Comparing AI agents and machine learning for predicting mechanical and rheological properties of dense food...

Food functionality prediction remains challenging because food materials are multicomponent, multiscale, and process-dependent. This challenge is particularly evident in plant-based meat analogues, in which composition, ingredient source, hydration, and processing history jointly determine structure and texture. This study compares three prediction paradigms for food mechanical measurements, namely deterministic machine learning (ML), inverse-distance-weighted (IDW) analog retrieval, and a knowledge-based AI agent that combined retrieval with explicit mechanistic priors. Two case studies were used. Case 1 involved out-of-distribution prediction of hardness and chewiness in plant-based meat analogues from proximate composition features. Case 2 involved prediction of storage modulus ( $G^{\text{'}}$ ) and maximum stress at 200% strain in plant protein-polysaccharide mixtures. Across both cases, the knowledge-based agent matched or exceeded the ML baseline. In the plant-based meat analogue case, the agent achieved the best performance for hardness (MAE = 4.66 N, $R^2$  = 0.782) and near-best performance for chewiness (MAE = 4.45 J, $R^2$  = 0.696). In the rheology case, the agent achieved the best performance for both $G^{\text{'}}$ (MAE = 152.30 kPa, $R^2$  = 0.695) and maximum stress at 200% strain (MAE = 17.21 kPa, $R^2$  = 0.765). Pure IDW retrieval performed consistently worse, indicating that the main predictive gain came from mechanistic knowledge rather than retrieval alone. These results demonstrate that explicit domain priors can improve food functionality prediction, especially under distributional shift and for targets whose mechanisms are only partly encoded in numeric features.