A machine learning approach to predicting thermal conductivity in nanoparticle-enhanced phase change materials

Thermal conductivity is a critical property of nanoparticle-enhanced phase change materials (NEPCMs), directly influencing their performance in thermal management applications. This study presents the development of an artificial neural network ANN model to accurately predict the thermal conductivity of NEPCMs, showcasing the potential of machine learning in advancing materials science. A dataset comprising 824 experimental samples from 21 relevant studies was compiled, covering a diverse range of thermophysical properties, including the thermal conductivities of nanoparticles and PCMs (W/m·K), nanoparticle concentration (wt%), the temperature of NEPCM (°C), the phase of PCM (solid or liquid), and the size of nanoparticles (nm). The model’s output was the predicted thermal conductivity of the NEPCM (W/m·K). To optimize predictive accuracy, the ANN model was fine-tuned for hyperparameters using RandomizedSearchCV, resulting in an optimized multi-layer feedforward architecture with two hidden layers, trained using the ADAM backpropagation algorithm. The ANN demonstrated outstanding performance, achieving a correlation coefficient ( R ) of 0.983586 and a mean squared error (MSE) of 0.019917 on the validation set, indicating minimal prediction error and strong agreement with actual thermal conductivity values. A sensitivity analysis using the SHapley Additive exPlanations (SHAP) framework revealed that the PCM phase had the most significant influence on thermal conductivity, followed closely by the thermal conductivity of the nanoparticles. These findings highlight the ANN model’s robustness and practical applicability in accurately predicting the thermal conductivity of various NEPCM compositions.