Machine learning for carbon dots: Capabilities, limitations, and the path toward rational design

Carbon dots (CDs), as significant luminescent carbon-based nanomaterials, exhibit broad application potential in sensing, biomedicine, and optoelectronic devices. However, their synthesis processes are highly nonlinear, and their structural heterogeneity is prominent, resulting in a long-standing lack of a decipherable framework for the synthesis-structure-property relationship, which has restricted on-demand design and controllable preparation. In recent years, machine learning (ML) has provided a data-driven paradigm to address this complexity, yet related research still faces challenges such as scattered data, unclear task boundaries, and insufficient model interpretability. This review systematically examines recent key advances in machine learning for CDs research, focusing on three core tasks: property prediction, synthesis and inverse design, and mechanism analysis. It critically analyzes the capabilities and limitations of various models in predicting emission wavelength, photoluminescence quantum yield, and phosphorescence lifetime. By comparing data sources, feature construction, and validation strategies, it points out that many current high-prediction accuracies primarily stem from statistical fitting rather than learning physical causality, particularly facing structural data bottlenecks in red or near-infrared emission and cross-system generalization. Furthermore, from the perspective of CDs application systems, the review systematically evaluates the practical enabling role of machine learning in CDs-related applications such as sensing, biomedicine, optoelectronics, and information encryption, clearly distinguishing between its role as a tool for performance optimization and its function as a key means for rational design of CDs materials. Finally, a future-oriented pathway for machine learning-driven CDs research is proposed: by constructing standardized, high-quality databases, introducing physically constrained and interpretable models, and combining active learning with closed-loop experimental validation, the field can advance from empirical trial-and-error toward a predictive, interpretable, and translatable paradigm of rational design.