MACHINE LEARNING APPROACHES FOR TARGETED DOUGH PROFILES IN INDUSTRIAL FLOUR PRODUCTION

Industrial flour production depends on the controlled use of additives or enzymes to achieve consistent performance despite variability in resources. However, formulation decisions are often based on trial-and-error approaches which limit proactive decision-making. This research is guided by a central question: to what extent machine-learning–based, stage-dependent models can support formulation and managerial decision-making by reducing uncertainty in industrial flour production. To address this question, the study pursues three objectives: (i) to assess the predictive capability of machine-learning models for dough behaviour based on flour characteristics and additive usage; (ii) to identify which stages of the dough rheology process are more reliably predictable; and (iii) to evaluate whether a stage-dependent modelling framework can support targeted formulation by informing additive dosing prior to laboratory testing. The research adopts a data-driven approach based on the CRISP-DM methodology, using industrial Mixolab data from a large-scale flour producer. Stage-dependent Random Forest models are developed to reflect the sequential nature of dough transformation and are evaluated using standard predictive metrics. In addition to forward prediction, the models are applied in an inverse setting to explore additive configurations aligned with desired dough profiles. The results show strong predictive performance for most stages, with lower accuracy in enzymatic phases. Importantly, the framework generates realistic additive recommendations, supporting proactive formulation. From a managerial perspective, this enables earlier feasibility assessment, reduces uncertainty, and improves coordination between commercial, R&D, and production functions.