Artificial intelligence algorithms for complex industrial optimization

1 Overview

Artificial intelligence (AI) in industry draws on learning, reasoning, search, planning, perception, and autonomous action, often in combination with human expertise. This paper focuses on one part of that broad field: the algorithms used to formulate, solve, and revise complex industrial optimization problems. Recent accounts of AI for optimization distinguish applications to parameter generation, model formulation, and solution methods, while emphasizing that AI complements the structured decision methods of operations research (Fan et al., 2026; Wiberg et al., 2026). Such algorithms operate within physical processes, engineering knowledge, organizational workflows, and operating constraints. Their central challenge is to formulate the right decision problem, solve it at operational scale, and revise decisions as the system changes. Many industrial problems are combinatorial and NP-hard, with tightly coupled decisions across production and supply networks.

Flexible manufacturing provides a clear example. Production resources interact with material and energy flows, transport, and maintenance over several time horizons. More accurate processing-time predictions do not guarantee an executable schedule if the optimization model omits an operator-qualification constraint or a breakdown shifts the bottleneck. Likewise, a schedule that improves factory-level performance may increase downstream inventory or cause a shipment to miss its departure. Such outcomes are more than implementation defects: missing data or a misspecified model can alter the feasible region and the value of a decision. The problem solved by the algorithm may then differ from the problem faced in operation.

Credible evaluation therefore distinguishes prediction quality, optimization quality, and operational value. Prediction error, average reward, and benchmark objective values are informative, but none establishes that a decision will work in practice. A useful industrial decision needs to be feasible, stable under minor perturbations, recoverable after disruption, and economically worthwhile. Decision-focused learning addresses part of this gap (Donti et al., 2017; Wilder et al., 2019), but an end-to-end loss cannot be assumed to encode all physical, safety, and organizational requirements. Figure 1 provides a compact map of the three substantive challenges and two cross-cutting requirements developed below.

We organize the research needs around three connected challenges: AI-assisted problem formulation and structural representation; scalable optimization of coupled resources and supply chains; and closed-loop adaptation in dynamic, uncertain, and multi-objective environments. Two requirements cut across all three challenges: structure-aware integration of optimization, learning, and simulation; and operational validation, human–AI collaboration, and accountability. In the latter, information, tasks, and decision authority are allocated among domain experts, operators, AI assistants and agents, optimizers, and simulators according to risk and context (Karapanagiotis et al., 2025).

2 Grand challenges and research needs

2.1 AI-assisted problem formulation and structural representation

Industrial optimization begins before a solver is selected. Relevant information is distributed across data systems, engineering documents, regulations, and expert practice. These sources differ in timing, vocabulary, and reliability. Model formulation requires identification of the decisions, objectives, constraints, uncertainties, and couplings that define the operational problem. Translating an agreed description into equations is only one part of that task.

Methods for constraint acquisition can infer candidate rules from examples and expert queries (Bessiere et al., 2017), while decision-focused approaches can optimize predictive models for downstream decision quality (Donti et al., 2017; Wilder et al., 2019). Rules inferred from data or text remain provisional until they are checked against actual operations. Each candidate rule requires verification of its provenance, scope, conflicts with other rules, and operational consequences before it enters a decision model. Human skills, decision rights, and handoff responsibilities may themselves be part of the modeled structure, particularly when tasks can move between people and automated components (Karapanagiotis et al., 2025).

Large language models (LLMs) and LLM-based agents can mediate between informal descriptions and formal optimization models. They can propose formulations, generate solver code, and assign modeling or verification tasks to specialized components (Zhang J. et al., 2024; Xiao et al., 2024; Yang et al., 2026). Recent work that combines LLM-based extraction with reusable modeling graphs and expert-defined templates illustrates a more controlled division of responsibility (Li et al., 2026). These outputs remain candidate models. Solver execution can expose coding errors or infeasibility, but it cannot establish semantic completeness or show that an omitted business rule is inconsequential.

A traceable modeling loop records how constraints are acquired and classified, who validates them, how uncertainty and problem structure are represented, and when operational change triggers model revision. If the resulting problem is infeasible, the algorithm needs to identify the conflicting assumptions and return either an interpretable relaxation or a safe recovery action. Evaluation then turns on downstream decision performance under missing or changing information; fluent model generation alone is insufficient.

2.2 Scalable optimization of coupled resources and supply chains

The computational burden of large industrial optimization problems arises from both problem size and coordination across resources, organizations, and time scales. Machines, workforce capabilities, energy supply, inventory, and transport capacity are coupled through shared constraints, but their influence on system performance is unequal and state dependent. A disruption can shift the bottleneck from equipment to technicians, spare parts, storage, or transport. Algorithms need representations and sensitivity measures that identify the binding couplings and detect when those relationships change.

In distributed scheduling, job-to-factory assignment and within-factory sequencing are interdependent (Naderi and Ruiz, 2010). Across a supply chain, procurement, production, inventory, and delivery decisions are made over different planning horizons. Solving these stages separately can shift costs, delays, or emissions downstream; even under restricted assumptions, integrated production–inventory–delivery models are computationally demanding (Gen et al., 2008; Li et al., 2017). Coordination also involves organizations, human planners, and AI agents that hold different information and decision rights. Recent supply-chain research on agentic AI accordingly treats human–agent roles and organizational constraints as design choices rather than assuming full autonomy (Trifone et al., 2026).

Decomposition and mathematical programming can exploit sparsity and explicit constraint structure, while evolutionary algorithms and large-neighborhood search can explore difficult discrete and nonconvex search spaces (Yu and Gen, 2010; Pisinger and Ropke, 2010). A central design question is how to balance global exploration with local improvement. Global moves reallocate orders, capacity, and supply-chain flows; local moves refine sequences, routes, and maintenance schedules. Learned components may select neighborhoods, warm starts, decompositions, or algorithm configurations. Their value depends on preserving global consistency while reducing end-to-end decision time and resource use. A learned policy may improve one subproblem yet render another infeasible when cross-stage constraints are omitted, represented too coarsely, or invalidated by a disruption. Learned guidance also needs to recognize when a fixed decomposition is no longer valid.

Multiple objectives introduce an additional tension between computational efficiency and decision effectiveness. Efficiency concerns whether an algorithm returns an answer before the decision deadline. Effectiveness covers feasibility, solution-set quality and diversity, and alignment with decision-maker preferences (Deb et al., 2002; Zhang W. et al., 2024). Automated program search can generate heuristics and search rules, including variants for multi-objective optimization (Romera-Paredes et al., 2024; Yao et al., 2025). Their performance remains uncertain when the evaluation function overlooks cross-stage couplings. Evaluation should establish whether local gains persist once supply-chain interactions are accounted for and whether a globally consistent plan can be produced on time.

2.3 Closed-loop adaptation in dynamic, uncertain, and multi-objective environments

Demand, resource availability, prices, and customer priorities change over different time scales. Parameter uncertainty, observed disruptions, distribution shift, and structural change each call for a different response. Repeatedly running a static optimizer after every event does not by itself create a closed-loop strategy.

A closed-loop algorithm links state estimation and prediction to model revision, optimization, execution, and feedback. Robust planning and reactive rescheduling play complementary roles and can be coordinated through rolling-horizon optimization, online learning, dynamic optimization, and calibrated simulation. Full regeneration may improve the objective value but increase schedule nervousness; local repair preserves stability but can lock in a poor global allocation (Vieira et al., 2003; Blackburn et al., 1986). Reinforcement learning can represent delayed consequences, but exploration in industrial systems must respect constraints during training and deployment. A high expected return does not establish acceptable tail risk or recoverability (García and Fernández, 2015; Wachi et al., 2024). Recent dynamic-scheduling work shows how a formal supervisor can restrict reinforcement learning to a safe, non-blocking operating space and support rapid recovery after disruptions (Wang et al., 2026).

Digital twins are useful for stress testing only if their models remain valid under the disruptions that drive decisions (Tao et al., 2019). Emerging architectures connect digital twins, multiple agents, online multi-objective optimization, and human–robot collaboration (Huang et al., 2026). These combinations strengthen the case for a closed loop, but they also increase the number of interfaces whose state, latency, and failure modes need monitoring.

Priorities assigned to cost, service, sustainability, resilience, and workload may also change during a disruption. Adaptation therefore requires more than repeated static Pareto optimization. The algorithm needs to choose among continuing or repairing the current plan, re-optimizing, updating the model, and escalating the decision to a human decision-maker. Task and decision authority may also need to shift as risk, AI confidence, human workload, and time pressure change. Evaluation needs to report feasibility under distributional and structural shifts, tail-risk and recovery metrics, and operational gains net of excessive plan changes.

2.4 Cross-cutting design principle: structure-aware integration of optimization, learning, and simulation

Hybrid algorithmic architectures are justified only when their components perform distinct, necessary functions, their interfaces can be validated, and the integrated method outperforms a strong but simpler baseline (Fan et al., 2026; Wiberg et al., 2026). Learning can estimate uncertain states or surrogate models and select search controls when equivalent information is unavailable within the operational deadline. Mathematical and constraint programming can enforce explicit feasibility logic and provide bounds. Evolutionary algorithms are useful for discrete, nonconvex, black-box, or preference-sensitive search, whereas simulation represents behavior that is difficult to express analytically.

The interfaces between components deserve equal scrutiny because they govern information exchange, uncertainty propagation, and failure detection. A learned initialization may accelerate a solver yet fail under distribution shift. Likewise, a high-fidelity simulation may consume the computation needed for adequate stress testing. Ablation studies need to test both individual contributions and failures at component interfaces under the same operational budget.

A learning-based or generative component may propose a formulation or heuristic; a solver, simulator, or explicit checker can then test it. The inclusion of every added component should be justified by a verifiable improvement in decision quality. This division of responsibility is especially important for LLM-based agents, whose ability to orchestrate tools does not remove the need for formal checks and controlled access to operational systems.

2.5 Cross-cutting evidence requirement: operational validation, human–AI collaboration, and accountability

Operational reliability depends on how an algorithm interacts with enterprise data flows, shop-floor execution systems, and human decision-makers. Stale state information, interface mismatches, human overrides, and execution deviations can all change the optimization problem. Monitoring, rollback, re-optimization, escalation, and revalidation are consequently part of the algorithm lifecycle.

The appropriate unit of analysis is the hybrid decision team rather than the AI component in isolation. Its performance depends on how information, tasks, and authority are distributed among operators, managers, AI agents, optimization solvers, and organizational units. Human participants can contribute tacit knowledge and handle exceptions, but poor handoffs or unclear authority can delay responses and obscure accountability. Experimental evidence from AI-assisted production indicates that team organization affects failure recovery through the frequency and quality of communication (Krzywdzinski et al., 2026). Semantic models for dynamic work sharing offer one way to represent changing human and artificial roles, although their contribution must ultimately be tested through team and system outcomes (Karapanagiotis et al., 2025).

Accordingly, evaluation should include communication quality, handoff latency, calibrated reliance, intervention workload, recovery performance, and operational outcomes alongside conventional algorithmic metrics. Explanations need separate scrutiny: in a job-shop scheduling experiment, brief AI rationales increased behavioral reliance without necessarily improving understanding (Saßmannshausen et al., 2026). Trust, reliance, and decision correctness are distinct measures.

Claims of improvement require comparison against competitive optimization, learning, and engineering baselines under matched end-to-end resource budgets. Those budgets include data preparation, model development, computation, simulation, and human effort. Experiments need to vary coupling strength, constraint tightness, disruption severity, distribution shift, and decision deadlines rather than instance size alone. Studies should also report prediction quality, optimization quality, and operational value separately. Optimization quality includes feasibility, optimality gap, regret, and Pareto-front approximation quality; operational value includes plan stability, service performance, recovery time, and integration effort.

Fluent formulations and executable code do not establish correctness. Grounded in solver functions, LLM interfaces can nevertheless support infeasibility diagnosis, sensitivity analysis, and what-if or why-not queries (Chen et al., 2025). High-consequence decisions still require traceable constraints, validated alternatives, controlled system access, and explicit human authority over rollback and escalation.

3 Concluding remarks

The three challenges span problem formulation and structural representation, scalable optimization of coupled resources and supply chains, and closed-loop adaptation. Success depends on a faithful operational model, coordination of coupled decisions within the available time, and revision of those decisions without sacrificing feasibility or stability. Structure-aware integration determines whether the algorithmic components work together; operational validation, human–AI collaboration, and accountability determine whether gains at one stage translate into system-level performance.

Prediction accuracy, average reward, isolated benchmark gains, and the addition of another AI component provide incomplete evidence of progress. The stronger standard is better decision quality under realistic constraints, coupling, and change, together with communication and authority arrangements that support timely intervention and recovery. Ultimately, an industrial AI algorithm needs to solve the right problem within the decision deadline and continue to produce feasible, stable, and controllable decisions as operating conditions change.

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 and reviewed by: Luis M. Camarinha-Matos, NOVA University of Lisbon, Portugal