Integrating artificial Intelligence-Based metaheuristic optimization with Machine learning to enhance Nanomaterial-Containing latent heat thermal energy storage systems

• A database of 22 laboratory studies is compiled to predict the latent heat of NePCMs. • Machine learning techniques are compared in terms of accuracy and calculation costs. • Hyper-parameters of all techniques are optimized to achieve maximum efficiency. • GA and PSO showed a significant improvement in the prediction performance. • The GA/PSO-based optimized multilayer perceptron ANN provided a reliable model. Progress in artificial intelligence and machine learning has significantly improved the capability to accurately predict the properties of nano-enhanced phase change materials (NePCMs). Machine learning can be a reliable solution to avoid the cost and time of repetitive tests to measure thermophysical characteristics. The present study proposes a novel strategy to provide optimal data-driven models for predicting the latent heat of NePCMs. In this regard, a database containing 22 laboratory studies with five input variables, including nanomaterials characteristics (mass fraction, density, and average diameter) and PCM properties (latent heat and density) is used. The proposed strategy combines machine learning algorithms, including multilayer perceptron neural network (MLPNN), generalized additive model (GAM), Gaussian kernel regression (GKR), support vector machine (SVM), and Gaussian process regression (GPR) with artificial intelligence-based metaheuristic optimization algorithms (PSO and GA) to optimize their structural/training parameters. Statistical criteria, margin of deviation plots, violin graphs, and external experimental datasets are utilized to assess the developed optimized models. Results showed a noticeable difference in the effectiveness of the GA and PSO algorithms in optimizing the structural and training parameters of various machine learning approaches. Both GA and PSO provided the same parameters for the MLPNN model, with an R-value of 0.9558 and a mean absolute percentage error of 3.7%, indicating the best accuracy among all models. The optimized models based on GAM and GPR were classified in the second accuracy level. On the other hand, the models generated using SVM and GKR were found to be the least accurate.