Towards a Generic Large Atomic Model (DPA Model Series) for Materials Simulations

The rapid advancements in artificial intelligence (AI) facilitate significant improvements in atomic modeling, simulation, and design. A series of AI-driven potential energy models (i.e., DeepMD) have demonstrated their capability to conduct large-scale, long-duration simulations with the accuracy of ab initio electronic structure methods. Targeting the bottleneck of model generation, we worked towards a model-centric strategy, wherein a large atomic model (LAM), pre-trained across multiple disciplines, can be efficiently fine-tuned and distilled for various downstream tasks. Specifically, the DPA (DPA-1 & DPA-2) architectures, pre-trained on a diverse array of chemical and materials systems (i.e., perovskite oxide systems and IIB-VIA semiconductors) using a multi-task approach, demonstrate superior generalization capabilities across various downstream tasks compared to the traditional single-task pre-training and fine-tuning methodologies.