Improving Nutrition and Sensory Goals: Utilizing Explainable Machine Learning and Multi-Objective Optimization to Optimize Quality of Osmanthus fragrans Extract

Abstract

Osmanthus fragrans extract (OFE) has significant potential for application in the beverage and cosmetic industries. However, the conventional extraction processes of OFE are affected by multiple factors, making it challenging to identify parameters that can be synergistically optimized for diverse qualities. This study aimed to establish a multi-objective optimization (MOO) method for the ultrasonic-assisted extraction of O. fragrans var. thunbergii. The effects of time, temperature, ultrasonic power, and solid-liquid ratio on yield, content of total phenylethanol glycosides, verbascoside and salidroside, and colorimetric parameters were compared using response surface methodology and machine learning algorithms. Multi-layer perceptron, XGBoost, and support vector regression were evaluated as base models for constructing an ensemble model, using the SHAP algorithm for interpretation. A MOO method based on BP-ANN and NSGA-II was developed to maximize the content of bioactive compounds in the extract while considering other quality indicators. The optimal conditions for the target quality attributes were determined to be a temperature of 54°C, an extraction time of 52 min, a solid-liquid ratio of 16 mL/g, and an ultrasonic power of 345 W. This study provides a novel approach that combines artificial intelligence and MOO for process optimization in the plant extraction industry.