An efficient strategy to construct general machine learning potentials for high-entropy ceramics

Molecular dynamics (MD) simulations powered by machine learning potentials (MLPs) have become an effective way to investigate complex systems. However, the construction of transferable MLPs with broad elemental applicability for HECs remains a challenge due to the vast compositional space. Taking high-entropy carbides (HECs) as the model, we propose a strategy to efficiently construct general neuroevolution potentials (NEPs) for HECs with up to ten principal elements based on carbides with low entropy. The trained NEP exhibits high accuracy and transferability for 3-10HECs with low testing errors of 15. 7 meV/atom and 301 meV/ for energy and force, respectively. Moreover, the accuracy, generalization, and reliability of our established NEP are further validated through the accurate predictions on structural, mechanical, and thermal properties of HECs with the comparison to first-principles calculations, experimental measurements, or the rule of mixture. Our work provides an efficient solution to developing general MLPs for high-entropy ceramics.