Enhanced Simulation of Complicated MXene Materials with Graph Convolutional Neural Networks

MXene, a notable two-dimensional transition metal carbide, has attracted increasing attention in materials science due to its unique attributes, driving innovations in energy storage, sensors, catalysts, and electromagnetic shielding. The property and application performance are determined by the electronic structure, which can be described based on the density of states (DOS). The conventional density functional theory (DFT) calculation is able to provide the DOS spectrum of a specific atomic structure. However, for complicated composition, such as the recently reported high entropy MXene, the DFT calculations in exhaustive structure space are resource-intensive. In this study, machine learning (ML) technique, specifically the crystal graph convolutional neural networks (CGCNN) model, is applied to generate DOS of these MXene models with complex compositions. By using calculations on M C and M C structures as training sets, the DOS of the complex high entropy MXene is well reproduced according to the atomic structure. Moreover, the adsorption energy of lithium is precisely predicted based on the DOS, which can be further employed to screen the potential electrode materials for lithium batteries. Herein, ML method not only streamlines predictions but also enhances the understanding of MXene's intrinsic properties.