A common challenge in accelerated material design is to apply machine learning (ML) methods that can handle data with different structures and dimensions, and also provide physical interpretability. Unfortunately, most existing ML methods are ‘black box’ models incapable of providing physical interpretation or dealing with missing dimensions data that are often encountered in materials science. To overcome this challenge, we propose an interpretable and extensible machine learning framework based on thermodynamically informed graphs and deep data mining from graph neural networks. We demonstrate our framework on the problem of predicting the martensite start (Ms) temperature, which depends on various factors (composition, austenite grain size, and outfield conditions). We construct a thermodynamically informed graph that captures the quantitative relationships between these factors and the Ms temperature using limited and incomplete data. The prediction results indicate that our framework provides clear physical insights because the thermodynamic mechanisms are embedded in the thermodynamic representation graph. Our framework has several advantages: 1) it incorporates thermodynamic mechanisms into the graph structure, 2) it can handle missing dimensions data by filling in the gaps with graph information, and 3) it can be easily extended to new features without requiring much additional data for training. Moreover, we derive a general empirical equation for the Ms temperature prediction from the trained graph neural networks for practical applications.
Read full abstract