• Artificial intelligence is transforming electron microscopy from a qualitative imaging tool into a quantitative platform for discovering structure–property relationships in complex materials. • Recent self-supervised and physics-guided learning strategies overcome data scarcity and poor transferability, enabling robust analysis of noisy experimental data across diverse material systems. • Unsupervised identification of atomic-scale motifs and domains introduces a new paradigm for linking microstructural heterogeneity directly to functional performance without manual annotation. • Machine learning–driven interpretation of high-dimensional diffraction data now allows direct mapping of strain, phase evolution, and internal fields that were previously inaccessible using conventional methods. Artificial intelligence (AI) is revolutionizing the analysis of (scanning) transmission electron microscopy ((S)TEM) data by enabling high-throughput, objective, and scalable interpretation of atomic-scale structures. Although initial AI applications often relied on large, labeled datasets and extensive training, exhibiting limited generalizability across diverse materials systems, recent advancements have introduced more efficient and physically meaningful approaches. However, challenges persist in achieving broad applicability and extracting interpretable insights from data-driven outputs. This concise review highlights emerging strategies that address these limitations, including unsupervised learning, physics-informed models, and domain-adaptive frameworks. By showcasing representative examples, we discuss how these approaches enhance the robustness, efficiency, and scientific relevance of AI-aided STEM analysis, ultimately advancing our capacity to uncover structure–property relationships in complex materials.
Sohn et al. (Sun,) studied this question.