Trustworthy and physics-informed artificial intelligence for food processing operations: validation, control relevance, and industrial readiness

Background Artificial intelligence (AI) applications in food processing are mostly limited to data-driven prediction models, while industry needs AI for trustworthy operational use. Much of the evidence still rests on dataset-level performance, whereas validation quality, physical credibility, uncertainty, interpretability, control relevance, and industrial readiness of AI models are examined less consistently. Scope and approach This critical review uses structured evidence mapping strategy to appraise recent reviews, methodological papers, and direct evidence studies across drying, thermal processing, extrusion, fermentation, cleaning, quality monitoring and process systems enabled by AI models. Related nonthermal, membrane, separation, and rheology-oriented studies are used only when they inform validation, process physics, monitoring, control, or industrial readiness. The final reference list contains 73 cited sources; 69 of these were coded as records in the structured evidence map, while the remaining sources supported background, conceptual framing, methodological context, or boundary positioning. Key findings and conclusions Predictive modeling and soft sensing are comparatively mature, especially when spectral, imaging, acoustic, or process data are used to estimate quality attributes or state variables. However, external validation across batches, products, equipment, and production periods remains inconsistent. Uncertainty is rarely quantified, and many claims involving digital twins or advanced control remain limited to conceptual frameworks, simulations, or laboratory-scale demonstrations. Physics-informed machine learning, hybrid mechanistic data-driven models, reduced-order models and digital twins offer promising pathways toward transferable and robust process AI, but only when underlying physical assumptions, validation domains, and deployment constraints are clearly defined. Future research should better align model claims with rigorous validation strategies, explicit physical grounding, uncertainty quantification, and measurable process-level benefits.