Machine learning (ML) is transforming the research paradigm of materials science. As a mathematical tool adept at solving high-dimensional problems, it enables efficient process optimization, on-demand customized synthesis and quantitative interpretation of multifactorial growth mechanisms. While numerous ML-assisted strategies have been developed, including active learning loops, supervised-learning workflows and transformer-based language models for functional material design and process optimization, existing ML applications in synthesis remain fragmented, lacking a holistic framework that integrates process optimization, customized synthesis, and multifactorial mechanism deciphering, especially under data scarcity.
A joint team led by Prof. Shanshan Wang from the National University of Defense Technology, Prof. Fangping Ouyang from Central South University, and Prof. Jin Zhang from Peking University has developed a data-knowledge dual-driven machine-intelligence framework for the experiment-efficient, full-chain synthesis of 2D ReSe 2 dendrites, a class of materials with broad prospects in catalysis and nonlinear optics. This approach integrates active learning, data-augmented XGBoost algorithm and interpretable ML to accomplish three objectives: (i) rapid process optimization, where the branching degree of products increased by 69.4% within one week; (ii) construction of a nonlinear recipe–morphology mapping for customized growth using less than 1.5% of all possible parameter combinations; and (iii) interpretable multifactorial mechanism decoding via combined SHAP analysis, multiscale characterizations, and phase-field simulations.
“This work establishes a human-AI collaborative paradigm that intertwines data-driven induction with knowledge-driven reasoning, which can be extended to diverse low-dimensional materials and general CVD systems by resetting optimization targets to material thickness, surface coverage and phase/stacking-related features, as well as incorporating extra process variables like chamber pressure and heating rate.” Notes Prof. Wang. The work entitled “Data-knowledge dual-driven intelligent framework for full-chain, experiment-efficient synthesis of 2D dendrites” has been accepted for publication in Science Bulletin (DOI: 10.1016/j.scib.2026.08.033).
Science Bulletin
Experimental study