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

Area of Science:

• Materials Science
• Artificial Intelligence
• Chemical Synthesis

Background:

• Covalent organic framework (COF) synthesis relies heavily on expert intuition and iterative experimentation, posing challenges for scalability and reproducibility.
• Current methods lack systematic approaches for rational design, hindering the efficient discovery of novel COF materials.

Purpose of the Study:

• To develop a chemist-guided human-AI workflow for digitizing and optimizing the COF synthesis reasoning loop.
• To enhance the prediction of optimal synthesis conditions (solvent, catalyst, temperature, time, stoichiometry) for COFs.
• To enable the rational design and scalable synthesis of complex reticular materials.

Main Methods:

• Constructed a COF synthesis knowledge base from over 800 publications (2709 protocols).
• Coupled a structured literature knowledge base with retrieval-augmented large language models for hypothesis generation.
• Integrated an experiment-aware diagnosis module using macroscopic observations and powder X-ray diffraction (PXRD) for iterative updates.

Main Results:

• The human-AI workflow demonstrated improved solvent-catalyst hit rates up to 0.83 in leave-one-out benchmarks on 60 held-out COFs.
• Successfully synthesized two fluorinated COFs (TAPPy-4F and TAPPy-8F) exhibiting crystallinity and permanent porosity.
• The system showed robust transferability beyond individual case studies, validating its generalizability.

Conclusions:

• The developed human-AI workflow offers a generalizable and scalable paradigm for the rational design of COFs.
• This approach integrates expert knowledge with model-driven exploration, accelerating materials discovery.
• The system effectively simulates the chemist's reasoning loop, improving synthesis efficiency and reproducibility.