical process decisions?”

9.9 Industrial-scale validation

The laboratory-to-industry gap remains one of the most important barriers to autonomous food processing. Models developed from controlled experiments may encounter substantially greater variability in industrial environments. Raw-material composition, production rates, equipment condition, sensor drift, ambient conditions, cleaning cycles, and operator interventions can all differ from laboratory conditions. In direct response to peer review, the barriers specific to this transfer are: (i) sensor drift and fouling accumulating over continuous industrial duty cycles far longer than typical laboratory trials (Section 8.7); (ii) raw-material variability across seasons, suppliers, and cultivars exceeding the range sampled during model development (Sections 5.1, 5.7); (iii) equipment aging and wear altering the physical relationships a mechanistic or hybrid model assumes (Section 6.7); (iv) cleaning-in-place and changeover cycles introducing periodic disturbances and downtime not represented in continuous laboratory operation; (v) communication failures and legacy-equipment interoperability limiting the sensor and actuator connectivity a closed-loop system requires (Sections 8.7, 8.9); and (vi) routine operator intervention, which both compensates for model limitations in current practice and complicates attribution of performance to the automated system versus human correction. A staged validation pathway is therefore recommended: Laboratory development → independent experimental validation → pilot-scale demonstration → prospective industrial validation → supervised closed-loop deployment → long-term performance monitoring. Validation should include abnormal operating conditions rather than only nominal production. Failures in sensors, communication, actuators, and model predictions should be deliberately evaluated where this can be done safely. Industrial-scale evidence should also report economic and environmental outcomes (Section 8.10, Section 8.11), since technical feasibility alone does not establish commercial value.

9.10 Regulatory and safety frameworks

Increasing autonomy changes the nature of process responsibility. When computational systems recommend or implement decisions affecting food safety, questions arise concerning validation, traceability, accountability, model updates, and human oversight. Existing food-safety principles remain applicable, but autonomous systems require additional consideration of software and data integrity. Future regulatory frameworks should address model validation, change control, data provenance, cybersecurity, auditability, uncertainty thresholds, human override, fail-safe behaviour, and post-deployment monitoring. Computational models should not be permitted to override validated critical safety limits solely because an optimization algorithm predicts an efficiency benefit. As requested by peer review, the “independent safety layer” introduced conceptually in Section 8.9 is developed here in more architectural detail: the safety layer should (a) enumerate the process’s critical control points and their validated limits as explicit, non-learned rules rather than parameters within a trainable model; (b) monitor every controller output against those rules before actuation, with authority to block, clip, or override any action that would violate a limit; (c) fail to a defined, pre-validated safe state (hold, revert, or shutdown) rather than to an unconstrained default when sensor data, model confidence, or communication falls outside validated bounds; (d) log every override event with the triggering condition, for auditability and post-incident review; and (e) remain subject to a separate, periodic revalidation and change-control process from the optimization or learning components it supervises, so that updates to the AI system cannot silently alter safety behaviour. The most defensible architecture is therefore likely to separate adaptive optimization from this independent safety layer entirely: the intelligent controller may optimize within an approved operating envelope, while deterministic constraints prevent operation outside validated limits.

9.11 Research priorities toward autonomous food factories

The next phase of research should move from isolated proof-of-concept models toward integrated systems demonstrated under realistic processing conditions. Eleven priorities emerge from the evidence reviewed:

• Develop multimodal sensing capable of estimating product state continuously rather than relying solely on equipment variables.
• Integrate mechanistic food-engineering knowledge with data-driven prediction.
• Establish uncertainty-aware models capable of recognizing conditions outside their validated operating domain.
• Develop self-updating digital twins with controlled model-drift management.
• Create computationally efficient models suitable for edge deployment and real-time control.
• Investigate safety-constrained reinforcement learning using validated virtual environments before physical deployment.
• Establish standardized datasets, reporting practices, and benchmarks for food-process intelligence, including the four-part validation standard proposed in Section 8.8.
• Validate intelligent controllers across products, batches, seasons, equipment, and manufacturing sites, using the laboratory-to-industry pathway detailed in Section 9.9.
• Integrate quality, safety, energy, water, waste, productivity, and carbon impacts within multi-objective control, supported by verified techno-economic evidence (Section 8.11).
• Establish cybersecurity, regulatory, governance, and human-oversight requirements appropriate to increasing autonomy, following the safety-layer architecture in Section 9.10.
• Demonstrate that higher autonomy provides measurable advantages over well-designed conventional control before pursuing full automation.

These priorities suggest that the future autonomous food factory should not be defined simply by the presence of advanced computational methods. Its defining characteristic should be the capacity to sense, understand, predict, decide, act, verify, and adapt within validated physical and safety constraints (Figure 5). An independent safety and governance layer should operate across all stages rather than appearing only at the end of the architecture. This roadmap also aligns with the Food Bioprocess Autonomy Framework proposed earlier in this review. Current systems largely occupy FBAF Levels 1–3 (monitoring, automation, prediction, and decision-support), with a growing number of Level 4 demonstrations in thermal processing and drying (Table 6), whereas progression toward Level 5 requires demonstrated closed-loop control, adaptation, and reliable operation under changing conditions. The principal research challenge is therefore not simply to develop increasingly complex computational models, but to establish whether these models can be integrated with physical food processes in ways that are safe, explainable, transferable, computationally practical, economically justified, and environmentally beneficial (Table 6).

TABLE 6. Critical challenges and research priorities toward autonomous food bioprocessing.

9.12 Ranking of food-bioprocessing operations for transition to autonomy

Peer review requested an explicit ranking of which food-bioprocessing operations are the most realistic near-term candidates for autonomous operation. Table 7 ranks the operations discussed in Section 8 against five criteria drawn directly from the evidence synthesized in this review: technological maturity (current FBAF level, from Table 7; Table 5), data availability (breadth of published sensing and modelling evidence), process complexity (number and nonlinearity of interacting variables), food-safety risk (consequence severity of an incorrect autonomous action), and demonstrated potential for closed-loop control (existence of a published closed-loop or digital-twin-MPC demonstration, as opposed to monitoring or offline optimization alone). Rankings are the authors’ qualitative synthesis of the evidence reviewed, not a quantitative index, and are offered as a prioritization starting point rather than a definitive industrial roadmap.

TABLE 7. Ranking of food-bioprocessing operations for near-term transition toward autonomy.

Read jointly with Table 4, this ranking indicates that thermal processing and drying currently offer the most realistic near-term pathway to demonstrated FBAF Level 4 operation, precisely because they combine comparatively mature sensing and modelling evidence with at least one published closed-loop or digital-twin-MPC demonstration; cold-chain and refrigeration applications offer a comparably strong case for autonomous monitoring and decision support specifically, short of full closed-loop control of physical refrigeration equipment; and fermentation and non-thermal processing, despite their scientific interest, remain constrained principally by the difficulty of continuously measuring the relevant biological or treatment-specific state variables rather than by control-algorithm limitations.

10 Conclusion

Food bioprocessing is progressing from conventional automation and offline quality assessment toward increasingly connected systems capable of real-time monitoring, prediction, optimization, and intelligent process control. Advances in smart sensing (including in-line PAT, hyperspectral imaging, and electronic-nose/tongue systems), machine learning, hybrid physics–data modelling, digital twins, and process optimization have created demonstrated opportunities to improve product quality, energy efficiency, productivity, and resource utilization. Current evidence, however, indicates that technological maturity remains uneven and is concentrated at FBAF Levels 1–3: most reported applications address monitoring, predictive modelling, and offline optimization, whereas validated closed-loop and self-learning systems remain comparatively limited, with a small but growing number of Level 4 demonstrations concentrated in thermal processing and drying. A central conclusion of this review is that predictive accuracy should not be equated with process autonomy. Reliable autonomous food bioprocessing requires integration of sensing, state estimation, physically credible prediction, uncertainty assessment, multi-objective optimization, control, actuation, and continuous feedback, evaluated against the four-part validation standard proposed in Section 8.8 (robustness, calibration, latency, and degradation) rather than accuracy metrics alone.

The principal gap between current intelligent food processing and true autonomy is therefore not the absence of increasingly sophisticated computational models, but the limited integration and industrial validation of these technologies as complete, safety-supervised closed-loop systems, together with a scarcity of verified techno-economic and life-cycle evidence (Section 8.10, Section 8.11). Progress toward greater autonomy must consequently be supported by external validation across products, batches, equipment, seasons, and production environments rather than relying predominantly on laboratory-scale predictive performance. Autonomous control should also remain bounded by validated physical and safety constraints, enforced through the independently governed safety layer described in Section 9.10, with appropriate human oversight. Higher levels of autonomy will be valuable only where they provide demonstrable, independently verified improvements over robust conventional control in terms of process consistency, resource efficiency, economic performance, safety, or sustainability not where autonomy is pursued as an end in itself. The Food Bioprocess Autonomy Framework proposed in this review, now situated explicitly against ISA-95, RAMI 4.0, and SAE J3016-derived scales with measurable per-level criteria (Section 7.9; Table 4), together with the application-level ranking in Section 9.12, provides a basis for benchmarking specific published and proposed systems rather than a general narrative of progress. The immediate research priority is not full autonomy as an end in itself, but the development of trustworthy systems that can progressively sense, predict, decide, act, verify, and adapt under realistic, economically justified industrial conditions.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Anbarasan Rajan, Defence Food Research Laboratory (DRDO), India

Reviewed by: Juan Gabriel Segovia Hernandez, University of Guanajuato, Mexico

Arun Sree Vijay, Defence Food Research Laboratory (DRDO), India

Vijay Dilip Kolate, Trinity Academy of Engineering, India

Bhavadharini Boopathy, Central Food Technological Research Institute (CSIR), India