Machine learning and COSMO-RS integration for predicting anthocyanin extraction from berries using eutectic solvents

The development of efficient, green extraction processes for bioactive compounds from fruits requires strategies that account for both solvent-solute interactions and the inherent chemical heterogeneity of natural biomass. We present a hybrid computational framework integrating the COSMO-RS with machine learning (ML) to predict anthocyanin yields from diverse berry matrices using deep eutectic solvents (DES). A dataset of 299 experimental points spanning 15 biomass types was used to train and validate seven ML algorithms; Gradient Boosting emerged as the most accurate (R² = 0.92). The model combines COSMO-RS-derived solvent descriptors, key process variables, and a single experimental biomass descriptor: the anthocyanin content obtained with ethanol. The model was validated with independent literature data and new extraction experiments using different DES and berry matrices, showing strong agreement between predicted and experimental yields. This approach minimizes experimental screening and enables rapid optimization of sustainable extraction processes for complex plant materials.