Enhancing materials property prediction by leveraging computational and experimental data using deep transfer learning

Dipendra Jha,Wei-Keng Liao,Kamal Choudhary,Ankit Agrawal,Carelyn Campbell,Francesca Tavazza,Alok Choudhary

doi:10.1038/s41467-019-13297-w

Abstract

The current predictive modeling techniques applied to Density Functional Theory (DFT) computations have helped accelerate the process of materials discovery by providing significantly faster methods to scan materials candidates, thereby reducing the search space for future DFT computations and experiments. However, in addition to prediction error against DFT-computed properties, such predictive models also inherit the DFT-computation discrepancies against experimentally measured properties. To address this challenge, we demonstrate that using deep transfer learning, existing large DFT-computational data sets (such as the Open Quantum Materials Database (OQMD)) can be leveraged together with other smaller DFT-computed data sets as well as available experimental observations to build robust prediction models. We build a highly accurate model for predicting formation energy of materials from their compositions; using an experimental data set of 1,643 observations, the proposed approach yields a mean absolute error (MAE) of 0.07 eV/atom, which is significantly better than existing machine learning (ML) prediction modeling based on DFT computations and is comparable to the MAE of DFT-computation itself.

Highlights

Our results demonstrate a significant benefit from the use of deep transfer learning; in particular, the proposed approach enables us to achieve an mean absolute error (MAE) of 0:07 eV/atom against an experimental data set containing 1,643 observations, which is significantly better than the mean absolute discrepancy of ~0:1 eV/atom of the DFTcomputational data sets compared against experiments, and MAE of ~0:15 eV/atom of the predictive models trained from scratch on either experimental data set or Density Functional Theory (DFT)-computed data sets
For the experimental data set, we use the experimental formation energy from the SGTE Solid SUBstance (SSUB) database; they are collected by international scientists[59] and contain a single value of the experimental formation enthalpy, which should represent the average of formation enthalpy observed during multiple experiments, and do not contain error bars
In this work, we demonstrated the benefit of leveraging both DFT computations and experimental observations to build more robust prediction models whose predictions are closer to the experimental observations compared with the predictive models built using only DFT-computed data sets

Summary

Introduction

(ML) techniques to accelerate the discovery/design of new materials with select engineering properties[21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45] Such predictive models enable reducing the size of the search space for material candidates and help in prioritizing which DFT simulations and, possibly, experiments, to perform. Kirklin et al compared the DFT-computed formation energy with experimental measurements of 1670 materials and found the mean absolute error (MAE) to vary from 0:096 to 0:136 eV/atom for OQMD10. DFT databases, such as OQMD and the Materials Project, reduce this systematic error by chemical potential fitting a 0.50

Methods

Results

Conclusion