Data-Driven Synthesis of Covalent Organic Frameworks via Machine Learning with Integrated Success–Failure Data

Abstract Covalent organic frameworks (COFs) have emerged as a versatile class of porous materials with promising applications in catalysis, energy storage, and gas adsorption. However, their synthesis remains a major bottleneck primarily due to the widespread reliance on inefficient trial-and-error approaches that waste resources and delay discovery. Herein, we address this challenge by integrating success and failure data: we curated 1822 in-house failed synthesis records and extracted 2603 successful cases from the literature. A random forest machine learning (ML) model, selected for its robustness with complex experimental data sets, achieved 91% accuracy in solvent prediction and demonstrated strong predictive performance for reaction temperature (R2 = 0.9129) and reaction time (R2 = 0.9596), significantly outperforming models trained solely on success-only or trial-and-error data sets. Building on this model, we developed the ML-COF toolkit, which automates the retrieval of known COF synthesis routes and accurately predicts conditions for novel frameworks. Experimental validation, conducted across eight structurally complex and synthetically challenging COFs, including six M(salen)-COFs and two M(Rabson)-COFs, confirmed the toolkit’s high predictive accuracy, with all targets successfully synthesized under predicted conditions. This work establishes a data-driven paradigm for COF synthesis, offering a practical alternative to trial-and-error methods and paving the way for accelerated discovery and future robot-assisted synthesis.