Enhancing crystal structure prediction by combining computational and experimental data via graph networks
Crystal structure prediction (CSP) stands as a powerful tool in materials science, driving the discovery and design of innovative materials. However, existing CSP methods heavily rely on formation enthalpies derived from density functional theory (DFT) calculations, often overlooking differences between DFT and experimental values. Moreover, material synthesis is intricately influenced by factors such as kinetics and experimental conditions. To overcome these limitations, a novel collaborative approach was proposed for CSP that combines DFT with experimental data, utilizing advanced deep learning models and optimization algorithms. We illustrate the capability to predict formation enthalpies that closely align with actual experimental observations through the transfer learning on experimental data. By incorporating experimental synthesizable information of crystals, our model is capable of reverse engineering crystal structures that can be synthesized in experiments. Applying the model to 17 representative compounds, the results indicate that the model can accurately identify experimentally synthesized structures with high precision. Moreover, the obtained formation enthalpies and lattice constants closely align with experimental values, underscoring the model's effectiveness. The synergistic approach between theoretical and experimental data bridges the longstanding disparities between theoretical predictions and experimental results, thereby alleviating the demand for extensive and costly experimental trials.