AI Agents Beat Machine Learning at Predicting Food Texture, Study Finds

Plant-based burgers, dairy-free cheeses, and meat analogues live or die by their texture, yet predicting how a new formulation will feel in the mouth remains one of the most stubborn challenges in food science. Mechanical and rheological properties of dense food structures emerge from tangled interactions among proteins, polysaccharides, moisture, and processing history, so two recipes with nearly identical compositions can behave dramatically differently in a texture analyzer. A new study published in Current Research in Food Science by Yizhou Ma of Wageningen University and Research suggests a way forward, and the result is striking: an artificial intelligence agent that combines simple data retrieval with encoded mechanistic food-science knowledge outperformed conventional machine learning models in predicting the texture of plant-based meat analogues and the rheology of protein-polysaccharide gels.

The research tackles a long-standing bottleneck. Food companies developing plant-based products typically rely on trial-and-error experimentation, because the relationship between formulation and final texture is nonlinear and context-dependent. Machine learning has recently been recruited to help. Earlier work demonstrated that hardness and chewiness of plant-based meat analogues can be estimated from proximate composition using supervised learning, and that machine-learning-assisted optimization can reduce experimental burden in high-moisture extrusion. But these models share a critical weakness: they are sensitive to dataset size, feature selection, and something statisticians call distributional shift. In food research, where experiments are expensive and sample numbers are small, new formulations often fall outside the range of the training data, and purely statistical models tend to fail exactly when researchers need them most.

Ma’s study compared three fundamentally different prediction approaches across two held-out prediction tasks. The first approach used standard machine learning baselines: gradient boosting, random forest, and TabPFN, a pre-trained transformer model suited to small tabular datasets. The second was pure inverse distance weighting, or IDW, a retrieval method that predicts each new sample by averaging the three most similar training samples in normalized composition space. The third was a knowledge-based AI agent that used the same retrieval logic but added an explicit layer of food-domain reasoning, drawing on a curated mechanistic knowledge base of literature-derived rules. The comparison was deliberately fair: all three methods received identical input features, so any difference in performance could be attributed to the mechanistic knowledge the agent carried with it.

The first case study used a dataset of 54 plant-based meat analogue samples characterized by seven compositional features, including protein, fat, carbohydrate, fibre, ash, moisture, and target moisture, alongside measured hardness and chewiness. To create a stringent test, the researcher held out 13 samples whose fibre and protein levels deliberately fell outside the training range, leaving only 41 samples for training. This out-of-distribution design mimics a common real-world scenario: a food developer introduces a novel protein source or an unusual fibre level, and existing models must extrapolate rather than interpolate. The prediction targets were hardness, the first compression peak force in texture profile analysis, and chewiness, the mechanical energy required to chew a sample to a swallowable state.

The results revealed a dramatic hierarchy. The knowledge-based agent achieved the lowest error for hardness, with a mean absolute error of 4.66 newtons and a coefficient of determination of 0.782, and for chewiness it reached a mean absolute error of 4.45 joules. Gradient boosting, the best machine learning baseline, was statistically comparable, though the agent was numerically ahead. Pure IDW retrieval, by contrast, performed catastrophically, with strongly negative R-squared values meaning its predictions were less informative than simply guessing the training average. The paired differences between the agent and pure retrieval were large and statistically significant, roughly 14 newtons for hardness and 11 joules for chewiness, with confidence intervals comfortably excluding zero. Notably, some machine learning models were highly unstable on the small, shifted test set, and gradient boosting behaved partly like a classifier, grouping its hardness predictions into three discrete levels that failed to capture the variation in the new formulations.

What made the agent’s success possible was the mechanistic knowledge base built into its reasoning stage. The agent, powered by a large language model, operated in two stages. First, it retrieved the three nearest training analogues using the same normalized Euclidean distance as the IDW baseline. Second, it processed these retrieved examples together with compact, literature-derived rules for plant-based meat analogues, including moisture-dependent protein network formation, the relative texturizing capacity of pea, soy, and wheat proteins under high-moisture extrusion, fibre-reinforced network stiffening, and fat-induced lubrication and softening. The model then chose an adjustment within a qualitative range defined by these rules, and when no rule applied, it simply emitted the retrieval-weighted average. Crucially, the test set remained entirely inaccessible during knowledge compilation, and no measured texture values of test samples were included in the retrieval base, ensuring strict train-test separation.

The second case study tested whether the approach would hold up on a larger dataset with different physics. It used 311 measurements on plant protein-polysaccharide mixtures, split into 246 training and 63 test samples, with two structural outliers excluded on data-inspection grounds before any predictions were run, after their maximum-stress values of roughly 4 to 7 pascals proved about three orders of magnitude below the rest of the distribution, indicating measurement failure. The prediction targets were the storage modulus, a small-deformation measure of gel stiffness, and the maximum stress at 200 percent strain, a large-deformation failure property. Here, all methods performed better thanks to the larger sample size, but the agent still came out on top. For the storage modulus it achieved a mean absolute error of 152.30 kilopascals and an R-squared of 0.695, edging out the best machine learning model, TabPFN. The gap widened dramatically for maximum stress, where machine learning models struggled badly, with the best one reaching only an R-squared of 0.412 and predictions concentrated in a narrow band that missed the full dynamic range of the test set. The agent posted a mean absolute error of 17.21 kilopascals and an R-squared of 0.765, following the parity line most closely across the measurable range.

Ablation tests confirmed where the agent’s advantage came from. When the mechanistic knowledge base was removed, leaving the language model with retrieval alone, its performance collapsed back to the IDW baseline in the meat analogue case and even fell below IDW for maximum stress. In other words, the improvement stemmed from the encoded food-science knowledge, not from the language model’s general reasoning abilities. This finding aligns with broader trends in artificial intelligence research, where retrieval-augmented and tool-using systems increasingly combine data-driven lookup with explicit reasoning, and it suggests a practical recipe: gray-box prediction that marries empirical data with mechanistic understanding, rather than relying on either alone.

The implications for food formulation are considerable. Predicting mechanical behavior from composition alone has always been fragile because texture depends on how water and biopolymer networks reorganize during processing, information that numeric feature vectors rarely capture. Knowledge-guided retrieval offers a way to stabilize predictions precisely under distribution shift, the condition that breaks conventional models, and could accelerate AI-assisted formulation screening for plant-based food design. The study also acknowledges limits: only two tabular case studies with scalar targets were examined, the agent’s quality depends on the accuracy of its encoded knowledge, and the implementation is not self-updating, so new observations must be added manually. Future work, Ma suggests, should explore continuous learning frameworks in which the agent autonomously refreshes its knowledge base as new experiments arrive, extending knowledge-guided retrieval to broader food systems, processing conditions, and mechanical targets.

Subject of Research: Comparing knowledge-based AI agents and machine learning for predicting mechanical and rheological properties of dense food structures

Article Title: Comparing AI agents and machine learning for predicting mechanical and rheological properties of dense food structures

Article References: Ma, Y. (2026). Comparing AI agents and machine learning for predicting mechanical and rheological properties of dense food structures. Current Research in Food Science, 13, Article 101578. https://doi.org/10.1016/j.crfs.2026.101578

Image Credits: AI Generated

DOI: 10.1016/j.crfs.2026.101578

Keywords: artificial intelligence, machine learning, plant-based meat analogues, food texture, rheology, food science, retrieval-augmented prediction, protein-polysaccharide gels, high-moisture extrusion, food formulation, large language models, mechanistic knowledge