COFFlow: Building-Block-Based Generative Modeling for Two-Dimensional Covalent Organic Framework Discovery

Abstract

Two-dimensional covalent organic frameworks (2D COFs) are promising porous materials with vast yet largely unexplored design space arising from diverse building blocks (BBs) and crystal topologies. Efficient crystal structure prediction (CSP) and inverse design of 2D COFs therefore remain highly challenging. Here, we develop COFFlow, a flow-matching-based generative framework for CSP of 2D COFs. A COF-specific deconstruction algorithm is introduced to consistently decompose periodic COFs into experimentally relevant BBs across diverse topologies, enabling direct generative assembly of periodic frameworks from BB representations. Systematic evaluations demonstrate that COFFlow can reliably and efficiently generate structurally consistent and chemically valid 2D COFs. Beyond the commonly used match rate (MR), we further introduce validity rate (VR) to assess the chemical validity of generated structures. The models exhibit nontrivial extrapolation capability toward unseen BBs and crystal topologies, although preserving chemical validity under strong distribution shifts remains challenging. To demonstrate the practical utility of the framework, we integrated COFFlow with data-driven screening and inverse design of 2D COFs for methane storage. Using deliverable capacity as the target metric for adsorbed natural gas application, we identified several previously unreported 2D COFs with performance among the highest reported to date. More broadly, this work establishes a scalable framework for AI-driven discovery and design of reticular materials.