| 초록 |
Recently, chemical space exploration and catalyst discovery have been accelerated with the help of machine learning techniques, which could bypass computationally expensive density functional theory (DFT) calculations. In addition, an inverse design strategy was suggested, which predicts novel materials with pre-defined target properties in a broad searching space. However, it is challenging to validate stability and synthesizability of a large number of materials through DFT calculations. To overcome this challenge, we developed Direction-based Crystal Graph Convolution Neural Network (D-CGCNN) using directional information as crystal graph representation. We found that the directional information is practically identical for unrelaxed and relaxed crystal structures, generating equivalent inputs for both structures. The convolutional neural network based on our representations predicted formation energies of unrelaxed structures more accurately compared to the state-of-the-art model (0.186 vs. 0.220 eV/atom for RMSE). We expect our approach will enable significant reduction in computational costs during the inverse design process. |
