Explainable machine learning in carbon dot research: From prediction to scientific understanding

Carbon dots (CDs) are among the most structurally complex nanomaterials, exhibiting heterogeneous architectures, nonlinear formation pathways, and multifunctional properties that challenge conventional structure–property paradigms. These complexities have stimulated increasing interest in machine learning (ML) as a tool for modeling synthesis–property relationships and accelerating materials development. However, predictive accuracy alone is insufficient for advancing scientific understanding when the underlying mechanisms remain unclear. This review examines the emerging role of explainable machine learning (XML) in CD research, highlighting its transition from a predictive framework to a knowledge-extraction approach capable of revealing hidden relationships within complex materials datasets. We discuss the structural heterogeneity of CDs, the nonlinear nature of their formation pathways, and the limitations of current characterization methods that hinder reliable interpretation of functional behavior. Recent studies employing explainable and interpretable ML techniques are analyzed to assess their contributions to understanding synthesis–property relationships, resolving structural ambiguities, and guiding application-oriented design. Current challenges related to data quality, descriptor representation, model reliability, and causal interpretation are also evaluated. Finally, future directions involving causal inference, physics-informed learning, and autonomous discovery platforms are discussed as key pathways toward transforming CD research from empirical optimization to a predictive and scientifically interpretable discipline.