Artificial Intelligence for Chalcogenide Materials: Evidence, Validation, and Emerging Closed-Loop Workflows

Chalcogenides challenge materials artificial intelligence (AI) because their properties depend on defects, disorder, metastability, interfaces, processing history, and operating conditions, as well as composition and structure. This review examines AI applications across transition-metal dichalcogenides, photovoltaic materials, thermoelectrics, phase-change materials, chalcogenide perovskites, and glasses. It combines structured bibliographic analysis with critical assessment of representative studies rather than exhaustive evaluation of all retrieved publications. Its central contribution is a claim-specific, five-level evidence framework: Level 0, method transfer or proposal; Level 1, internal retrospective validation; Level 2, domain-aware or external validation; Level 3, prospective computational or experimental validation; and Level 4, closed-loop operation in which new results guide subsequent decisions. Evidence levels characterize individual claims, not entire papers. Descriptor-based models, graph neural networks, and machine-learned interatomic potentials offer complementary capabilities, but predictive accuracy alone does not establish transferability or practical reliability. Prospective validation is documented in selected prediction, synthesis, and processing applications. Among the chalcogenide-specific studies examined, however, no workflow met the Level 4 criteria for repeated model-updated closed-loop operation. Progress requires traceable datasets, leakage-resistant evaluation, uncertainty assessment under relevant distribution shifts, physically meaningful validation, and stronger connections between modelling, synthesis, processing, and characterization. The framework clarifies existing evidence and priorities for future validation.