A process-level perspective on the impact of molecular force fields on the computational screening of MOFs for carbon capture

Abstract

Multiscale high-throughput computational screening (HTCS) has emerged as an important strategy to identify the most promising MOF adsorbents for carbon capture and sequestration (CCS). Within this approach, adsorption data from molecular simulations is passed on to detailed pressure swing adsorption (PSA) process models to evaluate the performance of candidate materials.1 One of the key design choices within the HTCS pipeline is the definition of a forcefield, which is required to perform molecular simulations of adsorption. Despite over a decade of searching, only one MOF, CALF-20, has been tasked to pilot-scale.2 Given that screening studies are meant to inform experimentalists on which MOFs to synthesize and test, this issue can be attributed, in part, to a lack of accurate, reproducible, and consistent implementations of these workflows. Therefore, contradictory conclusions on which materials to test, precipitated by inconsistent or inaccurate material property data, represents one of the current bottlenecks obstructing industrial-scale deployment of adsorption-based CCS technologies. The question we pose in this study is to what extent the ranking of materials, and the selection of top performers identified in PSA process modelling, depends on the choice of the commonly available forcefields. To answer this question, we first generated distributions of $CO_2$ and $N_2$ adsorption isotherms in 726 MOFs using six typical forcefields: the UFF or Dreiding sets of Lennard-Jones parameters, combined with partial charges derived from ab initio calculations (DDEC scheme) or by charge equilibration (EQeq or Qeq schemes).3 Here we reflect on the consistency of the adsorption data, and reason some of the important differences in molecular modelling approaches. Next, we determined the resulting distributions in process-level performances for each material. We then quantified the uncertainties which propagate from the material-level to the process-level and ranked materials by their performance according to different forcefield definitions. Finally, we explored potential pathways towards uncertainty mitigation in multiscale HTCS of MOFs for CCS and discuss future prospects for the field. Our results show that: (i) partial charge assignment is the prevailing source of uncertainty, and that charge equilibration schemes produce results which are inconsistent with the DDEC charge scheme; (ii) the uncertainties originating from partial charge assignment can be effectively mitigated using machine learning models trained to predict charges derived from ab initio calculations; (iii) the choice of Lennard-Jones parameters is still a considerable source of uncertainty in HTCS workflows. References: 1. A. H. Farmahini, Chem. Rev. (2021) 10666 – 10741 2. P. Hovington, Proc. 16th Greenhouse Gas Cont. Tech. Conf. (2022) 3. F. L. Oliveira, arXiv:2210.09456, (2022)