Materials Property Prediction for Limited Datasets

The combined progress in first-principles simulation codes and supercomputing capabilities has given birth to the so-called high-throughput ab initio approach. This has generated large amounts of materials data (mainly ground-state properties) which has been stored in various open databases. However, for more complex properties (e.g., linear or higher-order responses), this approach is still out of reach because of the required CPU time. To overcome this limitation, artificial intelligence has recently attracted much attention. However, in order to make accurate predictions, current machine-learning approaches generally require large amounts of data, which are precisely not available for complex properties. In this talk, I will introduce the MODNet framework which relies on a feedforward neural network and the selection of physically meaningful features. Next to being faster in terms of training time, this approach is shown to outperform current machine-learning models on small datasets. In particular, the vibrational entropy at 305 K for crystals is predicted with a mean absolute test error of 0.009 meV/K/atom (four times lower than previous studies). Furthermore, joint learning reduces the test error compared to single-target learning and enables the prediction of multiple properties at once, such as temperature functions. Finally, the selection algorithm highlights the most important features and thus helps to understand the underlying physics.