Machine learning models for predicting thermal conductivity of nano‑enhanced phase-change materials

Abstract

The low thermal conductivities of phase-change materials (PCMs) remain a major obstacle to their efficient application in thermal energy storage. Nano‑enhanced phase-change materials (NEPCMs) present a promising alternative; therefore, accurate prediction of their thermal conductivities is of great importance. In this work, six machine learning (ML) models were developed to predict the effective thermal conductivities of NEPCMs: multilayer perceptron neural network (MLPNN), radial basis function neural network, least squares support vector machine, decision tree, k‑nearest neighbors, and dendrite morphological neural network models. A dataset comprising 324 data samples with six input parameters, the temperature, nanoparticle concentration, nanoparticle size, nanoparticle thermal conductivity, PCM thermal conductivity, and PCM phase state, was employed. Based on statistical and graphical approaches, the performances of the ML models and five classical empirical correlations were evaluated and compared. Physical trend analysis was conducted to further verify the ML model performances. Among the models compared in this study, the MLPNN model achieved the best overall performance on the compiled dataset (coefficient of determination ( R 2 ) = 0.9924, root mean square error = 0.0476, and mean absolute percentage error = 3.10%), performs better than the other ML models and classical empirical correlations under the conditions examined. The MLPNN successfully captured the expected thermal conductivity variations with respect to each input parameter and performed adequately across most input ranges, with only slight increases in the error for high concentrations, extreme temperatures, and specific PCM types. This work provides a systematic benchmark for the selection of appropriate ML models and demonstrates the superiority of data‑driven approaches in predicting the thermal conductivities of NEPCMs.