[论文解读] Multi-fidelity Machine Learning Interatomic Potentials for Charged Point Defects

Xinwei Wang, Irea Mosquera‐Lois+1 more|arXiv (Cornell University)|Mar 5, 2026

Machine Learning in Materials Science被引用 0

一句话总结

The paper shows that foundation MLIPs trained on bulk data fail for charged point defects and introduces a charge-embedded, multi-fidelity MACE-based model that accurately predicts defect structures and transition levels, with significant cost reductions for defect landscape exploration.

ABSTRACT

Machine learning interatomic potentials (MLIPs) can now reproduce the energy, forces and stresses of bulk materials with high accuracy compared to first-principles calculations. The description of imperfections, where coordination environments and electron counts deviate from those found in pristine reference structures, remains a challenge. We find that the current generation of foundation MLIPs do not describe the defect physics of the semiconductor Sb2Se3. We introduce global defect charge embeddings that distinguish the bonding characteristics of different charge states. We further employ a multi-fidelity approach that combines low-cost (semi-local exchange-correlation functional) reference data with high-quality (non-local hybrid functional) energies and forces that describe well the subtleties of the defect energy landscape. The resulting defect-capable force fields can find stable structural configurations and predict charge-transition levels in quantitative agreement with direct quantum mechanical calculations, at a fraction of the computational cost.

研究动机与目标

• Demonstrate the failure of bulk-trained foundation MLIPs to describe charged point defects in semiconductors.
• Develop a defect-aware MLIP with a global charge embedding to distinguish multiple charge states.
• Introduce a multi-fidelity training approach to efficiently learn high-accuracy defect energetics.
• Show that the charge-embedded MF model identifies ground-state defect structures and transition levels in agreement with hybrid DFT.

提出的方法

• Benchmark foundation MLIPs against DFT for V_Sb in Sb2Se3 across five charge states to demonstrate failure in defect physics.
• Introduce a bespoke MACE-based model with a global defect charge embedding that maps total charge to a learnable vector added to atomic embeddings.
• Train on defect configurations across multiple charge states using hybrid functional data (HSE06) to provide reliable defect energetics.
• Apply a multi-fidelity strategy that densely samples the low-cost PBE PES with a sparse set of HSE06 corrections to learn the delta between low- and high-fidelity energies.
• Use the charge-embedded MF model to perform defect structure searches and compute configuration coordinate diagrams at hybrid-level accuracy with reduced cost.

实验结果

研究问题

• RQ1 Can foundation MLIPs trained on bulk data reliably identify ground-state structures of charged defects across multiple charge states?
• RQ2 Does incorporating a global charge embedding enable a single MLIP to distinguish different defect charge states and reproduce defect energetics?
• RQ3 Can a multi-fidelity training scheme efficiently capture high-accuracy defect landscapes at a fraction of the cost of full hybrid-functional calculations?
• RQ4 How well do charge-embedded MF MLIPs reproduce ground-state geometries and thermodynamic transition levels compared to HSE06 DFT?

主要发现

• Foundation MLIPs fail to reliably identify defect ground states across charge states, with RMSD typically 0.2–0.4 Å and no consistent ground-state recognition.
• A bespoke charge-embedded MACE model separates defect configurations by charge state in descriptor space and predicts ground-state geometries within 0.05 Å, with energies and forces matching hybrid-DFT in RMSE ~0.48 meV/atom and ~20.15 meV/Å respectively.
• Training on five charge states with HSE06 data yields defect energetics and thermodynamic transition levels within ~0.01 eV of hybrid functional references.
• A multi-fidelity training strategy combining dense PBE sampling and sparsely sampled HSE06 corrections can discover global minima missed by standard defect screening, reducing HF cost by ~3 orders of magnitude while maintaining accuracy.
• The MF model enables efficient generation of configuration coordinate diagrams that reproduce HSE06-level results at negligible cost, enabling high-throughput defect screening and property prediction.