Chemist-Guided Human-AI Workflow for Covalent Organic Framework Synthesis

The synthesis of covalent organic frameworks (COFs) is still largely driven by chemists' literature-informed intuition and iterative trial-and-error, which can be difficult to scale and reproduce. Here we present a chemist-guided human-AI workflow that digitizes this reasoning loop─search, hypothesis formation, and iteration─by coupling a structured literature knowledge base with retrieval-augmented large language models and experiment-aware updates. We first construct a COF synthesis knowledge base containing 2709 protocols extracted from over 800 publications. Given an unseen linker combination, the workflow retrieves a Top-K neighborhood and assembles evidence through stratified sampling and context permutation, generating a range-type synthesis prior over solvent system, catalyst, temperature, time, and stoichiometry. A diagnosis module then interprets macroscopic observations together with powder X-ray diffraction (PXRD) files using a failure taxonomy and proposes targeted updates and next-round experiments. In leave-one-out benchmarks on 60 held-out COFs, the best context-assembly and self-consensus settings improve solvent-catalyst hit rates from baseline levels to up to 0.83, supporting robust transfer beyond individual case studies. We demonstrate the workflow by synthesizing two fluorinated COFs, TAPPy-4F and TAPPy-8F, both exhibiting crystallinity and permanent porosity. By simulating the chemist's reasoning loop, this human-AI system integrates expert knowledge with model-driven exploration, offering a generalizable and scalable paradigm for the rational design of complex reticular materials.