Predicting melting temperature of inorganic crystals via crystal graph neural network enhanced by transfer learning

Abstract

The melting temperature, a crucial material property, is particularly challenging to measure accurately for inorganic crystals. The data-driven approach emerges as a potential solution, although its effectiveness is hindered by the limitation of available data. To counter this challenge, we implement transfer learning, leveraging a vast computational database of atomization energy. We first pre-train a geometric-information-enhanced crystal graph neural network (GeoCGNN) using atomization energies of approximately 36,000 materials that are computed by the density functional theory. Subsequently, the pre-trained model is fine-tuned using melting temperatures measured for 799 crystals, encompassing 83 elements, ranging from unary to quaternary systems. This transfer learning strategy decreases the root mean square error from 407 to 218 K, attesting to a marked improvement in prediction accuracy. Furthermore, transfer learning significantly mitigates error variability across unary, binary, and ternary (or higher-order) systems, thereby enhancing the reliability of predictions across a broader range of crystals. We also show that transfer learning allows effective task adaptation by leveraging representation learned from pre-training. Therefore, it can achieve better prediction performance even with a limited number of data for predicting melting points.