MeltNet: Predicting alloy melting temperature by machine learning

Abstract

Thermodynamics is fundamental for understanding and synthesizing multi-component materials, while efficient and accurate prediction of it still remain urgent and challenging. As a demonstration of the "Divide and conquer" strategy decomposing a phase diagram into different learnable features, quantitative prediction of melting temperature of binary alloys is made by constructing the machine learning (ML) model "MeltNet" in the present work. The influences of model hyperparameters on the prediction accuracy is systematically studied, and the optimal hyperparameters are obtained by Bayesian optimization. A comprehensive error analysis is made on various aspects including training duration, chemistry and input features. It is found that except a few discrepancies mainly caused by less satisfactory treatment of metalloid/semimetal elements and large melting point difference with poor liquid mixing ability between constituent elements, MeltNet achieves overall success in prediction, especially capturing subtle composition-dependent features in the unseen chemical systems for the first time. The reliability, robustness and accuracy of MeltNet is further largely boosted by introducing the ensemble method with uncertainty quantification. Based on the state-of-the-art underlying techniques, MeltNet achieves a prediction mean average error (MAE) as low as about 120 K, at a minimal computational cost. We believe the present work has a general value for significant acceleration of predicting thermodynamics of complicated multi-component systems.

Machine Learning-Guided Exploration of Glass-Forming Ability in Multicomponent Alloys

“…We opted for a brute force search method to tune the hyperparameters over other possibilities, such as Bayesian optimization, due to the relatively small model size and hyperparameter search space. 38 The details of the hyperparameter explorations are provided in the supplementary Tables S1 and S2. Firstly, the values of all the 13 features were normalized and served as the values of input layer neurons.…”

Differentiable thermodynamic modeling

“…For this end, there have emerged many encouraging progresses in recent years. One notable direction is the usage of machine learning (ML) techniques [1][2][3][4][5][6][7][8][9][10][11][12][13] , the generalizability of which allows predictions based on limited amount of data. So far, the methods along this direction can be classified into two types based on the types of training data and predicted quantities of the ML models.…”

“…So far, the methods along this direction can be classified into two types based on the types of training data and predicted quantities of the ML models. The type-I models are trained on and predict thermochemical quantities [1][2][3][4][5][6][7][8] , while the type-II models are trained on and predict phase equilibria [9][10][11][12][13] . Naturally, one may ask if there can be a cross-type learning, i.e., learning thermochemical quantities from phase equilibria (or learning phase equilibria from thermochemical quantities, which is essentially within type-I, since the former can be derived if the latter are fully determined, thus this type is ignored here).…”