The Open DAC 2023 Dataset and Challenges for Sorbent Discovery in Direct Air Capture

Anuroop Sriram, Sihoon Choi, Xiaohan Yu, Logan M. Brabson, Abhishek Das, Zachary W. Ulissi, Matt Uyttendaele, Andrew J. Medford, David S. Sholl

Introduction

Direct Air Capture (DAC) is increasingly recognized as a vital component in global efforts to reach net-zero emissions. By removing CO2 directly from the atmosphere, DAC offers a pathway to mitigate hard-to-abate sectors and potentially reverse historical emissions. However, the widespread deployment of DAC is hindered by the lack of efficient, low-cost sorbent materials. Metal-Organic Frameworks (MOFs) are promising candidates due to their high surface area and tunable chemistry, but finding the ideal MOF for DAC—one that selectively binds dilute CO2 in the presence of atmospheric humidity—is a "needle in a haystack" problem.

To accelerate this discovery process, researchers from Meta AI, Georgia Tech, and Oak Ridge National Laboratory have introduced Open DAC 2023 (ODAC23). This dataset represents the largest collection of high-fidelity quantum mechanical calculations for MOFs to date, designed specifically to train machine learning (ML) models that can screen millions of materials at a fraction of the cost of traditional simulations.

Overview of the Open Direct Air Capture 2023 (ODAC23) dataset, highlighting the MOFs, adsorbates, ML tasks, and target applications.

The Challenge: Accuracy vs. Scale

The search for DAC materials faces two primary technical hurdles. First is the Accuracy Gap. Traditional "Force Field" (FF) simulations are computationally efficient but often fail to accurately model the strong chemical interactions (chemisorption) necessary for capturing CO2 at low concentrations (approx. 420 ppm). While Density Functional Theory (DFT) provides the required accuracy, its high computational cost makes it impractical for large-scale screening.

Second is the Humidity Problem. Most existing MOF datasets focus on CO2 capture in a vacuum. In reality, water vapor is far more abundant in the air and often competes for the same binding sites. A material that performs well in dry conditions may be completely ineffective in the humid environments typical of real-world DAC operations.

The ODAC23 Dataset

The ODAC23 dataset addresses these challenges by providing a massive ground-truth library for ML training. It encompasses over 38 million single-point DFT calculations and 176,000 structural relaxations, covering approximately 8,400 MOF materials derived from the CoRE MOF 2019 database.

Distribution of DFT calculations across training, validation, and test sets for the S2EF and IS2RS/IS2RE tasks, categorized by molecular composition.

A key feature of ODAC23 is its focus on realism. The dataset includes:

  • Defective Structures: Simulating missing linkers or metal sites to reflect the imperfections found in synthesized materials.
  • Co-adsorption: Explicitly modeling the interaction of CO2 and H2O together within the MOF pores.
  • Out-of-Domain Benchmarks: Specialized test sets to evaluate how well models generalize to entirely new molecular topologies and chemical environments.

Machine Learning for Material Discovery

The researchers benchmarked several state-of-the-art Graph Neural Network (GNN) architectures, including EquiformerV2, GemNet-OC, and eSCN. These models were trained on three primary tasks:

  1. S2EF (Structure to Energy and Forces): Predicting the energy and atomic forces of a given configuration.
  2. IS2RE (Initial Structure to Relaxed Energy): Predicting the final energy of a system from its initial, unrelaxed state.
  3. IS2RS (Initial Structure to Relaxed Structure): Predicting the final equilibrium geometry of molecules within the framework.

The EquiformerV2 (large) model emerged as the top performer, achieving a Force Mean Absolute Error (MAE) of 8.20 meV/Å. More importantly, these ML models can approximate DFT-level results orders of magnitude faster, enabling the screening of materials in seconds that would otherwise require centuries of supercomputing time.

Key Insights and Discoveries

The analysis of the ODAC23 dataset yielded several significant insights for material scientists:

  • Failure of Classical Models: The study confirmed that classical force fields are fundamentally limited to "physisorption" (weak interactions) and cannot reliably predict the chemisorption required for effective DAC.
  • Defects as a Design Tool: Interestingly, the data suggests that structural defects (like missing linkers) can sometimes enhance a MOF's CO2 affinity, suggesting that "imperfect" materials might actually be superior for carbon capture.
  • Identifying Water-Resistant MOFs: By explicitly modeling co-adsorption, the team identified a rare subset of approximately 240 MOFs that maintain a preference for CO2 even in the presence of water, providing a high-priority shortlist for experimental validation.

Conclusion and Future Work

The release of ODAC23 marks a major milestone in the application of AI to climate technology. By providing the data necessary to train high-accuracy ML models, the researchers have created a powerful engine for material discovery.

Future work will focus on extending these models to "high loading" scenarios—understanding how MOFs behave as their pores fill up—and integrating them into Grand Canonical Monte Carlo (GCMC) simulations to predict full adsorption isotherms. These steps are crucial for translating molecular-level predictions into the design of industrial-scale DAC processes.

Summary updated

BibTeX
@inproceedings{sriram2024open,
  author = {Anuroop Sriram and Sihoon Choi and Xiaohan Yu and Logan M. Brabson and Abhishek Das and Zachary W. Ulissi and Matt Uyttendaele and Andrew J. Medford and David S. Sholl},
  title = {{The Open DAC 2023 Dataset and Challenges for Sorbent Discovery in Direct Air Capture}},
  booktitle = {ACS},
  year = {2024},
  doi = {10.1021/acscentsci.3c01629},
}
Figure