Design, Synthesis, and Screening of COFs for CO₂ Adsorption by Gaussian Process

Effective CO₂ capture requires careful design of Covalent Organic Frameworks (COFs). Current computational approaches often rely solely on simulated properties, neglecting critical chemical and synthetic factors that determine real-world COF performance. We present an integrated computational-experimental framework combining machine learning with experimental validation. Our study analyzes 240 unique COFs (617 samples) with experimentally measured CO₂ adsorption capacities across varied synthesis conditions. Gaussian Process and CatBoost models were developed to predict CO₂ adsorption by simultaneously considering chemical structures, synthesis parameters, and measurement protocols. The GP model demonstrates improved generalization and uncertainty quantification compared to CatBoost. SHAP analysis reveals the model's focus on COF type and synthesis conditions. Using a database of 181 building blocks, we generated 5557 COF structures with synthesis condition recommendations based on experimental similarity. Experimental validation confirmed the predictions for three synthesized COFs, demonstrating the framework's practical utility for COF design and optimization.