Computational optimization of food formulation and extrusion conditions for nutritional quality and antioxidant capacity.

Processed foods are important to modern dietary patterns, yet current classification systems are qualitative and provide limited guidance for designing healthier products. Here, we present a computational framework integrating an ingredient-level nutrient and bioactive compound database, mechanistic kinetic models, a machine learning predictor, and global optimization to improve nutritional quality and antioxidant capacity of processed foods, using extrusion as a case study. An large language model (LLM)-assisted data extraction workflow curated a database of 503 ingredient records from 2000 peer-reviewed papers and commercial supplier data. Bioactive and nutrient concentrations varied by several orders of magnitude among ingredients, with a median within-ingredient coefficient of variation of 36% across supplier-level entries. Ingredient selection alone accounted for a mean Nutrient Rich Foods index (NRF9.3) variability of 28% across formulations. Under a simulated reference extrusion condition, the model predicted that NRF9.3 decreased while Ferric Reducing Antioxidant Power (FRAP) increased across all 20 formulations. Variability in ingredient sourcing typically contributed 3 to 11times more to changes in NRF9.3 than extrusion, although extrusion had a greater effect in formulations rich in heat-labile vitamins. Formulation-specific optimization of extrusion temperature and residence time was predicted to improve the combined NRF9.3-FRAP score by a mean of 10% (up to 44%) over the reference condition. As a whole, this work provides a generalizable framework that integrates formulations and process parameters for systematic optimization of nutritional and bioactivity targets in food processing.