DAC-SIM: Accelerating CO₂ Direct Air Capture Screening for Metal-Organic Frameworks with a Transferable Machine Learning Force Field

Yunsung Lim1,*, Hyunsoo Park2,*, Aron Walsh2,†, and Jihan Kim1,†
KAIST1 Imperial College London2
*Indicates Equal Contribution
arXiv DAC-SIM code DAC-SIM Doc
MY ALT TEXT

Abstract

Direct air capture (DAC) of CO₂ is necessary for climate change mitigation, but it faces challenges from low atmospheric CO₂ concentrations and competition from water vapor. Metal-organic frameworks (MOFs) are attractive candidates for DAC owing to their exceptionally high surface area, tunable porosity, and potential for adsorption-based capture processes with relatively low regeneration cost. Identifying optimal MOFs is hindered by their structural complexity, the vastness of their chemical space, and the expense of accurate simulations. Here, we present a machine learning force field (MACE-DAC) tailored for CO₂ and H₂O interactions in MOFs by finetuning the foundation model MACE-MP-0. To address smoothing issues and catastrophic forgetting, we curated the diverse GoldDAC dataset and introduced a continual learning loss function. To efficiently sample gas configurations, we developed the DAC-SIM package that uses MLFFs to achieve ab initio quality thermodynamics based on Widom insertion at computational speeds comparable to classical force fields. High-throughput screening on more than 8,000 synthesized MOF structures was performed to identify optimal MOFs and extract important chemical features. This approach overcomes prior limitations in describing CO₂/MOF and H₂O/MOF interactions, providing a scalable and accurate framework for accelerating DAC research for porous materials.

Overview

MY ALT TEXT

DAC-SIM package
✅ Developing a transferable machine learning force field (MLFF) for CO₂ and H₂O interactions with MOFs
✅ Integrate the MLFF into the DAC-SIM package to run Widom insertion (Monte Carlo) simulations

A Transferable MLFF in MOFs for DAC

MY ALT TEXT

Two main issues in finetuning the foundation model MACE-MP
✅ Smoothing issues
➡️ GoldDAC dataset including whole potential energy surface of interactions between CO₂ and H₂O with MOFs

✅ Catastrophic forgetting
➡️ continual learning loss function to preserve the knowledge of the pretrained model during finetuning

Potential energy surface of CO₂ and H₂O interactions with MOFs

MY ALT TEXT

Comparison of classical force field and MLFF
✅ The classical force field approach (UFF+DDEC) struggles to capture the interactions between CO₂ and H₂O with MOFs
✅ While the MLFF approach (MACE-DAC) provides a more accurate description of the interactions

High-throughput screening of MOFs for DAC

MY ALT TEXT

Screening on CoRE MOF database
✅ DAC-SIM identified 167 promising MOFs for DAC applications (Heat of Adsorption, Qst > 50 kJ/mol and CO2/H2O selectivity > 1)
✅ While only few MOFs were identified by the classical force field approach (UFF+DDEC) according to the paper by Findley et al. (2021)

Distribution of Heat of Adsorption

MY ALT TEXT

Comparison to classical force field simulations
✅ The distribution of Qst for CO₂ is quite similar between the MLFF approach (MACE-DAC) and the classical force field approach (UFF+DDEC)
✅ While the distribution of Qst for H₂O is significantly different between the two methods
🚀 check out the analysis at the section "Comparison to classical force field simulations" in our paper

Chemical Features for DAC applications

MY ALT TEXT

Important chemical features for DAC applications
✅ We categorize 7 chemical features from the 167 promising MOFs for DAC applications
🚀 check out the detailed chemistry analysis at Discussion section in our paper

BibTeX

@article{Lim2024accelerating,
    title={Accelerating CO2 Direct Air Capture Screening for Metal-Organic Frameworks with a Transferable Machine Learning Force Field},
    author={Lim, Yunsung and Park, Hyunsoo and Walsh, Aron and Kim, Jihan},
    journal={ChemRxiv},
    doi={10.26434/chemrxiv-2024-7w6g6},
    year={2024},
    }