Source-linked AI summary
How to represent crystal structures for machine learning: towards fast prediction of electronic properties
K. T. Schütt, H. Glawe, F. Brockherde, A. Sanna, K. R. Müller, E. K. U. Gross
TL;DR
High-throughput electronic-structure calculations are too costly for systematic materials exploration, motivating fast machine-learning prediction of solid-state properties. The paper develops crystal representations and learning methods for periodic systems, showing meaningful DOS-at-the-Fermi-energy prediction while identifying representation and training-set constraints.
Problem
High-throughput KS-DFT calculations become prohibitively costly as materials and unit-cell complexity grow, while periodic crystals lack a unique learning representation.
Method
The authors train a DOS-at-the-Fermi-energy predictor on LSDA calculations using periodic crystal features based on partial radial distribution functions and kernel ridge regression with nested cross-validation.
Results
PRDF features consistently outperform the Bravais-matrix-plus-Coulomb-matrix description, and spd DOS-at-the-Fermi-energy predictions achieve an average error smaller than 6% of the DOS value range.
Takeaways & Limitations
The PRDF representation supports fast prediction for crystals with arbitrary numbers of atoms per unit cell, including cases where conventional DFT calculations would be prohibitive.
Takeaways & Limitations
Higher prediction accuracy requires larger training sets, potentially reaching the limits of present computing facilities, and Coulomb matrices are unsuitable for periodic crystals.
Abstract
from arXiv · showhide
High-throughput density-functional calculations of solids are extremely time consuming. As an alternative, we here propose a machine learning approach for the fast prediction of solid-state properties. To achieve this, LSDA calculations are used as training set. We focus on predicting metallic vs. insulating behavior, and on predicting the value of the density of electronic states at the Fermi energy. We find that conventional representations of the input data, such as the Coulomb matrix, are not suitable for the training of learning machines in the case of periodic solids. We propose a novel crystal structure representation for which learning and competitive prediction accuracies become possible within an unrestricted class of spd systems. Due to magnetic phenomena learning on d systems is found more difficult than in pure sp systems.