From solar heat to digital drying: integrating computational fluid dynamics, process optimisation, and digital twins for sustainable tuber-chip drying

1 Introduction

Drying remains one of the most important preservation operations in agricultural and food processing because removal of water suppresses microbial growth, slows biochemical deterioration, reduces transport mass, and increases storage stability. The process is nevertheless governed by simultaneous heat and mass transfer, so performance depends on drying-air temperature, airflow, relative humidity, material geometry, moisture diffusivity, and changing thermophysical properties rather than on temperature alone (Ekechukwu, 1999; Ekechukwu and Norton, 1999). Cassava (Manihot esculenta), yam (Dioscorea spp.), and sweet potato (Ipomoea batatas) are especially relevant because their high moisture content and rapid postharvest deterioration create a strong need for stabilisation. Tuber-drying studies show that temperature, air velocity, slice thickness, cultivar, pretreatment, and internal diffusion can materially alter drying time and final product properties (Cosme-De Vera et al., 2021; Ju et al., 2016; Kamal et al., 2020; Ojediran et al., 2020; Rashid et al., 2022; Uthayakumaran et al., 2024). These responses make tuber drying a multi-objective problem in which rapid moisture removal must be balanced against energy demand, colour, texture, starch functionality, rehydration, and nutritional quality.

Open-sun drying remains attractive because of its low direct energy cost, but exposure to dust, insects, rewetting, fluctuating weather, and poorly controlled temperature and airflow compromises reproducibility. Enclosed solar dryers retain the renewable-energy advantage while improving process control. Forced-convection systems can further intensify heat and mass transfer, but their behaviour depends on collector performance, fan-system matching, chamber geometry, tray resistance, product loading, and transient solar input (Tiwari, 2024; Venkateswarlu and Reddy, 2024; Yahya et al., 2016). Cassava studies in both mobile and comparative drying systems illustrate the practical importance of dryer configuration, energy source, and operating conditions (Menya et al., 2020; Olalusi et al., 2012; Yahya et al., 2016).

The central engineering difficulty is spatial non-uniformity. Air entering a multi-tray chamber may bypass portions of the product bed, recirculate in corners, or create large tray-to-tray differences in velocity, temperature, humidity, and convective transfer. Chamber averages can therefore conceal the conditions that actually determine product moisture and quality. CFD addresses this limitation by numerically resolving mass, momentum, energy, and, when appropriately formulated, species and moisture transport. Reviews of food and solar drying identify CFD as a valuable tool for airflow diagnosis, chamber redesign, heat-transfer analysis, energy reduction, and quality-oriented optimisation, while also emphasising the need for realistic material properties and experimental validation (Getahun et al., 2021; Kidane et al., 2025a; Malekjani and Jafari, 2018; Shrivastava et al., 2025).

A second transition is now occurring from offline simulation towards intelligent drying. Statistical optimisation, machine learning, sensor fusion, and digital twins can potentially convert validated process models into rapid prediction and adaptive control. AI reviews describe growing use of neural networks, ensembles, fuzzy methods, and hybrid physics-data models for drying, while digital-twin research is beginning to establish architectures for real-time state estimation and decision support (Arefi et al., 2026; Guo et al., 2026; Miraei Ashtiani and Martynenko, 2024; Wang et al., 2025). The opportunity is particularly relevant to solar drying because irradiance and ambient conditions vary continuously.

A recurring difficulty in this interdisciplinary literature is that much of the methodological evidence for CFD, optimisation, AI, and digital twins was generated for other foods or for dryers that were not loaded with tuber chips, and is then discussed as if it applied directly to cassava, yam, and sweet potato. This mini-review therefore examines how tuber-specific drying physics, active solar-dryer design, CFD, validation, optimisation, AI, digital twins, and sustainability can be assembled into a coherent engineering pathway, while stating explicitly, for each strand, which evidence is tuber-specific and which is transferred. The organising proposition is that computational sophistication should follow, rather than precede, physical credibility: product properties and drying kinetics must inform CFD; CFD must be verified and validated; optimisation must be confirmed experimentally; and AI or digital-twin control should be built on that validated foundation. This is a reasoned engineering position derived from the synthesis, not a finding established by any single study.

The review pursues five explicit objectives: (1) to synthesise tuber-specific drying physics and kinetics for cassava, yam, and sweet potato; (2) to evaluate how CFD, numerical verification, and experimental validation can resolve the spatial non-uniformity that limits solar-dryer performance, and to compare critically how validation has been practised; (3) to assess, with a critical comparison of approaches, how response-surface and multi-objective optimisation, machine learning, and digital twins can extend a validated physical model towards prediction and control; (4) to examine, and compare critically, the energy, exergy, economic, environmental, and quality evidence needed to judge whether a given drying pathway is genuinely sustainable; and (5) to identify and prioritise the research gaps that follow from the evidence. The Conclusion returns to these five objectives in turn.

2 Methodology

2.1 Review design and reporting framework

This article adopts a structured narrative mini-review design suited to an interdisciplinary evidence base that spans postharvest engineering, food drying, solar thermal systems, CFD, optimisation, intelligent control, and sustainability. Narrative-review guidance was used to maintain a clearly defined scope and critical synthesis, while principles from SANRA and PRISMA informed transparent reporting without presenting the work as a systematic review or formal meta-analysis (Baethge et al., 2019; Ferrari, 2015; Page et al., 2021; Snyder, 2019). The guiding review question was: what does the available evidence show about how tuber-specific drying physics, active solar-dryer design, CFD, validation, optimisation, AI, digital twins, and sustainability assessment can be combined for cassava, yam, and sweet-potato chip drying, and how much of that evidence is itself tuber-specific?

2.2 Data sources and search strategy

Six sources were searched: Scopus, Web of Science Core Collection, ScienceDirect, PubMed/MEDLINE, IEEE Xplore, and Google Scholar (Table 1). Scopus and Web of Science Core Collection served as the principal multidisciplinary indexes; ScienceDirect was searched for engineering and food-science journals; PubMed/MEDLINE was included for nutrition- and quality-related tuber literature; IEEE Xplore was included for sensing, control, and digital-twin work; and Google Scholar was used to supplement retrieval and to locate early-access or conference items. Search terms were organised into five concept blocks (tuber crops; drying technology; CFD and transport modelling; optimisation and digitalisation; sustainability) and combined with Boolean operators in two complementary searches. The first (S1) targeted tuber-specific evidence; the second (S2) deliberately omitted tuber terms so that transferable computational, optimisation, and control methodology from other foods could be identified and labelled as such. Both searches were supplemented by backward and forward citation tracking from key reviews and from the most closely related primary studies. Searching was iterative and was last updated in 2026. No lower date limit was imposed, so that foundational drying theory, solar-dryer classifications, and experimental-design references could be captured, but priority was given to studies from approximately the last decade.

TABLE 1. Databases, search concepts, and search strings used to identify sources.

Truncation (*), the wildcard (?), and quotation marks follow scopus-style syntax and were adapted to the syntax of each interface. Searching was iterative and was last updated in 2026.

2.3 Eligibility criteria

Eligibility criteria were defined before full-text assessment and are summarised in Table 2. Evidence was included when it contributed directly to tuber drying, solar-dryer design, CFD methodology, numerical verification, experimental validation, drying kinetics, energy or exergy assessment, statistical optimisation, AI, sensing, digital twins, or scale-up. Studies of non-tuber products were retained only where they offered transferable computational, optimisation, or control methodology, or where they synthesised solar-dryer evidence across products.

TABLE 2. Inclusion and exclusion criteria applied during screening.

Criteria were applied at title and abstract stage and again at full-text stage (Section 2.4).

2.4 Screening and study selection

Records were screened in two stages by the author: titles and abstracts were first assessed against the criteria in Table 2, and the full texts of potentially eligible records were then read. Duplicates across databases were removed manually by comparing titles and digital object identifiers. Sources identified through citation tracking were subjected to the same two-stage screening. A source was retained only if it met at least one inclusion criterion and no exclusion criterion. Because searching was iterative and concept-driven, record counts were not tabulated at each stage, and a PRISMA-style flow diagram with numbers of records is therefore not presented. The outcome of selection is instead reported transparently as the complete evidence base of 45 cited sources, each assigned to an evidence class in Table 3.

TABLE 3. Evidence classes and the sources assigned to each.

T = tuber-specific; X = transferred; F = foundational. Classification describes each source’s relationship to tuber-chip drying and is not a judgement of study quality. Maphosa (2026), class X1 concerns pomegranate arils and Li et al. (2025), class X2 concerns apple slices.

2.5 Evidence classification and critical appraisal

Every source was classified by the strength of its relationship to tuber-chip drying (Table 3). Tuber-specific evidence (classes T1 and T2) refers to studies whose drying, dryer-performance, or modelling data were obtained on cassava, yam, sweet potato, or potato, or that review the drying of such crops. Transferred evidence (classes X1 and X2) refers to studies of other foods or product-independent dryer systems whose methods or findings are applied to tuber chips by analogy. Foundational theory and reporting guidance (class F) is kept separate because it supports method rather than findings. The classification describes the relationship of a source to the review question, not its quality. Potato is a tuber but not one of the three focal crops, so evidence from potato is treated as tuber-specific but only indirectly transferable to cassava, yam, and sweet potato.

Critical appraisal was qualitative and SANRA-informed rather than based on a formal risk-of-bias instrument. Each source was read for the product and dryer studied, whether the dryer was loaded or empty, which variables were validated and under what conditions, whether measurement or model uncertainty was reported, and whether performance metrics were defined in a way that permits comparison with other studies. These appraisal questions determined how much weight and what wording each source received in the synthesis; sources were not excluded on appraisal grounds.

2.6 Data extraction and synthesis

For each source the following were extracted where reported: product and geometry, dryer type and operating range, modelling approach, validation variables and loading condition, performance metrics and their definitions, and stated limitations. Because system architectures, crops, climates, numerical assumptions, and outcome definitions were heterogeneous, findings were synthesised thematically rather than pooled quantitatively. Numerical values from different studies are reported side by side only to illustrate the scale of reported effects; they are compared on design, definition, and context (Tables 4–7) and are not ranked unless the metric and system boundary were demonstrably equivalent. Statements that a study type was “not identified” refer to the evidence base assembled for this review and do not exclude the existence of studies outside it.

TABLE 4. Comparison of CFD validation practice in the cited solar-dryer studies.

Error definitions, validated variables, and measurement locations differ among sources, so reported percentages are not directly comparable. All four rows are transferred evidence (class X1, Table 3).

3 Statistical and analytical methods applied in the reviewed literature

Although this is a narrative synthesis rather than a primary experimental study, the underlying literature relies on a recurring set of statistical and analytical tools to characterise and compare drying parameters, and these are summarised here for transparency. Moisture-loss kinetics are typically fitted with thin-layer semi-empirical models selected by goodness-of-fit statistics such as the coefficient of determination (R²), root-mean-square error (RMSE), sum of squared errors (SSE), reduced chi-square (χ²), and mean absolute error (MAE) (Kamal et al., 2020; Ojediran et al., 2020; Okunola et al., 2023). Effective moisture diffusivity is estimated from Fick’s second law applied to the falling-rate period, and activation energy is subsequently derived from an Arrhenius-type relationship between diffusivity and drying-air temperature (Hassini et al., 2007; Okunola et al., 2023). Process-optimisation studies use response-surface methodology built on central-composite or Box–Behnken experimental designs, with analysis of variance (ANOVA) used to test the significance of individual factors, interaction terms, and lack of fit before a quadratic response surface is accepted (Bezerra et al., 2008; Box and Behnken, 1960; Ferreira et al., 2019; Szpisják-Gulyás et al., 2023). Machine-learning and digital-twin studies extend this evaluation logic with predictive-accuracy metrics including R², relative RMSE, mean bias error, and mean absolute percentage error together with error-based validation of CFD predictions against independent experimental measurements (Li et al., 2025; Loksupapaiboon et al., 2025). Because these statistical approaches originate from the primary studies rather than from new pooled computation, they are reported descriptively here; no meta-analytic or pooled statistical test was performed on the included literature. The use of common metric names does not imply that values are comparable across studies, since data-splitting strategy, operating range, and metric definition differ.

4 Tuber-chip drying as a coupled biological and transport process

4.1 Product properties, kinetics, and moisture transport

Root and tuber tissues are heterogeneous biological materials whose drying behaviour reflects composition, cellular structure, porosity, starch content, initial moisture, density, heat capacity, thermal conductivity, and shrinkage. These properties determine heat penetration and internal moisture migration and can evolve substantially during dehydration. Consequently, assigning generic constant food properties to tuber chips may be adequate for preliminary airflow studies but becomes increasingly problematic when CFD is expected to predict local moisture or quality (Cosme-De Vera et al., 2021; Hassini et al., 2007). Work on potato, a tuber outside the three focal crops, indicates that the treatment of shrinkage is an important consideration when moisture diffusivity is estimated from convective drying kinetics (Hassini et al., 2007), and diffusivities for cassava, yam, and sweet potato should be interpreted with that in mind.

During the falling-rate period that dominates many tuber-drying processes, internal moisture transport becomes increasingly important. Thin-layer models provide practical descriptions of temporal moisture loss, whereas effective diffusivity provides a transport-oriented parameter for Fickian approximations. Yam experiments demonstrate strong effects of drying temperature, relative humidity, air velocity, and thickness on moisture ratio, effective diffusivity, internal temperature, and rehydration behaviour (Ju et al., 2016; Kamal et al., 2020; Ojediran et al., 2020). Reported effective moisture diffusivities for yam span more than an order of magnitude, from roughly ×6.4 10⁻⁹ to 1.6 × 10⁻⁷ m² s^-1, with activation energies of 11–55 kJ mol^-1, and water yam reaches up to 4.2 × 10⁻⁸ m² s^-1 (Ju et al., 2016; Kamal et al., 2020; Ojediran et al., 2020; Okunola et al., 2023; Uthayakumaran et al., 2024). A spread of this size is unlikely to reflect species alone; differences in temperature range, slice thickness, air velocity, the geometry assumed in the Fickian solution, and the treatment of shrinkage plausibly contribute. Diffusivities are therefore best treated as method- and condition-specific parameters rather than species constants, and the yam values cited here derive predominantly from controlled convective drying rather than from fluctuating solar input. Water-yam studies further show why kinetic performance should be interpreted together with energy and exergy indicators rather than moisture removal alone (Okunola et al., 2023).

4.2 Crop-specific implications for cassava, yam, and sweet potato

4.2.1 Cassava

Cassava chips require rapid stabilisation after harvest, and their drying response depends strongly on airflow, thermal conditions, product thickness, and dryer configuration. Solar-dryer and technology evaluations show that system architecture can materially influence drying rate and thermal performance, which supports the use of forced convection and computational airflow analysis where uniformity is a priority (Menya et al., 2020; Olalusi et al., 2012; Yahya et al., 2016). The cassava evidence assembled here consists mainly of dryer-performance evaluations; cassava-specific CFD or diffusivity studies were not identified, so arguments about airflow uniformity in cassava-chip dryers rest on transferred evidence.

4.2.2 Yam

Yam is a particularly useful model of the trade-off between process intensification and product quality. Increasing temperature accelerates moisture removal and can increase effective diffusivity, yet the same thermal exposure can alter flour functionality and physicochemical attributes. More recent work on purple yam similarly reinforces the need to couple drying kinetics with nutrient and functional quality (Kamal et al., 2020; Uthayakumaran et al., 2024). Yam offers the most detailed kinetic evidence of the three crops, but the studies cited on moisture ratio, diffusivity, and ANFIS modelling were conducted in controlled convective dryers (Ju et al., 2016; Ojediran et al., 2020; Okunola et al., 2023); solar operation with its fluctuating inputs is not represented in that evidence.

4.2.3 Sweet potato

Sweet potato has comparable complexity because cultivar-dependent sugars, pigments, starch, and cellular structure influence browning, texture, nutrient retention, shrinkage, and rehydration. Reviews of sweet-potato drying technologies emphasise that energy-efficient operation must be evaluated alongside quality-related attributes (Rashid et al., 2022). A recent indirect solar-drying study is especially important because it combined sweet-potato drying with machine learning and 4E analysis, providing a direct bridge between tuber drying, intelligent modelling, and sustainability assessment (Abdelkader et al., 2024). It is, however, the only study in the assembled evidence base to do so for a tuber in a solar dryer, and it describes a single system and crop.

4.2.4 Evidence provenance

Across the three crops, tuber-specific evidence (classes T1 and T2 in Table 3) accounts for 12 of the 45 cited sources. It concentrates on yam kinetics in laboratory convective drying, cassava dryer performance, and one integrated sweet-potato solar study. For CFD validation, optimisation under uncertainty, surrogate modelling, and digital twins, the sections below draw largely on transferred evidence (classes X1 and X2), and this is stated where it applies.

5 Active solar dryers and CFD-enabled design

5.1 Solar-energy capture, forced convection, and chamber design

Solar dryers can be direct, indirect, mixed-mode, or hybrid, with airflow driven naturally or mechanically. Active systems use fans or blowers to impose airflow, improving controllability but adding pressure-drop and electrical-energy considerations. Recent reviews highlight thermal storage, hybrid heating, improved collectors, photovoltaic integration, and active control as major routes for extending useful drying beyond periods of strong irradiance (Saha et al., 2024; Tiwari, 2024; Venkateswarlu and Reddy, 2024; Zuluaga-Domínguez and Nieto-Veloza, 2025). These reviews address fruits, vegetables, fish, and agricultural products broadly rather than tuber chips specifically, so their conclusions are transferred evidence here.

For tuber chips, chamber design must distribute heated air through trays and product beds with minimal bypass and acceptable pressure loss. Empty-chamber behaviour can differ markedly from loaded operation because the product changes aerodynamic resistance, thermal capacity, evaporation, and humidity. Porous-medium representations provide a practical compromise when resolving every chip is computationally prohibitive, but resistance coefficients should be experimentally supported.

Collector optimisation should also avoid equating high outlet temperature with superior performance. Useful heat depends jointly on mass flow and temperature rise, while the chamber requires sufficient flow to transport both sensible heat and evaporated moisture. This is why collector, fan, duct, chamber, tray, and product load should be treated as one coupled flow network rather than as independent components.

5.2 CFD formulation, verification, and physical realism

CFD offers spatial access to velocity, pressure, temperature, turbulence, and water-vapour fields that cannot be reconstructed reliably from a small number of inlet and outlet measurements. For solar dryers, common Reynolds-averaged turbulence formulations include standard, RNG, and realisable k-epsilon models and k-omega SST. The literature reviewed does not support one universally superior model; selection should reflect expected flow physics and should be tested where conclusions are sensitive to turbulence formulation (Getahun et al., 2021; Kidane et al., 2025a). Both sources concern solar dryers for fruits and vegetables or solar drying chambers in general, so this guidance is likewise transferred to tuber chips.

High-quality CFD requires more than attractive contour plots. Geometry, mesh, material properties, boundary conditions, radiation treatment, product loading, discretisation, convergence criteria, and conservation checks determine numerical credibility. Grid-independence analysis is a verification step; agreement with measured temperatures, velocities, humidities, or drying curves is validation. These concepts should not be conflated. Experimental–numerical work on solar-dryer uniformity demonstrates the value of directly testing CFD-predicted improvements in the physical system (Kidane et al., 2025b).

Recent CFD-assisted prototyping, in this case for pomegranate arils, illustrates a useful translational model: numerical design is used to identify transport limitations, but the final engineering claim is anchored in physical implementation and measured performance (Maphosa, 2026). Future tuber studies would be strengthened by adopting the same logic while adding product-specific moisture transport and quality responses.

5.3 Critical comparison of CFD validation practice

The CFD studies cited here differ less in numerical technique than in what was validated, against what, and under what loading (Table 4). Three levels of validation stringency can be distinguished. The first compares simulated airflow and temperature with measurements in an unloaded chamber, as in the CFD–ANN–genetic-algorithm study of a household direct solar dryer, where predictions agreed with no-load experiments to within 7.3% error before being used in optimisation (Loksupapaiboon et al., 2025). The second tests whether a design change predicted by CFD produces the predicted improvement in the physical system (Kidane et al., 2025b). The third carries the numerical design through to a physical prototype and assesses measured performance (Maphosa, 2026). Reviews converge on the need for realistic material properties, experimental validation, and consistent reporting of grid independence and convergence (Getahun et al., 2021; Kidane et al., 2025a; Malekjani and Jafari, 2018; Shrivastava et al., 2025), but as secondary sources they do not themselves establish CFD accuracy for tuber chips.

Two cautions follow. First, the 7.3% error should not be read as a general accuracy benchmark: it applies to no-load airflow and temperature fields in a specific dryer, and percentage errors depend on which variable, measurement location, and error definition were used, so they cannot be compared across studies without matching these. Second, agreement in an empty chamber does not test the aspects that matter most for tuber chips, namely, product moisture, humidity, evaporative cooling, and the flow resistance of a loaded bed. The evidence assembled for this review did not include a tuber-chip CFD study validated against product moisture and quality in loaded operation. The defensible conclusion is that CFD is an established tool for diagnosing airflow non-uniformity in solar dryers, but not yet an evidence-backed predictor of moisture and quality for cassava, yam, or sweet-potato chips.

6 Experimental validation and process optimisation

6.1 Validation as an evidence gate

A dryer model should be validated against the variables it claims to predict. Temperature-only validation is insufficient for a model presented as a drying model if humidity, product moisture, or evaporation are central to the stated mechanism. Validation should therefore progress from airflow and temperature to moisture trajectories and, where feasible, product-quality indicators. Sensor placement is best informed by expected spatial gradients, with measurements at inlets, recirculation regions, tray centres and edges, downstream positions, and exhaust locations.

Model accuracy should be reported with direct error metrics such as RMSE, MAE, relative error, or MAPE rather than relying on R² alone. Calibration, validation, and independent testing should be separated where data permit. A model validated at only one operating condition may reproduce calibration-specific behaviour but still fail when airflow, load, temperature, or weather changes.

Instrument uncertainty is equally important. Agreement between prediction and experiment should be interpreted in relation to measurement accuracy, particularly for humidity and low-velocity airflow. Numerical verification, experimental uncertainty, and model-form uncertainty together define the confidence that can reasonably be placed in CFD-supported design decisions.

6.2 RSM and multi-objective optimisation

Once the physical model has been validated, response surface methodology provides an efficient way to map interactions among a manageable number of operating and design variables. Box–Behnken and central composite designs remain established approaches for constructing quadratic response surfaces and reducing the number of experimental or computational cases required relative to exhaustive factorial testing (Bezerra et al., 2008; Box and Behnken, 1960; Ferreira et al., 2019; Szpisják-Gulyás et al., 2023). This guidance derives largely from analytical chemistry, food analysis, and general experimental design rather than from tuber drying.

For tuber-chip solar drying, candidate factors include airflow velocity, drying-air temperature, chip thickness, loading density, tray spacing, collector geometry, and thermal-storage settings. Responses should extend beyond drying time to moisture uniformity, energy consumption, pressure drop, quality, and thermodynamic efficiency. Domestic solar-dryer optimisation and recent thermodynamic optimisation studies show that the design space can be treated explicitly rather than relying on one-factor-at-a-time adjustment (Atalay et al., 2026; Gupta et al., 2026; Shimpy et al., 2024).

The next methodological step is robust rather than merely nominal optimisation. Uncertainty-aware RSM, demonstrated for thin apple slices, shows how process settings can be constrained by variability instead of being reported as a single mathematically exact optimum (Li et al., 2025). For outdoor solar drying, such thinking is especially important because irradiance, ambient humidity, product properties, and loading can change from batch to batch.

6.3 Critical comparison of optimisation approaches

Four families of optimisation approach appear in the cited sources (Table 5), and they answer different questions. Response-surface methodology yields a local quadratic approximation that is efficient when the number of factors is modest and each experiment or simulation is costly, but it describes a limited region of the design space and, when applied nominally, ignores weather and batch variability. Multi-objective and thermodynamic optimisation of solar dryers (Atalay et al., 2026; Gupta et al., 2026; Shimpy et al., 2024) treats trade-offs between performance criteria explicitly; these studies concern solar dryers generally, and their optimal settings cannot be assumed to hold for tuber chips whose quality responses differ. Simulation-driven surrogate optimisation, as in the CFD–ANN–genetic-algorithm framework (Loksupapaiboon et al., 2025), can explore a wider design space than RSM at low computational cost once trained, but its optimum is only as credible as the CFD model that generated the training data, which in that study was validated against no-load measurements. Uncertainty-aware, constrained machine-learning RSM (Li et al., 2025) is the only one of these approaches that treats variability as a design constraint, yet it was demonstrated for apple slices under controlled convective drying rather than for solar operation or tubers.

TABLE 5. Comparison of optimisation approaches identified in the cited literature.

ANN, artificial neural network; GA, genetic algorithm. Class codes refer to Table 3.

Objective functions also differ between studies (drying time, uniformity, energy, thermodynamic efficiency, or quality), so optima reported in different papers are not directly comparable. Confirmatory experiments at the recommended settings are the minimum standard for accepting any optimum, and none of the four approaches has yet been shown, within the evidence assembled here, to deliver a robust and experimentally confirmed optimum for tuber-chip solar drying.

7 Artificial intelligence, machine learning, and digital twins

7.1 AI as an accelerator of validated drying science

Machine learning is increasingly used to predict moisture ratio, drying time, energy performance, and quality from nonlinear combinations of process variables. Reviews of intelligent food drying describe artificial neural networks, tree-based ensembles, support-vector methods, fuzzy systems, and deep learning as increasingly important for monitoring and control (Guo et al., 2026; Miraei Ashtiani and Martynenko, 2024; Wang et al., 2025). In tuber drying, ANFIS has shown strong predictive capability for yam moisture kinetics in a convective hot-air dryer, while ML-based analysis has been applied directly to indirect solar drying of sweet potato (Abdelkader et al., 2024; Ojediran et al., 2020).

The role of ML should nevertheless be defined carefully. Random splitting of time-correlated drying measurements can leak information between training and test sets, generating optimistic performance. More defensible evaluation holds out complete drying runs, operating conditions, or crops. Models should also be benchmarked against simpler thin-layer, linear, or response-surface models before additional complexity is justified.

A particularly promising strategy is surrogate modelling. High-fidelity CFD can generate spatially rich training data, after which an ML model approximates the CFD response rapidly enough for optimisation or online prediction. A recent combined CFD–ANN–genetic-algorithm study demonstrates this simulation-driven concept for solar-dryer optimisation (Loksupapaiboon et al., 2025). The final design, however, still requires physical confirmation.

7.2 From digital model to digital twin

A digital twin is not synonymous with an offline simulation. It requires an operational link between the physical dryer and its virtual representation, together with data updating, state estimation, prediction, and a decision or control function. Food-drying digital-twin research remains at an early stage, but the conceptual architecture is becoming clearer: sensors observe the process; a mechanistic or hybrid model estimates hidden states; a surrogate predicts future behaviour; optimisation selects an action; and the physical response updates the model (Arefi et al., 2026).

Real-time feasibility is a major constraint because transient multiphysics CFD is generally too slow for frequent control decisions. Nazari and Malvandi (2026) addressed this problem using a region-aware graph neural network for real-time simulation of coupled heat and mass transfer in a food-drying digital-twin framework. Graph-based surrogates are attractive because they retain spatial structure rather than reducing the dryer to a small set of scalar averages.

Smart solar-drying research also emphasises automation, communication, remote monitoring, and sustainability. The systematic review by Nkolokosa et al. (2026) indicates that intelligent solar dryers can improve efficiency and drying time while supporting real-time control, although implementation remains sensitive to sensor cost, maintenance, communication reliability, and local technical capacity. A staged progression from digital model to sensor-connected digital shadow and only then to closed-loop digital twin is therefore more defensible than prematurely describing any connected simulation as a twin.

7.3 Critical comparison of AI applications

AI studies in the cited literature differ in what is predicted, from what data, and how performance is judged (Table 6). The tuber-specific applications predict lumped outputs: ANFIS for yam drying kinetics in a laboratory convective dryer (Ojediran et al., 2020), and machine learning combined with 4E analysis for an indirect solar dryer processing sweet potato (Abdelkader et al., 2024). Surrogate approaches, by contrast, learn from physics-based simulation, whether to predict objective values for optimisation (Loksupapaiboon et al., 2025) or to reproduce coupled heat and mass transfer fields in real time (Nazari and Malvandi, 2026), and they inherit the fidelity of the underlying model. Uncertainty-aware machine-learning RSM contributes a different capability, namely, constraining optima by variability (Li et al., 2025).

TABLE 6. Comparison of AI applications in the cited drying literature.

ANFIS, adaptive neuro-fuzzy inference system; 4E = energy, exergy, environmental, and economic. Class codes refer to Table 3.

Performance is usually summarised with R², RMSE, or related error metrics, but the values depend on data-splitting strategy, dataset size, and operating range, and they cannot be ranked across studies, particularly where random splits of time-correlated data may inflate apparent accuracy (Section 7.1). Within the evidence assembled here, tuber-specific AI evidence is limited to the two studies above, neither of which, as characterised in the sources reviewed, involves closed-loop control. Digital-twin architectures (Arefi et al., 2026) and real-time spatial surrogates (Nazari and Malvandi, 2026) are transferred from other foods. The reviews document rapid growth while also emphasising early-stage maturity and implementation constraints (Guo et al., 2026; Miraei Ashtiani and Martynenko, 2024; Nkolokosa et al., 2026; Wang et al., 2025). The defensible reading is that AI has shown promise for prediction, but its benefit for tuber solar drying over validated physics-based or simpler statistical models has yet to be demonstrated.

8 Energy, exergy, economic, environmental, and quality performance

8.1 Thermodynamic, economic, and environmental indicators

Thermal performance should be assessed alongside aerodynamic and moisture performance. Useful heat, collector efficiency, overall dryer efficiency, specific energy consumption, and specific moisture extraction rate quantify first-law behaviour, but they do not identify where energy quality is destroyed. Exergy analysis adds a second-law perspective by locating irreversibilities associated with finite temperature differences, pressure loss, heat leakage, mixing, and moisture-transfer processes.

The sweet-potato study of Abdelkader et al. (2024) is notable because it integrated machine learning with energy, exergy, environmental, and economic analysis in one indirect solar-drying system, reporting 15% overall thermal efficiency, a specific energy consumption of 4.5 kWh per kilogram of moisture removed, and a 29-month economic payback period. More broadly, recent reviews argue that solar-dryer evaluation should combine thermodynamic, techno-economic, and environmental metrics rather than treating renewable-energy input as sufficient evidence of sustainability (Roy et al., 2025; Saha et al., 2024; Venkateswarlu and Reddy, 2024).

Multidimensional assessment also helps expose trade-offs. Thermal storage may extend operation but adds cost and embodied impacts; higher airflow may improve external mass transfer but increase fan energy and pressure drop; auxiliary heat may reduce drying time but decrease the solar fraction. Mugi and Chandramohan (2024) extend this logic through a 7E + Q framework that explicitly includes product quality alongside energy, exergy, economic, environmental, and related cost-performance dimensions.

For decentralised tuber processing, affordability and maintainability are critical. Life-cycle and economic claims should use explicit system boundaries, locally relevant energy prices, realistic utilisation rates, and an appropriate baseline for avoided emissions. Recent sustainability reviews emphasise that low-cost design, local adaptation, training, and field validation are central to adoption, particularly where connectivity, spare parts, and reliable electricity cannot be assumed (Zuluaga-Domínguez and Nieto-Veloza, 2025).

8.2 Critical comparison of reported sustainability metrics

The metrics summarised in Table 7 show why sustainability claims require care. The 15% overall thermal efficiency, 4.5 kWh kg^-1 specific energy consumption, and 29-month payback (Abdelkader et al., 2024) describe one indirect solar dryer processing sweet potato. The exergy efficiencies of 58%–75% (Okunola et al., 2023) describe hot-air drying of water yam in a different, non-solar system, and they are second-law quantities. These values are not in conflict, and they should not be read as evidence that one dryer is better than the other: they rest on different thermodynamic definitions, reference states, energy sources, and system boundaries. Economic payback additionally depends on local energy prices, utilisation, and the baseline technology displaced, so a 29-month figure cannot be transferred to another setting without recalculation. No like-for-like comparison of solar tuber dryers using a common metric set was found in the evidence assembled here. The 7E + Q framework (Mugi and Chandramohan, 2024) offers a route to a common structure by adding product quality, but it needs consistent boundaries and datasets before it can be used to rank systems.

TABLE 7. Comparison of reported sustainability and performance metrics and their comparability.

Values are reproduced from the cited studies to illustrate scale and are not ranked. Class codes refer to Table 3.

A minimum reporting set would make future tuber-dryer studies comparable: (i) first-law indicators (collector efficiency, overall dryer efficiency, specific energy consumption or specific moisture extraction rate) with their definitions and boundaries; (ii) exergy indicators with the stated reference environment; (iii) solar fraction and the quantity and source of auxiliary and fan energy; (iv) economic indicators (payback or cost per kilogram of product) with energy prices, utilisation, and the baseline technology; (v) environmental indicators with an explicit life-cycle boundary and avoided-emission baseline; and (vi) product-quality indicators appropriate to the crop, such as colour, rehydration, and starch functionality.

9 Challenges, knowledge gaps, and future directions

9.1 Persistent methodological gaps

A recurrent limitation across the evidence assembled here is not a shortage of proposed dryer geometries but incomplete integration of product physics, validated transport modelling, and field evidence. Tuber-specific thermophysical data appear sparse across moisture and temperature ranges. Shrinkage and evolving porosity are often neglected in the CFD studies and reviews examined. Boundary conditions are frequently simplified despite transient solar input. Empty-chamber CFD may be used to infer loaded performance. Grid independence and convergence are reported inconsistently, and validation may be limited to a few temperature measurements (Getahun et al., 2021; Kidane et al., 2025a; Shrivastava et al., 2025).

Quality integration also remains weak. Airflow and temperature uniformity are useful engineering indicators, but the process objective is a safe, stable, and commercially acceptable product. Future CFD should therefore connect local thermal and moisture histories to quality kinetics wherever defensible data exist. This is especially relevant to starchy tubers, in which colour, texture, rehydration, and functional properties may respond differently to process intensity.

AI introduces additional risks of data leakage, small datasets, weak external validation, and poor transferability across crops. A model developed for one variety or geometry should not be presumed to generalise to another. Digital twins add further requirements for sensor calibration, fault detection, data reconciliation, uncertainty-aware state estimation, cybersecurity, and safe fallback control.

9.2 Prioritised research agenda

The gaps identified in Sections 4–8 do not carry equal weight. They are prioritised in Table 8 using three criteria applied through the author’s judgement: dependency (whether other research steps rely on closing the gap), evidence deficit (how little tuber-specific evidence exists in the assembled literature), and feasibility in the low-resource settings where much tuber processing occurs. Priority 1 gaps are foundational, because model credibility, optimisation, and AI all depend on them. Priority 2 gaps can proceed once Priority 1 evidence is in place. Priority 3 items are conditional on the earlier stages and should not be pursued in isolation.

TABLE 8. Prioritised research gaps, rationale for ranking, and minimum evidence needed.

Priorities reflect the author’s judgement based on the criteria stated in Section 9.2, not a formal consensus process.

Beyond these priorities, future optimisation should target robust operating regions rather than fragile single-point optima. Climate-responsive models can use measured and forecast solar conditions to schedule fan speed, thermal-storage discharge, or auxiliary heat. Physics-informed learning may help constrain data-driven predictions by mass and energy conservation, bounded moisture, and realistic product temperatures. Intelligent control should be compared experimentally against conventional fixed-operation strategies rather than assumed to provide benefit. Open benchmark geometries and datasets would allow turbulence models, moisture formulations, surrogate algorithms, and control strategies to be compared on common evidence, and multi-season field trials are needed to determine whether laboratory gains persist under realistic weather, loading, cleaning, maintenance, and operator conditions.

9.3 Limitations of this review

Several limitations should be borne in mind. The narrative design, single-author screening, and qualitative appraisal introduce the possibility of selection bias, and no formal risk-of-bias instrument was applied. Only English-language sources were considered, and coverage depends on the six databases searched and on the search strings in Table 1. Because searching was iterative, record counts were not logged at each stage. The classification of sources into evidence classes reflects the author’s reading of each source, and some sources could reasonably be assigned differently. A substantial share of the most relevant literature on digital twins and intelligent solar drying is very recent, including early-access items that have not yet been tested by wide citation. Finally, statements that a study type was not identified describe the evidence base assembled here and should not be read as proof that no such study exists.

10 Conclusion

Solar tuber drying appears to be moving from empirically operated equipment towards an integrated postharvest engineering platform in which product behaviour, renewable-energy capture, airflow, moisture transport, computation, and process control are treated as one system. The evidence synthesised above is drawn together below against the five objectives set out in the Introduction, with attention to how much of it is tuber-specific.

• Tuber-specific drying physics and kinetics. Cassava, yam, and sweet-potato drying appears to be governed by composition, porosity, shrinkage, and internal moisture diffusion rather than by drying-air temperature alone. Reported effective diffusivities for yam span roughly 6.4 × 10⁻⁹ to 1.6 × 10⁻⁷ m² s^-1 (up to 4.2 × 10⁻⁸ m² s^-1 for water yam) with activation energies of 11–55 kJ mol^-1, and they derive mainly from controlled convective drying. This spread supports crop- and condition-specific property measurement, rather than a generic constant-property assumption, as the starting point for a credible drying model (Ju et al., 2016; Kamal et al., 2020; Ojediran et al., 2020; Okunola et al., 2023; Uthayakumaran et al., 2024).
• CFD, verification, and experimental validation. CFD is particularly valuable because it resolves spatial non-uniformities such as bypass, recirculation, and tray-to-tray gradients that chamber averages conceal. Its credibility depends on explicit grid-independence checks and on validation against measured moisture, humidity, and temperature fields rather than temperature alone. A quantitative example is agreement to within 7.3% of no-load airflow and temperature measurements for a household direct dryer (Loksupapaiboon et al., 2025), and CFD-assisted prototyping has been illustrated for a non-tuber product (Maphosa, 2026). Neither demonstrates loaded, product-specific validation, which for tuber chips remains to be shown.
• Optimisation, machine learning, and digital twins. Once a physical model is validated, response-surface and multi-objective optimisation can help identify robust operating regions that balance drying time, uniformity, energy, pressure loss, and quality rather than maximising one variable in isolation. Machine-learning surrogates and, at a later maturity stage, sensor-connected digital twins may then make these validated relationships computationally tractable for real-time prediction and control, but only when trained and evaluated on validated CFD or experimental data rather than substituted for it. The tuber-specific AI evidence is limited to yam kinetics and one sweet-potato solar study, and the remaining support is transferred from other foods (Arefi et al., 2026; Li et al., 2025; Nazari and Malvandi, 2026).
• Energy, exergy, economic, environmental, and quality evidence. A dryer is not genuinely optimised merely because it is faster or solar-powered. Reported system-level values, for example, 15% overall thermal efficiency, 4.5 kWh kg^-1 specific energy consumption, and a 29-month payback period for an indirect sweet-potato solar dryer, or exergy efficiencies of 58–75% for hot-air-dried water yam, arise from different systems and definitions and cannot be ranked against one another. They show that energy, exergy, economic, and environmental performance should be reported on consistent bases and together with product quality before a drying pathway is judged sustainable (Abdelkader et al., 2024; Mugi and Chandramohan, 2024; Okunola et al., 2023).
• Prioritised research gaps. Tuber-specific evidence accounts for only 12 of the 45 sources cited. The most consequential gaps are therefore foundational: crop-specific thermophysical and kinetic data measured under solar-relevant conditions, and loaded-dryer CFD for tuber chips validated against product moisture. Robust optimisation, quality-linked modelling, and standardised sustainability reporting follow as near-term priorities, and rigorous AI evaluation, digital shadows, and closed-loop digital twins are conditional on those foundations (Table 8).

Taken together, these strands support a single staged pathway: tuber-specific properties and drying kinetics should inform the model; CFD should be numerically verified and experimentally validated; optimisation, machine learning, and digital-twin control should be built on that validated foundation and confirmed physically; and the resulting design should be judged against reproducible physical accuracy, robust product quality, affordable operation, and measurable sustainability at the scale where cassava, yam, and sweet-potato processing actually occurs.

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: Seydi Yıkmış, Namik Kemal University, Türkiye

Reviewed by: Bogala Madhu, Siddharth Institute of Engineering & Technology, India

Muhammad Hanif, University of Agriculture, Peshawar, Pakistan