This review examines how machine learning advances the modeling of material defects, covering point defects, dislocations, surfaces, grain boundaries and interfaces. It summarizes defect-data construction, machine-learning interatomic potentials, configuration searches, active learning, high-throughput screening and multiscale simulations. Emphasis is placed on matching models and training data to defect physics, accounting for finite-temperature and long-range effects, and validating defect energetics, transport and structural evolution through calibrated uncertainty and independent benchmarks.
This review provides a systematic overview of machine learning in material defect research, connecting the physical description of defects with data construction, model selection and validation. It covers point defects, dislocations, surfaces, grain boundaries and interfaces, and discusses how their distinct atomic and electronic environments affect formation, migration, transport and structural evolution.
The discussion begins with the theoretical foundations and computational requirements of different defect classes. Defect formation energies and charge-transition levels are introduced alongside the treatment of chemical potentials, finite-size corrections and band-edge references. These foundations clarify why models must be selected according to the target property, particularly when local reconstruction, charge states, long-range interactions or interfacial segregation govern the relevant mechanism.
Defect-data construction is examined from initial structure generation to reference calculations and metadata recording. The coverage includes symmetry-based enumeration, structural perturbations and searches for stable and metastable configurations. Particular attention is given to defect cores, migration pathways, reconstructed interfaces and thermally accessible structures, which may be poorly represented in datasets dominated by nearly perfect crystals. Consistent computational settings and traceable data provenance are emphasized throughout.
Machine-learning interatomic potentials form a central part of the review. Their ability to learn energies and forces from first-principles data supports more extensive simulations of defect relaxation, migration and collective evolution within validated domains. Graph neural networks, sparse Gaussian-process models and committee approaches are discussed together with strategies for describing chemically diverse environments, long-range electrostatics and configuration-dependent charge response.
Configuration search, active learning and high-throughput screening are considered as complementary approaches to exploring complex defect landscapes. Uncertainty estimates help identify unfamiliar structures for additional first-principles calculations, followed by model retraining and validation. Defect configuration atlases provide a way to organize related structures, compare their energetics and identify candidate states or pathways for subsequent kinetic analysis.
The treatment of defect dynamics extends from molecular dynamics and finite-temperature free-energy sampling to accelerated dynamics and kinetic Monte Carlo. Connections with dislocation dynamics, phase-field methods and continuum models show how atomistic quantities can inform larger-scale structural evolution. Migration barriers, mobilities, interfacial energies and their uncertainties must be assessed when information is transferred between these descriptions.
Representative applications span semiconductors, halide perovskites, solid electrolytes, battery cathodes, irradiation damage and catalytic nanoparticles. These examples illustrate the use of machine learning to study defect thermodynamics, ion migration, defect interactions and degradation. They also demonstrate why finite-temperature effects, rare configurations and changes in local chemistry require explicit consideration in model development and assessment.
Validation is organized around the physical quantities that control the target mechanism. Formation energies, charge-transition levels, migration barriers, diffusion coefficients and kinetic observables are considered alongside energy and force errors. A reporting checklist addresses training-domain coverage, uncertainty calibration, physical consistency, transferability and reproducibility. Independent tests on previously unseen defect types or conditions help establish the range in which predictions can support reliable conclusions.
The outlook considers physical constraints, learning from data at different levels of accuracy, standardized data infrastructure and increasingly automated workflows. Fully autonomous defect discovery remains an emerging direction. Across these developments, the central conclusion is that reliable machine-learning-assisted defect research depends on combining computational efficiency with defect-specific data, physical understanding and independent first-principles or experimental validation.
This review, "Empowering material defect research with machine learning," was published in AI & Materials .
Ren F, Hu S, Zhang G, Cai Z, Han D. Empowering material defect research with machine learning. AI Mater. 2026(3):0012, https://doi.org/10.55092/aimat20260012.
AI & Materials
Literature review
Not applicable
Empowering material defect research with machine learning
28-Sep-2026