Crystal graph attention networks for the prediction of stable materials

Abstract

Graph neural networks for crystal structures typically use the atomic positions and the atomic species as input. Unfortunately, this information is not available when predicting new materials, for which the precise geometrical information is unknown. We circumvent this problem by replacing the precise bond distances with embeddings of graph distances. This allows our networks to be applied directly in high-throughput studies based on both composition and crystal structure prototype without using relaxed structures as input. To train these networks, we curate a dataset of over 2 million density functional calculations of crystals with consistent calculation parameters. We apply the resulting model to the high-throughput search of 15 million tetragonal perovskites of composition ABCD 2 . As a result, we identify several thousand potentially stable compounds and demonstrate that transfer learning from the newly curated dataset reduces the required training data by 50%.

A dataset of 175k stable and metastable materials calculated with the PBEsol and SCAN functionals

“…Our starting point was the dataset used in the machine learning study of ref. 21 . This included PBE calculations stemming from the Materials Project database 2 , AFLOW 1 , and our own calculations.…”

“…In a previous work 21 we combined data from the AFLOW database 1 , the Materials Project 2 and from our own group to create a rather complete convex hull of thermodynamic stability at the PBE level. The details of the selection of the dataset can be found in ref.…”

“…The details of the selection of the dataset can be found in ref. 21 . Specifically, we selected all materials that were calculated with the same functional, pseudopotential, as well as U -parameters used in the Materials Project.…”

Formula Graph Self‐Attention Network for Representation‐Domain Independent Materials Discovery

“…[38] The attention mechanism has been adapted in several ML architectures for materials property prediction with improved accuracy. [28,30,33,37,[39][40][41][42] We show that Finder can outperform some state-of-the-art stoichiometry-only models such as Roost and compete with crystal graph models such as MEGNet and CGCNN on diverse benchmark databases curated from the Materials Project (MP) repository. Compared to other models revisited in this work, our model displays faster convergence and achieves lower errors at all training set sizes explored.…”

Machine learning–enabled high-entropy alloy discovery

“…We based our approach on the use of machine learning (ML) techniques, with a focus on probabilistic models and artificial neural networks. Limited by the amount of available composition-property data, conventional ML approaches in alloy design have to predominantly rely on simulation data, often with only limited experimental validation ( 9 , 10 ). As the experimental microstructure database continues to expand, ML obtains higher accuracy in predicting the phase or microstructure of materials ( 11 ).…”