Artificial Intelligence in Food Product Development: A Critical Review of Flavour Prediction, Texture Prediction, Ingredient Substitution and Formulation Optimisation

Abstract

Food product development remains a slow, iterative and expensive activity in which candidate formulations are screened through repeated bench trials and human sensory assessment. Machine learning and related artificial intelligence methods are increasingly promoted as a means of compressing this cycle by predicting sensory outcomes from composition, proposing ingredient replacements and directing experimental effort towards promising regions of formulation space. The purpose of this review is to examine, critically and across four connected problem domains, what the available evidence actually supports. Literature was identified through structured searching of open scholarly indexes and citation registries, supplemented by backward and forward citation tracking, and appraised for design adequacy, validation practice, data provenance and relevance to product development rather than to quality control or supply-chain analytics. The evidence is strongest and most mature for molecule-level prediction of odour and taste class, where large curated datasets, prospective validation against trained panels and reproducible benchmarks now exist. It is substantially weaker for product-level flavour, where a small number of well-designed studies contrast with a larger body of work resting on single-site panels, modest sample sizes and correlational designs. Texture prediction is less developed still: models map composition and process variables onto instrumental parameters with useful accuracy, but the link between those parameters and perceived texture remains poorly resolved, and recent evidence indicates that the mechanical descriptors most often modelled are not those most strongly associated with consumer perception. Ingredient substitution has advanced through knowledge graphs, co-occurrence networks and language models, yet is almost always evaluated against recipe-corpus proxies rather than measured sensory or functional equivalence. Formulation optimisation shows the clearest translational gains, with active learning and robotic experimentation reducing experimental burden relative to classical designed experiments in several controlled comparisons. Recurring weaknesses cut across all four domains: reliance on internal cross-validation, scarce external validation, undocumented dataset overlap, and an absence of shared benchmarks linking composition to perception. Progress will depend less on architectural novelty than on multimodal datasets that connect formulation, instrumental measurement and human perception under transparent reporting.