High-precision machine learning for predicting latent heat in diverse multicomponent molten salts

Molten salts in phase change materials offer significant advantages, including high thermal storage density, a wide operational temperature range, and low cost. However, the development of novel high-latent-heat molten salts remains largely empirical. Machine learning offers the potential to expedite theoretical advancements and enable precise, cost-efficient performance predictions. Nonetheless, the diversity of molten salt s complicates the accuracy and generalizability of machine learning models. This study proposes a novel latent heat prediction methodology that integrates data analysis and machine learning. A comprehensive dataset encompassing various inorganic salts was systematically analyzed to extract key features influencing latent heat. Subsequently, a predictive model was constructed by combining a backpropagation neural network (BPNN) with particle swarm optimization (PSO). The PSO-BPNN model demonstrated high predictive accuracy, achieving R2 values of 0.9389 and 0.9413 for binary and ternary molten salts, respectively, with experimental validation indicating prediction errors within 10 %. This approach establishes a high-precision, scalable framework for predicting the latent heat of multicomponent molten salts, thereby advancing the design of salts with tailored thermal properties and offering a valuable reference for predicting other thermophysical characteristics.