Machine-learning-guided predict and design of polymeric carbon nitride for H2O2 photosynthesis

The multifactor effects (e.g., catalyst preparation conditions, catalyst features, and photocatalytic reaction conditions) on hydrogen peroxide photosynthesis by polymeric carbon nitride present a complex yet unexplored avenue. Herein, this study presents a data-driven approach based on machine learning algorithms to understand the key correlations between multi-factors and hydrogen peroxide yield. A closed-loop intelligent system was developed to bridge the gap between data and practical application. The training model features are carbon nitride synthesis, photocatalytic process, and material property, in which the element characteristics were considered for the first time. This method allows classification and quantitative prediction of the hydrogen peroxide yield from modified carbon nitride under multifactor effects. After training and comparison, the random forest model demonstrated the best performance, as R 2 rose to 0.9266 , exhibiting excellent predictive performance. By employing SHapley Additive exPlanations analysis and element influence decoding, the model was interpreted to understand the inner workings of the machine learning "black box". This study revealed critical insights into modifier selection and feature-importance hierarchy, providing an in-depth understanding for the efficient design and regulation of carbon nitride photocatalysts for efficient H 2 O 2 production. Based on this model, three carbon nitride materials were prepared under the model prediction according to a simple calcination method. An optimizing yield of 13194 μM h −1 , surpassing 90 % of the values in previous research, was achieved in the K-doped catalyst with urea as the precursor.