Understanding wheat dough behavior through prediction of quality profiles with explainable Machine Learning

While technological measurements help anticipate dough behavior during breadmaking, expert sensory evaluations remain the main quality indicator. Due to their subjective nature, predicting these test outcomes is challenging, but recent advances have simplified this by identifying a limited number of consistent data patterns known as Quality profiles (QPs). This study explores the prediction of wheat QPs using Machine Learning based on standard technological measurements. The goal is both practical (to assist in decision-making) and analytical (to identify the most relevant indicators and their optimal value ranges). A dataset of 287 wheats samples was analyzed, using 14 commonly used parameters (eg. protein content, damaged starch, hardness, falling number, gluten index, farinograph® and alveograph® parameters). Among tested models, the Support Vector Machine model (SVM) delivered the best results. To interpret model outputs, the SHapley Additive exPlanation (SHAP) approach was used, revealing each variable’s influence. The Alveograph Elasticity Index (Ie) parameter was the most informative for predicting extreme QPs. However, for Ie values between 40–60%, results were less clear, suggesting the need for additional indicators and further research. • Quality profiles summarize baking test results • UMAP analysis guarantees the genericity of the training data set • ML algorithms predict quality profiles from standard technological measurements • Index of elasticity from Alveograph® is the best predictor of Quality profiles • Explainable AI techniques allow to determine the range of interpretable values