Dresden 2020 – scientific programme
The DPG Spring Meeting in Dresden had to be cancelled! Read more ...
Parts | Days | Selection | Search | Updates | Downloads | Help
MM: Fachverband Metall- und Materialphysik
MM 35: Topical Session: Data Driven Materials Science - Descriptors (joint session MM/CPP)
MM 35.4: Talk
Wednesday, March 18, 2020, 12:30–12:45, BAR 205
Hierarchical SISSO: predicting complex materials properties building on simpler ones — •Lucas Foppa1, Sergey V. Levchenko2,1, Matthias Scheffler1, and Luca M. Ghiringhelli1 — 1Fritz-Haber-Institut der MPG, Berlin, DE — 2Skolkovo Institute of Science and Technology, Moscow, RU
Symbolic regression is a promising tool to identify analytical models (descriptors) for predicting materials properties that are otherwise accessed via rather expensive ab initio calculations. In this context, the sure-independence screening and sparsifying operator (SISSO),[1] which combines the systematic generation of large feature spaces with compressed sensing, has been successfully applied, e.g., to the prediction of the (meta)stability of binary systems and perovskites from atomic properties only. However, if the relationship between the features and the target property is too complex, the descriptor search can become very inefficient. Here, we tackle this issue via a hierarchical approach: features that are easily computed (e.g., atomic properties) are used for predicting simple properties (e.g., lattice constant) and the resulting descriptors are in turn used as candidate features for modeling more complex properties (e.g., bulk modulus, position of band centers or band gaps). We demonstrate the hierarchical approach by analyzing a dataset of >700 cubic simple (ABO3) and double (A2BB′O6) perovskites for predicting mechanical and electronic properties. The learned models require only atomic features as inputs and are therefore suitable for high-throughput screening of such materials.
[1] R. Ouyang, et al., Phys. Rev. Mater. 2, 083802 (2018).