Data-driven design of carbon dots: Property prediction, optimization, and prospects for autonomous discovery

This review provides a comprehensive analysis of machine learning (ML)-assisted and data-driven design strategies for carbon dots (CDs), covering the transition from property prediction and synthesis-parameter optimization toward emerging inverse-design and autonomous-discovery frameworks. CDs exhibit complex structural heterogeneity and nonlinear synthesis–structure–property relationships, which limit conventional rational design approaches and create challenges for reliable data-driven modeling. This review systematically discusses the applications of machine learning methods, including ensemble learning, active learning (AL), Bayesian optimization (BO), uncertainty quantification, and explainable artificial intelligence, for predicting optical properties, quantum yield, emission behavior, and application-related performance of CDs. The distinctions among forward property prediction, parameter optimization, AL, inverse design, closed-loop experimentation, and autonomous discovery are critically evaluated. Current studies mainly focus on prediction and optimization within predefined chemical spaces, whereas complete inverse design requires target-property definition, candidate generation, physical and synthesizability constraints, ranking strategies, and experimental validation. Remaining challenges include data standardization, descriptor representation, model validation, uncertainty calibration, interpretability, and cross-system transferability. Future progress will rely on integrating machine learning with physically informed models, standardized multimodal datasets, and automated experimentation to establish more reliable and transferable principles for data-driven carbon-dot design.