Exploration of Potential Energy Surfaces of Multi-Element Materials Using Machine Learning and Ab-Initio methods

The discovery of novel functional materials requires efficient exploration of complex multielement systems, where conventional density functional theory (DFT)-based crystal structure prediction (CSP) is computationally expensive. Modern machine-learning interatomic potentials (MLIPs) provide high computational efficiency, but do not offer sufficient accuracy for reliable phase stability prediction without system-specific fine-tuning. This project aims to develop an AIaccelerated CSP framework by integrating fine-tuned MLIPs, artificial-intelligence-based structure generation, and targeted evolutionary search. A multi-level dataset will be constructed to train models such as MACE, MatterSim, GAP, SNAP, and NequIP with high energy prediction accuracy. The proposed approach will enable large-scale screening, identification of low-energy structures, and guided searches using USPEX, resulting in an efficient workflow for discovering stable and metastable materials.