Improved Deep Potential Model for Fast and Accurate Molecular Dynamics Simulations*

We present updates to the Deep Potential neural network architecture, incorporating a message-passing framework and higher-order tensor descriptors. Our model achieves accuracy on par with top machine learning force fields while delivering very fast inference speeds. These, combined with its stability, make it ideal for simulating large systems over long durations. We provide a convenient, flexible, and efficient GPU implementation. We demonstrate the model's effectiveness through its applications in specific accuracy-demanding condensed matter systems.