Large databases of crystalline materials and associated properties, typically computed using density functional theory (DFT), have become widely available and used for training machine learning (ML) models. Thermodynamic stability, band gap, and elastic mechanical properties are now available in large volumes with sufficient accuracy to train effective models. Magnetic properties, however, require special attention to model using DFT and are only currently available in datasets that are orders of magnitude smaller in size. ML models can therefore be used to screen candidate materials for magnetic properties if their effectiveness can be demonstrated in lower data regimes. In this work, we compare the ability of multiple model architectures to predict saturation polarization, testing a prevailing assumption that knowledge of crystal structure is critical to predict magnetic properties. While a prototypical crystal-structure-based model (CGCNN) did on average outperform a model based on composition only [Representation learning from Stoichiometry (RooSt)], we found that RooSt performed nearly as well. For example, in predicting saturation polarization, we observed a mean absolute error from RooSt models (0.15 T) only worse than from CGCNN (0.13 T): still sufficiently low for useful screening and with the significant benefit of not requiring crystal structure to be known a priori. Further, we found that models trained on DFT-computed values of saturation polarization are also usefully predictive of experimentally tested values, as confirmed in samples fabricated by arc melting. This suggests composition-based models can enable screening of novel candidate material spaces without prior knowledge of crystal structure, which would drastically accelerate screening of materials with promising magnetic properties.

Close

PrevNext

View figure in article

• Magnetism
• Convolutional neural networks
• Magnetization measurements