Prediction of Internal Moisture in Post-Harvest Cocoa (Theobroma Cacao) Using a Machine Learning System, Applicable to Farms That Use Solar Drying

This article explores the potential for predicting moisture content in cocoa post-harvest using a machine learning system that incorporates experimental data and accounts for varying temperature conditions. Key postharvest processes, such as fermentation and drying, significantly influence the quality of the final product, but humid and rainy climates in cocoa-growing regions complicate these processes, necessitating improved technologies. The methodology involves developing a prediction engine utilizing a Multilayer Perceptron (MLP) with MLPClassifier and MLPRegressor classes, employing experimental field data and two data-generating models: an automatic moving boundary model for isothermal food dehydration and the Brusselator model for autocatalytic reactions. A total of approximately 2,500 experimental data points were collected from six tests conducted between June and December 2023. The neural network was designed to accept a convergence interval of experimental results for 70 % of the training data, with the remaining $\mathbf{3 0 \%}$ used for verification. The prediction engine achieved an accuracy range of 78 % to 82 %. The implementation of Adrover's isothermal dehydration model proved beneficial, preventing non-convergence and showcasing effective convergence-accelerating algorithms. The system can make predictions for up to $\mathbf{4 2 0}$ hours based on experimental results. Ultimately, the study presents a predictor for postharvest cocoa conditions using an artificial neural network that effectively accounts for temperature variability, making it suitable for artisanal cocoa processing. However, it emphasizes the need for additional data to improve mean squared error (MSE) and $\mathbf{R}^{2}$ values, and suggests future research should include genetic diversity of cocoa lots to enhance prediction accuracy.