114年-2-9-實作學生：資訊工程學系【譚永祺】

參賽序號：2-9

海報主題

MXene-Specific Machine-Learned Potential — A DFT Surrogate for Structure Relaxation and Surface Chemistry​

系級

資訊工程學系

指導老師及參賽學生

指導老師：林新
參賽學生：譚永祺

構想說明

This research presents a MXene-specific machine-learned potential (MLP) designed as a surrogate for density functional theory (DFT) calculations. Thousands of MXene-like structures from the Meta Open Materials 2024 (OMat24) dataset were curated, balanced by composition and termination, and converted into datasets containing energies, forces, and stresses. An E(3)-equivariant graph neural network (MACE) was trained to enable structure relaxation and surface-chemistry predictions. The framework incorporates uncertainty ensembling and an active learning loop to identify out-of-distribution cases and prioritize new DFT labeling. Deployed as an ASE calculator, the model significantly accelerates geometry optimization and high-throughput screening, while preserving DFT-level accuracy in lattice parameters, interlayer spacing, work-function shifts, and adsorption energies. With per-structure total-energy errors below 10 eV, this MXene-focused surrogate provides a powerful tool for rapid exploration of the composition–termination–adsorbate design space, with promising applications in electrocatalysis and sensing.

我要投票

    分享：