An efficient rule-based screening approach for discovering fast lithium ion conductors using density functional theory and artificial neural networks

The density functional theory (DFT) method is a widely used tool that can guide targeted searches exploring numerous possible chemistries according to any property of interest. However, acquiring accurate DFT results from a large chemical search space is still a major challenge for computationalists because it requires considerable time and resources. Therefore, advances in this field are urgently needed. In particular, the development of new materials for Li ion batteries would benefit greatly because of the increasing demand for power, energy density, stability during operation, and safety. We have previously demonstrated the use of multivariate partial least squares (PLS) regression to augment DFT calculations and accelerate material screening. However, the linear function scheme in the PLS method does not accurately reproduce the material values near the extreme ends of the attribute dataset. In this study, we examined neural network (NN) modeling, which offers a more flexible framework, to improve the accuracy of the values. We used the compositional space for LiMXO4 (M - main group elements, X - group XIV and group XV), which is a candidate solid electrolyte material. The evolved NN models were substantially more accurate and could generalize better than the PLS models for values inside and outside the dataset that they were trained on. We also explored NN modeling with common literature data as descriptors for target properties and found that the predictive capability was comparable with that of DFT data-based modeling. The relevance of the input variables was then identified using two of the most common techniques: the causal index method and sensitivity analysis. Our method offers a simple, practical approach for merging theoretical and experimental databases to accelerate the screening of a wider variety of materials.