2026/02/14 by Connor W. Edwards, Fengxu Yang, Konstantin Stracke +1 · 1 voice · 1 citation
Physics and Astronomy · #cond-mat.mtrl-sci
Grand canonical Monte Carlo (GCMC) simulations are essential for screening metal-organic frameworks (MOFs) for gas adsorption, yet their accuracy is limited by underlying interatomic potentials. Universal machine-learned interatomic potentials (MLIPs), trained on diverse chemical datasets, promise zero-shot prediction without system-specific training. We introduce MLIP-MC, an open-source Python framework to conduct GCMC simulations with MLIPs, and use this framework to benchmark a series of universal models, including MACE-MP-0, ORB-v3, and fairchem ODAC, for CO2 adsorption on ZIF-8, ZIF-4, and Mg-MOF-74. All universal models exhibit systematic biases, consistently over- or underestimating adsorption energetics. Crucially, accuracy depends on training data composition: only models trained on MOF-adsorbate interactions achieve reasonable agreement with a density functional theory derived reference. Errors grow linearly with CO2 uptake, reflecting compounding inaccuracies in adsorbate-adsorbate interactions. Our results demonstrate that current universal MLIPs require finetuning for quantitative adsorption predictions and demonstrate the power of MLIP-MC to rapidly test models.