The prediction of the resistant starch production using transformer-based deep learning models

The production efficiency of resistant starch (RS) as a functional ingredient is strongly influenced by starch structure, pretreatment strategy, and reaction conditions. This study provides a reliable data-driven framework for optimizing eco-friendly starch modification strategies and enhancing RS production. In this study, environmentally friendly physical pre-treatments—freezing–thawing (FTF), hydrodynamic cavitation (HC), and ultrasonication (US)—were combined with organic acid esterification (citric, malic, and lactic acids), starch/acid ratio (0.5–1%), and incubation temperature (60–100°C) were applied to enhance RS formation in potato starch. Thereafter, their combination effects on the degree of substitution (DS), crystallinity, pasting behavior, thermal properties, and in vitro digestibility were investigated. The developed transformer-based machine learning models demonstrated high predictive performance for RS production. The Tab-Transformer achieved test R² values of 0.94–0.98 for viscosity parameters, 0.92–0.98 for thermal properties, 0.95–0.97 for digestibility fractions, and 0.89–0.97 for degree of substitution and crystallinity. The FT-Transformer showed superior accuracy, with test R² values ranging from 0.97 to 0.99 for most outputs, including pasting value, gelatinization temperatures, RS (%), and crystallinity. For RS prediction, the FT-Transformer achieved an R²-test of 0.97, a mean absolute error (MAE) of 1.43, and a mean absolute percentage error (MAPE) of 2.49%. Thermal properties were also accurately predicted, with FT-Transformer R²-test values of 0.98–0.99 for onset, peak, and conclusion gelatinization temperatures and 0.95 for gelatinization enthalpy. Crystallinity prediction achieved an R² of 0.99 and a MAE of 0.51%, indicating strong model robustness. The DS was predicted with an R² of 0.96 and a MAPE of 11.18%. The FT-Transformer outperformed the Tab-Transformer in predicting physicochemical and digestibility properties, indicating the effectiveness of transformer-based deep learning for modeling complex structure–property relationships in starch modification processes.