Self-pressurized Rapid Visco Analyzer-integrated machine learning for predicting high-moisture extrusion responses and texture of plant protein-starch blends

Modelling of high-moisture extrusion (HME) process-product relationships is challenging due to the non-linear behavior of biopolymer systems. This highlights the need for practical, data-driven tools applicable in industrial HME. For this we present a preliminary feasibility study of a machine learning (ML) approach that integrates pasting/gelling data from pre-extrusion plant protein-starch blends, obtained using a self-pressurized Rapid Visco Analyzer (RVA) at 135 °C (resembling barrel temperatures during HME). We seek to predict the instrumental texture of extrudates and relate these predictions to HME process behavior. HME trials were conducted using blends of protein concentrates (pea, soy, and hemp) and starches (maize starches with different amylose contents and potato starch) at varying barrel moisture contents (50–70%) and screw speeds (350–450 rpm), yielding a total of 158 extrudates. The dataset integrated in line extrusion setpoint data with RVA profiles measured at two solids concentrations, together with HME responses [specific mechanical energy (SME) and die pressure] and extrudate instrumental texture. XGBoost and Ridge regression, showed varying results for the blends tested. When using a leave-one-blend out validation, R 2 values were high for some blends and different dependent variables. However, in some cases large R 2 variations were observed across blends with hemp systems being unpredictable. Analysis of variables highlighted the contributions of RVA and in line extrusion setpoints. Shapley Additive Explanations and standardized coefficients were compared and revealed consistent variable contributions across models supporting interpretability. Findings highlighted the potential of RVA-informed ML modelling for anticipating HME process and product outcomes.