Can Machine Learning Predict the Liquidus Temperature of Binary Alloys?

Despite significant efforts in developing model descriptions of alloys mixing behavior, the liquidus temperature of an alloy is generally not predictable through theoretical models but instead determined experimentally. Here we explore if machine learning strategies can be used to predict them. We use random forest and consider various representations of the alloy through features vectors based on a prior known information. We found that when features based on physical insights into alloys’ mixing are used, an average prediction with 8% error can be achieved compare to 13% when only using the properties of A and B elements as features. Such error is essentially identically to a linear extrapolation of known melting temperatures of A and B to predict the alloy’s liquidus temperature. The poor predictability even under the best circumstances is most dramatically reflected in the fact that even when over 99.8% of all data considered for training of the algorithm, the error of prediction into the remaining 0.2% is only 8%. Our analysis reveals that the major challenges in predicting the liquidus temperature through ML algorithms originates from the challenge to represent the relevant characteristics of an alloy through which we argue is a common challenge in complex materials science problems. Further, the discrete nature of atoms and their corresponding features, constitutes the most fundamental challenge in applying machine learning strategies for complex materials science problems.