Machine Learning for the Calculation of Interatomic Potentials in WS2 and MoS2 Using the Gaussian Approximation Potential (GAP)

This work presents the development and validation of a Gaussian Approximation Potential (GAP) for MoS2 and WS2, trained on Density Functional Theory (DFT) reference data. The aim was to bridge the gap between quantum-mechanical accuracy and the efficiency needed for large-scale atomistic simulations. DFT calculations were performed using the Perdew–Burke–Ernzerhof (PBE) functional with various dispersion correction schemes. Analysis of the elastic constants showed that the D3(BJ) correction offered the best agreement with literature values and was selected as the reference framework for database generation. The GAP was trained using the Smooth Overlap of Atomic Positions (SOAP) descriptor, enabling machine-learning representation of local atomic environments. Validation against an independent DFT dataset showed excellent agreement, achieving a mean absolute energy error of 0.0002 eV/atom and a force RMSE of 0.20 eV/Å. The parity and error plots confirmed the model’s reliability across diverse configurations. The very high accuracy is attributed to the structural similarity between the training and validation data, both representing low-distortion geometries. Nevertheless, the computational gain is remarkable: a single DFT calculation for the 96-atom supercell required about 1 hour and 21 minutes, while GAP reproduced equivalent results in under 0.05 seconds, exceeding a 105-fold speedup. Overall, the developed GAP successfully reproduces DFT-level accuracy at a fraction of the cost, establishing a solid foundation for future work.