Graph Neural Networks for Accurate Prediction of Ionic Transport Properties of Solid-State Electrolytes

In the pursuit of energy-efficient and environmentally friendly energy storage devices, solid-state electrolytes (SSE) have emerged as promising candidates due to their substantial stability and performance. The discovery of new materials with enhanced ionic diffusion is essential for the advancement of SSEs. To address this challenge, we introduce a novel approach based on Graph Convolutional Neural Networks (GCNNs) to predict the ionic diffusion coefficient of materials in large datasets such as the Materials Project database.<br/><br/>Our GCNN model is trained on a large and diverse database of density functional theory ab initio MD (DFT-AIMD) simulations [1,2], comprising several families of SSE and tens of millions of atomic configurations, what allows extracting valuable insights from the underlying crystallographic and structural relationships.<br/><br/>The predictive power of our GCNN-based approach is demonstrated through extensive validation on the dataset, showcasing its ability to accurately forecast the electrolyte behavior of both known and yet to discover materials. The integration of GCNNs into the materials discovery pipeline holds great promise for the development of next-generation SSEs technologies, introducing here a framework which allows predicting on any desired dataset.<br/><br/>The scripts resulting from this study have been made publicly available as a Python package, that is user-friendly and easily adaptable [3] to any desired target database.<br/><br/>[1] C. López, A. Emperador, E. Saucedo, et al., Universal ion-transport descriptors and classes of inorganic solid-state electrolytes, Materials Horizons, 2023, doi: 10.1039/D2MH01516A<br/>[2] C. López, A. Emperador, E. Saucedo, et al., DFT-AIMD database, 2023, url: https://superionic.upc.edu<br/>[3] C. López, R. Rurali, C. Cazorla, Repository with all the developed codes, 2023, url: https://github.com/IonRepo/IonPred