Multi-objective optimization of low moisture food extrusion processing through active learning and robotics

Abstract

As low-moisture extrusion processing is very complex, especially due to the high number of process variables and their strong interdependence, experimental approaches in product development typically involve numerous iterations accompanied by off-line product testing. These processes are resource-intensive, time-consuming, and require expert knowledge. To overcome these limitations, this study presents a closed-loop framework that links automated product characterization with multi-objective optimization to configure the extruder’s operating variables for achieving specific product characteristics. For this purpose, an on-line automated analytical system based on gravimetric and visual techniques was developed, with results directly fed into the Thompson Sampling Efficient Multi-Objective Optimization (TSEMO) algorithm. The process parameters to be optimized were the barrel zone temperatures, screw and cutter speed, total feed moisture and the feed rate. Objectives for the total throughput, bulk density, shape and expansion ratio of the extrudates were pre-defined. The results of this study demonstrate an efficient approximation of those target properties within 15 iterations, while identifying optimal extrusion settings in a high-dimensional process space. This approach highlights the potential of integrating automation and active learning algorithms for the optimization of low moisture extrusion processes and offers a promising tool to accelerate the process development of directly expanded food products.