Machine learning-based predictive modeling of transient temperature distribution in nanoenhanced phase change material during solidification

Accurately predicting the transient temperature distribution (TTD) of OM08 fatty acid, a phase change material (PCM), during solidification is critical for maximizing the efficiency of thermal energy storage devices. Twelve machine learning (ML) techniques employing linear (LASSO, Ridge, and Elastic Net), instance-based (K-Nearest Neighbor), Kernel-based (Support Vector Regression), tree-based (Decision Tree), ensemble (Random Forest and Extra Trees), boosting (Gradient Boosting Regressor, Extreme Gradient Boosting, and Light Gradient Boosting Machine), and artificial neural network (Multilayer Perceptron) techniques were developed for predicting the TTD of OM08-PCM in a spherical shell during solidification. The predictive accuracy of these ML algorithms was comprehensively compared using the root mean square error and coefficient of determination ( R 2 ) as metrics, in conjunction with the error distribution analysis, residual analysis, and computational costs, such as the training time and memory consumption. Among the ML models, the Extra Tree algorithm afforded the best performance, characterized by the lowest root mean square error (≈0.0056), and highest R 2 (≈0.9987). Error and residual analyses further confirmed that the Extra Tree method is robust, with minimal bias and variance. Moreover, the Extra Trees approach offered the best compromise of computational efficiency and performance, making it better than more computationally demanding models such as Extreme Gradient Boosting and Random Forest models. This experiment demonstrated the advantages of ensemble learning in predicting the solidification behavior of a spherically encapsulated PCM.