Context-Aware Deep Learning for Defect Classification in Atomic-Resolution STEM

2026-06-08

AI Summary

Defect classification in atomic-resolution scanning transmission electron microscopy (STEM) images is often ambiguous when relying solely on image contrast, as it neglects critical contextual factors such as material composition and imaging conditions. This work proposes the first context-aware deep learning framework that integrates STEM images with multidimensional experimental metadata—including chemical composition, electron beam energy, and detector geometry—transforming an ill-posed classification problem into a physically interpretable inference task. Trained on a large-scale simulated dataset comprising 55 million image patches spanning 96 doped monolayer transition metal dichalcogenides, the method achieves over 98% accuracy on simulated data and demonstrates strong agreement with expert annotations on experimental data, reducing posterior entropy by 94%. This approach establishes a generalizable paradigm for multimodal autonomous materials characterization.

Abstract

Artificial intelligence is rapidly advancing materials characterization, yet most applications in electron microscopy rely solely on image contrast, overlooking the chemical and experimental context that shapes image formation. This limitation makes defect classification inherently ambiguous, as similar contrasts can arise from different materials or imaging conditions. Here we develop a context-aware learning framework that integrates image-derived contrast with metadata describing composition, beam energy, and detector geometry. Using a systematically constructed dataset of ~55 million simulated patches spanning 576 cases across 96 doped monolayer transition-metal dichalcogenides, we show that conditioning on contextual variables transforms defect classification from an ill-posed image-only task into a well-posed, physically grounded problem. The framework achieves over 98% accuracy on simulations and near-human agreement on experimental data, with a 94% reduction in posterior entropy. By emphasizing contextual grounding over architectural complexity, this approach links experimental image contrast to the underlying chemical and imaging conditions, supporting physically grounded defect assignments and a general pathway toward multimodal AI models for autonomous materials characterization.

Problem

• defect classification
• context-aware learning
• atomic-resolution STEM
• image contrast ambiguity
• materials characterization

Innovation

• context-aware learning
• defect classification
• atomic-resolution STEM
• multimodal AI
• materials characterization