Inverse design of experimentally synthesizable crystal structures by leveraging computational and experimental data

Abstract

Crystal structure prediction (CSP) drives the discovery and design of innovative materials. However, existing CSP methods rely heavily on formation enthalpies calculated by density functional theory (DFT) and ignore the differences between DFT and experimental values, resulting in predicted structures that may be limited in experimental synthesis. To overcome these limitations, a novel collaborative approach was proposed for CSP that utilizes advanced deep learning models and optimization algorithms combined with experimental data. The formation enthalpies obtained from our approach are highly consistent with the actual experimental observations by transfer learning from the experimental data. By incorporating the experimentally synthesizable information of crystals, our approach is able to inverse design crystal structures that can be synthesized experimentally. Applying the model to 17 representative compounds, the results indicate that the approach can accurately identify experimentally synthesized structures with high precision. Moreover, the obtained formation enthalpies and lattice constants closely align with experimental values, underscoring the approach's effectiveness. The synergistic approach bridges the longstanding disparities between theoretical predictions and experimental results, providing guidance for experimentally synthesizing new materials and alleviating the demand for extensive and costly experimental trials.