Machine learned features from density of states for accurate adsorption energy prediction

Victor Fung,P Ganesh,Guoxiang Hu,Bobby G Sumpter

doi:10.1038/s41467-020-20342-6

Victor Fung, P Ganesh + Show 2 more

Open Access

https://doi.org/10.1038/s41467-020-20342-6

Copy DOI

Abstract

Materials databases generated by high-throughput computational screening, typically using density functional theory (DFT), have become valuable resources for discovering new heterogeneous catalysts, though the computational cost associated with generating them presents a crucial roadblock. Hence there is a significant demand for developing descriptors or features, in lieu of DFT, to accurately predict catalytic properties, such as adsorption energies. Here, we demonstrate an approach to predict energies using a convolutional neural network-based machine learning model to automatically obtain key features from the electronic density of states (DOS). The model, DOSnet, is evaluated for a diverse set of adsorbates and surfaces, yielding a mean absolute error on the order of 0.1 eV. In addition, DOSnet can provide physically meaningful predictions and insights by predicting responses to external perturbations to the electronic structure without additional DFT calculations, paving the way for the accelerated discovery of materials and catalysts by exploration of the electronic space.

Highlights

Materials databases generated by high-throughput computational screening, typically using density functional theory (DFT), have become valuable resources for discovering new heterogeneous catalysts, though the computational cost associated with generating them presents a crucial roadblock
We develop a machine learning (ML) model, DOSnet, which takes the density of states (DOS) directly as the input and extracts these features using convolutional neural networks (CNNs) as part of the training process (Fig. 1)
The resolution of the DOS used in this work is 0.01 eV, though we find a similar performance even as a lower resolution is used, which can tuned in DOSNet by an initial average pooling layer

Summary

Introduction

Materials databases generated by high-throughput computational screening, typically using density functional theory (DFT), have become valuable resources for discovering new heterogeneous catalysts, though the computational cost associated with generating them presents a crucial roadblock. We note that while there are certainly exceptions to this trend with respect to the number of training data required, this general order of features usually holds for problems of the same scale and complexity This is likely due to structural features being more physically removed from the property being predicted (i.e., adsorption energy) than electronic features, and require more training data to learn their relationships. We focus on the electronic features for adsorption energy prediction, as it generally offers a good compromise between training and screening cost and requires a training data set size which is commonly obtainable in current highthroughput studies (~102–103 entries) It is well-known that the electronic structure of a surface is closely linked to its surface chemistry, many parallels to frontier molecular orbital theory have been made[43,44]. These pre-existing features must be discovered or selected prior to the study, which may not always be possible when moving to unexplored chemical spaces

Objectives

Methods

Results

Conclusion