An interpretable machine learning approach for process optimization and characterization of key model-derived parameter effects in sunflower oil diacylglycerol synthesis

Diacylglycerol (DAG) has potential metabolic and cardiovascular health benefits. An interpretable machine learning framework was developed to predict and optimize enzymatic DAG synthesis from sunflower oil. Among 16 model configurations, XGBoost-GS was numerically top-ranked under repeated exact-condition-grouped cross-validation (overall R² = 0.910, RMSE = 2.942 pp., and MAE = 2.175 pp). Through process parameter optimization using differential evolution, the optimal conditions were determined to be 11.53 h, 72.11 °C, 8.29 wt% lipase loading, and a weight ratio of 8.93, with a predicted relative DAG content of 54.22% (ensemble predictive dispersion: 1.06%). Experimental verification at the optimization-selected condition yielded 54.68 ± 0.20% relative DAG content. Interpretation of the final fitted model indicated that time had the highest permutation importance, temperature had the largest averaged perturbation sensitivity, and the time-weight ratio pair had the largest positive pairwise interaction metric. This framework supported model-assisted process optimization within the investigated domain.