Drying kinetic for moisture content prediction of peels Tahiti lemon (Citrus latifolia): Approach by machine learning and optimization - genetic algorithms and nonlinear programming

The application of a versatile approach for modeling and prediction the moisture content of dried peels was evaluated using both empirical and semi-empirical equations (Lewis, Page, Henderson and Pabis, Modified Page, Logarithmic, and Modified Logistic) as well as machine learning models (K-nearest neighbor | KNN, Decision Tree | DT, Artificial Neural Network | ANN and Support Vector Regression | SVR). Heuristic optimization methods, including genetic algorithms (GA) and nonlinear programming (NLP), were employed to identify the best empirical and semi-empirical models for estimating moisture content during the drying process of lemon peel layers. The parameters of the drying kinetics models were optimized using GA to achieve the best results. It was found that as the number of model parameters increases, particularly in models such as the logarithmic one, the optimization problem becomes more complex. Consequently, accurate initial guesses become increasingly important, emphasizing the need for heuristic methods like genetic algorithms. This optimization approach provided excellent performance metrics (R² > 0.9715, SSR 〈 0.0625 and MSE < 0.0026 for endocarp and R² 〉 0.9678, SSR < 0.0755 and MSE < 0.0030 for epicarp). The models proposed in this study achieved the best results with the modified logistic equation (R² > 0.9923, MSE 〈 0.0001 and SSR < 0.0013 for endocarp and R² 〉 0.9905, MSE < 0.0001 and SSR < 0.0013 for epicarp). In particular, the multilayer perceptron neural network of the machine learning proved to be the optimal choice as it best accounts for the complexity of the drying kinetics of lemons. This neural network model outperformed traditional empirical and semi-empirical models, demonstrating superior performance metrics (R² > 0.9979, MSE 〈 0.0002 and SSR < 0.0012 for endocarp and R² 〉 0.9989, MSE < 0.0001 and SSR < 0.0008 for epicarp) when tested against validation data.