DP-GEN: An Optimal Active Learning Protocol for Constructing Interatomic Potentials - Materials Simulation

Explore the groundbreaking DP-GEN (Deep Potential Generator) protocol for constructing interatomic potentials in materials simulation in this 45-minute conference talk by Roberto Car at #ICBS2024. Delve into how machine learning is revolutionizing molecular modeling by enabling simulations with ab initio quantum mechanical accuracy at nearly the cost of empirical force fields. Learn about the three-step iterative process of exploration, labeling, and training that forms the core of DP-GEN, designed to create a minimal training dataset for uniformly accurate deep potential models. Discover the key advantages of DP-GEN, including its flexibility in exploration methods, on-the-fly training dataset construction, and efficient selection of configurations for labeling. Examine real-world applications, such as modeling water's potential energy surface across various phases and thermodynamic conditions. Gain insights into the protocol's effectiveness for pure and mixed materials, ordered and disordered phases, and its integration with structure prediction algorithms. Understand the challenges and solutions for rare events and non-Boltzmann sampling techniques in the DP-GEN framework.