Integrated machine learning assisted multi-objective optimization of spray drying conditions for plant-based probiotic powder: Comparative modelling with GC-MS and FTIR validation

This study established an integrated strategy that combines spray drying, machine-learning-based prediction, and multi-objective optimization to enhance the functional and physicochemical quality of a plant-based probiotic powder. A Box-Behnken design was used to evaluate the effects of maltodextrin concentration, gum Arabic concentration, inlet temperature, and feed flow rate on seven key responses. Predictive performance was compared among least squares support vector machines (LSSVM), artificial neural networks (ANN), and random forest (RF), using the conventional quadratic response surface model as a reference. Compared with RSM, LSSVM and ANN generally showed lower prediction errors for several key responses, whereas RF displayed weaker robustness across some outputs. ANN remained competitive for several indicators, but showed slightly lower predictive stability and a limited tendency toward overfitting for some responses. The three machine learning models were further integrated with NSGA-II and TOPSIS to identify compromise optimal spray-drying conditions. Experimental validation focused on the LSSVM-NSGA-II and ANN-NSGA-II optima. The LSSVM-NSGA-II optimum, corresponding to 16.42% maltodextrin, 1.33% gum Arabic, 121.74 °C inlet temperature, and 225.0 mL/h feed flow rate, showed consistently low relative errors, indicating strong predictive reliability. In contrast, the ANN-NSGA-II optimum, obtained at 22.31% maltodextrin, 2.37% gum Arabic, 158.63 °C, and 280.2 mL/h feed flow rate, exhibited higher relative errors, particularly for probiotic viability, moisture content, and zeta potential. GC-MS and FTIR analyses further showed that the LSSVM-derived optimum better preserved thermally sensitive alcohols and structural features, whereas the ANN-derived optimum promoted the formation of acids, carbonyl compounds, and Maillard-related volatiles.

Text 125901.pdf - Published Version Restricted to Repository staff only Download (9MB) Official URL or Download Paper: https://linkinghub.elsevier.com/retrieve/pii/S2212...

Actions (login required)

View Item