Analysis of Machine Learning Algorithms for the Computer Simulation of Moisture Sorption Isotherms of Coffee Beans

Abstract

Digital twin–based machine learning (ML) techniques can improve the control of the storage conditions of dried products, strengthening the classical water sorption isotherm–based approach by including additional process variables. In this study, water sorption isotherms of dried parchment and green coffee beans were experimentally determined at 25, 35, and 45 °C using the dynamic dew point (DDI) method. Experimental data (both coffee bean types and temperatures) were simultaneously modeled by means of three ML techniques, support vector machine (SVM), random forest (RF), and artificial neural networks (ANN), with 75% of data used for model training and 25% for validation. The hyperparameters were identified by minimizing the mean square error (MSE). The ML model’s accuracy was addressed by a multiway ANOVA on the mean relative error (MRE), the coefficient of determination (R²), and the computation time (CT). The sorption isotherms were significantly (p-value < 0.05) affected by the type of coffee and the temperature. The SVM model provided the best fit (MRE < 1% and R² > 99%) in a reasonable CT (< 13 s). These results revealed the potential of ML models as a robust tool for the fast prediction of the equilibrium moisture content, including additional variables such as the type of coffee stage (dried parchment or green) and temperature; this paves the way for their industrial-level implementation to assist storage management.