Gaussian processes enhanced active learning for efficient atomic cluster expansion interatomic potential development

Accurate interatomic potentials are crucial for simulating the behavior of materials at the atomic scale, particularly under extreme conditions. However, developing high-fidelity interatomic potentials typically demands an extensive database of atomic configurations obtained through computationally intensive density functional theory (DFT) calculations. To address this, we present an active learning strategy for constructing atomic cluster expansion (ACE) [1] interatomic potentials across a wide range of temperatures and pressures.



Our approach integrates Gaussian processes (GPs) to guide the selection of the most informative atomic configurations for DFT calculations, significantly reducing the computational cost of building the training database. The active learning model iteratively samples atomic configurations, assesses their uncertainty using GPs, and identifies cases where additional DFT calculations provide the most value. By dynamically updating the training set, the model ensures that the ACE potential is trained on a diverse and representative set of configurations, enhancing the accuracy and transferability of the interatomic potential across various conditions.



We applied this strategy to develop an ACE potential for platinum, demonstrating its effectiveness in accurately capturing complex phenomena such as defect properties and bulk behavior. Our results indicate that the active learning approach reduces the required number of DFT calculations by over 30% compared to traditional sampling methods while maintaining or improving the accuracy of the ACE potential.



[1] Ralf Drautz, Phys. Rev. B, 99, 014104 (2019).