A critical examination of robustness and generalizability of machine learning prediction of materials properties

Recent advances in machine learning (ML) have led to substantial performance improvement in material database benchmarks, but an excellent benchmark score may not imply good generalization performance. Here we show that ML models trained on Materials Project 2018 can have severely degraded performance on new compounds in Materials Project 2021 due to the distribution shift. We discuss how to foresee the issue with a few simple tools. Firstly, the uniform manifold approximation and projection (UMAP) can be used to investigate the relation between the training and test data within the feature space. Secondly, the disagreement between multiple ML models on the test data can illuminate out-of-distribution samples. We demonstrate that the UMAP-guided and query by committee acquisition strategies can greatly improve prediction accuracy by adding only 1% of the test data. We believe this work provides valuable insights for building databases and models that enable better robustness and generalizability.