TL;DR
This paper introduces the Open DAC 2023 (ODAC23) dataset, a massive collection of quantum chemistry calculations for CO2 and H2O adsorption in over 8,000 metal-organic frameworks (MOFs). This dataset is crucial for accelerating the discovery of new MOFs for direct air capture (DAC) by enabling the development of more accurate machine learning models that account for realistic conditions like humidity and MOF flexibility.
What Problem Does This Paper Address?
The urgent need to combat climate change has spurred interest in direct air capture (DAC) technologies, which remove CO2 directly from the atmosphere. Metal-organic frameworks (MOFs) are a promising class of materials for DAC due to their tunable structures and high porosities. However, discovering optimal MOFs for DAC has been hampered by several limitations in previous research:
- Inaccurate Modeling: Many studies relied on empirical force fields (FFs), which can lead to qualitative and quantitative inaccuracies, especially for complex interactions.
- Ignoring Humidity: Previous work often neglected the crucial role of water vapor (H2O) in the air, which can compete with CO2 for adsorption sites and even degrade MOFs.
- Focus on Hypothetical Materials: Many datasets included purely hypothetical MOFs, making it difficult to assess their practical synthesizability.
- Excluding Defects: Real-world MOFs often contain defects that can significantly influence their adsorption properties, but these were rarely considered.
This paper aims to overcome these limitations by providing a comprehensive dataset of high-fidelity quantum chemistry calculations that account for these factors, paving the way for more reliable MOF discovery for DAC.
Methods
The authors employed a multi-pronged computational approach:
- Quantum Chemistry Calculations: They performed extensive density functional theory (DFT) calculations using the PBE+D3 functional, which treats electronic interactions quantum mechanically and includes van der Waals forces. This approach offers higher accuracy than empirical force fields.
- MOF Structure Generation: They started with the experimentally-derived CoRE MOF database and introduced "missing linker" defects to create a diverse set of defective MOFs. They also included some "ultrastable" MOFs from another database for broader coverage.
- Adsorbate Placement and Relaxation: CO2 and H2O molecules were placed within the MOF structures, and then the entire system (MOF + adsorbates) was relaxed using DFT to capture the dynamic interactions and potential framework distortions.
- Machine Learning (ML) Model Training: The generated DFT data was used to train various graph neural network (GNN) architectures (e.g., GemNet-OC, eSCN, EquiformerV2) for predicting adsorption energies and atomic forces. This allows for rapid screening of new MOFs.
Datasets
- ODAC23 Dataset: This is the primary contribution of the paper. It comprises over 38 million single-point DFT calculations and ~176,000 adsorption energies for CO2, H2O, and their mixtures on approximately 8,000 MOFs. This dataset includes both pristine and defective MOFs.
- CoRE MOF Database: Used as the starting point for generating MOF structures.
- Ultrastable MOF Database: A subset of MOFs from this database was included to test model generalizability.
- Open Catalyst Project (OCP) Datasets: The ML models and training infrastructure were adapted from those developed for the OCP, which focuses on catalysis.
Key Insights and Ideas
- Humidity is Critical: The study strongly emphasizes that ignoring water vapor in DAC material screening is a significant oversight, as it often competes with CO2 for adsorption sites and can even lead to less favorable CO2 adsorption.
- Defects Matter: Defects in MOFs can have a profound, and sometimes counterintuitive, impact on adsorption properties, either enhancing or diminishing their suitability for DAC. This highlights the need to consider imperfect materials.
- FFs Fall Short for Chemisorption: Classical force fields, while useful for physisorption, struggle to accurately model the stronger chemical interactions (chemisorption) that are crucial for selective CO2 capture at low concentrations.
- ML Models Outperform FFs: The developed ML models demonstrate significantly better accuracy than classical FFs, especially in the chemisorption regime, making them a more promising tool for high-throughput screening.
- Generalizability is Key: The inclusion of out-of-domain test sets (novel linkers, topologies, and larger MOFs) is vital for assessing how well ML models will perform on materials not seen during training.
Key Results
- Massive Dataset: The ODAC23 dataset contains ~176,000 adsorption energies and ~38 million single-point DFT calculations, making it the largest such dataset for MOFs and DAC.
- Promising MOFs Identified: The dataset enabled the identification of 135 pristine MOFs that bind CO2 more favorably than H2O, and 107 defective MOFs with similar promising characteristics.
- ML Model Performance: The best ML models achieved force MAEs of ~8.20 meV/Å and energy MAEs of ~0.15 eV on the in-domain test set. These models showed good generalization to new linker chemistries and larger structures but struggled more with unseen topologies.
- FF vs. ML Accuracy: ML models showed significantly lower errors than FFs in the chemisorption regime, with an average error of 0.10 eV for ML compared to 0.49 eV for FFs in this critical region.
Limitations and Future Work
- Computational Cost: While DFT provides high accuracy, it is computationally expensive, limiting the number of MOFs that can be screened.
- FF Limitations: Classical FFs are still useful for initial screening due to their speed but are not accurate enough for definitive selection, especially for chemisorption.
- ML Model Generalizability: While models generalize well to some out-of-domain sets, performance degrades for unseen topologies, indicating room for improvement.
- Future Work:
- Expand Dataset: Incorporating more data, especially for complex chemical environments like open metal sites and defects, will further improve ML model accuracy.
- Transfer Learning: Leveraging knowledge from other large datasets (like those from the Open Catalyst Project) can enhance model performance.
- Tailored Architectures: Developing ML architectures specifically designed for DAC applications, potentially incorporating vdW interaction information, could be beneficial.
- Predicting Isotherms: Extending ML models to predict adsorption isotherms (not just single-point energies) is crucial for realistic DAC process modeling.
- Experimental Validation: The identified promising MOFs will need experimental validation to confirm their suitability for DAC.
Summary updated