Transferable Neural Network Potential for Elastic and Vibrational Properties of Carbon, Hydrogen, Oxygen, and Nitrogen-Based Two Dimensional Covalent Organic Frameworks

Covalent organic frameworks (COFs) are promising porous materials for a wide range of applications; however, first-principles simulations of COFs remain challenging due to their large unit cells compared with inorganic materials. Here, we present a machine learning interatomic potential (MLIP), named “COF-NN”, specifically developed for two-dimensional COFs composed of carbon, hydrogen, oxygen, and nitrogen. By combining a molecular-cluster training strategy, constructed from COF-relevant monomer condensation chemistry, with active learning and uncertainty quantification, this artificial neural network-based MLIP achieves both high efficiency and robust performance across a set of structurally diverse 2D COFs while retaining near–DFT accuracy. Benchmarking results show that COF-NN accurately reproduces equilibrium structures, elastic constants, and phonon frequencies, with performance comparable to that of two state-of-the-art universal MLIPs. Beyond COFs, our cluster-based active learning framework provides a general strategy for constructing interatomic potentials that transfer local chemical environments learned from molecular clusters to periodic, complex, low-symmetry porous materials. We believe this work enables scalable atomistic simulations and high-throughput discovery of COF materials with predictive fidelity that was previously inaccessible.