Search papers, labs, and topics across Lattice.
This paper develops and benchmarks machine learning interatomic potentials (MLIPs) based on the MACE architecture, specifically QCOF models, trained on DFT data of COF conformations. The QCOF models demonstrate superior accuracy, efficiency, and transferability compared to general-purpose MLIPs, particularly in predicting forces in defective COFs and phonon dispersion. Large-scale non-equilibrium molecular dynamics simulations using the best QCOF model reveal the impact of defects on thermal conductivity and mechanical properties of CTF-1 and COF-LZU1 systems.
COFs can withstand defects surprisingly well: mechanical properties remain stable even with defects, but thermal conductivity plummets, revealing design trade-offs.
Covalent Organic Frameworks (COFs) are versatile two-dimensional (2D) materials for flexible electronics, catalysis, and sensing, owing to their tunable architectures and large surface areas. However, like most materials, COFs inevitably contain synthesis-induced defects, which-similar to graphene-can strongly influence intrinsic properties, such as thermal transport and mechanical strength. To address this challenge, we have assessed the performance of a set of machine learning interatomic potentials (MLIP) capable of efficient large-scale simulations of COFs with quantum accuracy. In doing so, QCOF models (Quantum COF) were developed by tuning the state-of-the-art MACE architecture on an extensive dataset of non-equilibrium COF conformations generated from high-fidelity density functional theory calculations. The accuracy, computational efficiency, memory footprint, and transferability to unseen chemical environments of these models were benchmarked against general-purpose MACE models and their fine-tuned variants. Our results show that an invariant QCOF model with a small descriptor dimensionality and cutoff outperforms all other models in most validation tasks, including scalability to large systems, force prediction in defective COFs, and phonon dispersion calculations. The best-performing QCOF model was then used to run large-scale simulations of thermal conductivity for defective CTF-1 and COF-LZU1 systems via non-equilibrium MD, revealing a more pronounced sensitivity of CTF-1 to structural defects. Stress-strain curves were also investigated, showing that the mechanical response remains nearly invariant at low defect densities, while asymmetric behaviour emerges at large strains. This work thus provides a foundation for the design of robust quantum-informed MLIP for large-scale property simulations of defective of extended network materials.
