An open dataset of gas adsorption in deformable metal-organic frameworks for direct air capture
summary
The gist
Identifying useful sorbent materials for direct air capture (DAC) from humid air remains a challenge.
In short
ODAC25 expands computational modeling for direct air capture by providing a massive dataset of gas adsorption in 15,000 MOFs. It includes new adsorbates like N2 and O2, functionalized materials, and GCMC configurations. Training machine learning models on this data improves predictions for energy, forces, and selectivity in complex systems.
Key concepts
- ODAC25 Dataset
- This is a large collection of nearly 60 million Density Functional Theory (DFT) calculations detailing the adsorption of CO2, N2, O2, and H2O across 15,000 Metal-Organic Frameworks (MOFs). It significantly increases the scale and diversity of existing computational data for sorbent discovery.
- Machine Learning Force Fields (MLFFs)
- These are advanced AI models like EquiformerV2 and eSEN trained on the ODAC25 dataset. They learn to predict adsorption energies and forces accurately, allowing researchers to screen vast chemical spaces much faster than traditional simulation methods.
- Deformable MOF Modeling
- This involves modeling how the metal-organic framework structure changes or flexes when gas molecules adsorb onto it. The models trained on ODAC25 can account for this structural movement, which is crucial because rigid models often fail to find the best sorbents.
- Henry’s Coefficients
- These coefficients are used to evaluate how well the machine learning predictions match experimental data for gas adsorption. By calculating these using the Widom insertion method, researchers can assess the model's accuracy in predicting selectivity and performance under various pressures.
Terminology used across episodes
This episode discusses
- An open dataset of gas adsorption in deformable metal-organic frameworks for direct air capture · Paper Radio
- EquiformerV2: Improved Equivariant Transformer for Scaling to Higher-Degree Representations
- Learning Smooth and Expressive Interatomic Potentials for Physical Property Prediction
- UMA: A Family of Universal Models for Atoms
- Comparing classical and machine learning force fields for modeling deformation of solid sorbents relevant for direct air capture
- A foundation model for atomistic materials chemistry
The paper
An open dataset of gas adsorption in deformable metal-organic frameworks for direct air capture · Read on arXiv
Anuroop Sriram, Logan M. Brabson, Xiaohan Yu, Sihoon Choi, Kareem Abdelmaqsoud, Elias Moubarak, Pim de Haan, Sindy Löwe, Johann Brehmer, John R. Kitchin, Max Welling, Á. Lawrence Zitnick, Áhar Zachary Ulissi
FAIR at Meta School of Chemical and Biomolecular Engineering Georgia Institute of Technology University of Tennessee-Oak Ridge Innovation Institute Oak Ridge National Laboratory CuspAI Department of Chemical Engineering Carnegie Mellon University
Engineering economical and scalable processes for direct air capture (DAC) from humid air relies on atomistic modeling to search and identify promising sorbent materials from a vast chemical space. Machine learning has emerged as a powerful tool for predicting material properties in an efficient yet accurate manner, but these models require large sets of carefully curated training data. We present the Open DAC 2025 (ODAC25) dataset comprising nearly 60 million DFT single-point calculations for CO2, H2O, N2, and O2 adsorption in 15,000 MOFs. ODAC25 introduces functionalized MOFs, high-energy adsorbate placements, and synthetically generated frameworks to maximize the versatility of models trained on the dataset. ODAC25 improves upon the accuracy of DFT calculations in previous works and applies engineering intuition from recent work regarding the treatment of adsorbate-induced MOF flexibility. We release state-of-the-art machine-learned interatomic potentials trained on ODAC25 and demonstrate their utility via Henry's coefficient predictions. ODAC25 enables MOF screening for DAC at a resolution previously unattainable.
DOI: 10.1038/s44286-026-00445-5
Transcript
Introduction to the show: ident: Condensed Matter. The day in condensed matter: materials, phases and the measurements under them.
Kai: We're looking at the paper: "An open dataset of gas adsorption in deformable metal-organic frameworks for direct air capture".
Title and authors: Kai: So, we're looking at this paper today, "An open dataset of gas adsorption in deformable metal-organic frameworks for direct air capture," and it’s about expanding what we can do with these materials. We're going to talk about how they built this new dataset and what that means for finding better sorbents.
Mira: Right, it’s a big push forward because they are tackling the challenge of identifying useful sorbent materials for direct air capture from humid air, which is super relevant right now. They show how chemical and structural diversity really matters when you're looking at this problem.
Lev: And what they’re showing us is that we need to look beyond just the basic MOFs we've seen before; they are bringing in a lot more complexity than before.
Kai: Exactly, they introduced functionalized MOFs, and they used these high-energy Grand Canonical Monte Carlo simulations to generate placements that were really high energy, which means those are configurations we might not have seen experimentally or in simpler databases.
Mira: That’s interesting because those GCMC derived placements give the models a lot more situations to learn from, which is crucial for modeling real-world adsorption processes like what happens in DAC.
Lev: And they also synthesized new frameworks, which extends the structural diversity beyond just what was found in existing experimental databases, giving the AI something completely new to train on.
Kai: They've also done a lot of heavy lifting on data quality control too; they performed validation checks using MOFChecker v0 point 9.six and corrected for systematic errors by re-relaxing each bare MOF structure to account for adsorbate deformation, which is a necessary step for good DFT calculations <ref:2508.03162#pg1>.
Mira: I noticed they also improved the convergence in reciprocal space by using a correction method that reduced energy errors due to low k-point density to about zero point zero one electron volts across all frames, which is important for getting reliable DFT results.
Lev: That level of accuracy on the DFT side is what allows the machine learning force fields, or MLFFs, to be trained effectively in the first place.
Kai: And they expanded this dataset from ODAC23 to ODAC25 substantially; it now has nearly sixty million DFT single-point calculations for CO2, H2O, N2, and O2 adsorption across fifteen thousand MOFs.
Title and authors: Mira: That expansion is significant because it moves us from just modeling carbon dioxide and water vapor to including nitrogen and oxygen adsorption as well, which is what you actually get in the air.
Lev: Including N2 and O2 changes the problem a lot because modeling those interactions adds another layer of complexity to how the sorbent behaves when dealing with a real gas mixture.
Kai: They also introduced two new adsorbates, N2 and O2, specifically to enhance coverage of situations relevant to modeling DAC, which directly addresses one of the key limitations they found in ODAC23.
Mira: And it’s not just about adding molecules; they’ve also looked at functionalized MOFs with both linker and open metal site functionalization, which brings in more chemical diversity than before.
Lev: That combination of structural diversity and chemical functionalization is what allows the AI to screen a significantly broader space for novel sorbents, which is what we need for real discovery.
Kai: They also incorporated those high-energy multi-component adsorption configurations derived from GCMC simulations, generating over two point seven million single-point DFT calculations, which really boosts the scale of what the models are learning from.
Mira: So it’s not just about having more data points; it’s about capturing the full range of chemical and structural possibilities relevant to how these materials interact with air.
Lev: And this leads into how they trained their machine learning force fields, which is a major part of what makes this paper important for acceleration.
Kai: They release a suite of state-of-the-art MLFFs, like EquiformerV2 and eSEN, all trained on this ODAC25 dataset to optimize a loss function that balances energy and force losses against the total DFT energy.
Mira: The fact that they used the total DFT energy as the training target is smart because it keeps the training focused on what we care about most for predicting actual adsorption.
Lev: And I see eSEN models trained on this filtered ODAC25 dataset achieve very low force mean absolute errors of zero point zero three two electron volts per angstrom, which is quite strong for a force field prediction.
Kai: When they benchmarked these MLIPs against the Structure to Energy and Forces task, the eSEN models showed lower energy and force mean absolute errors on all metrics compared to prior models they tested.
Title and authors: Mira: They also used the Widom insertion method to calculate Henry’s coefficients, which is a physical way to evaluate how well these machine learning force fields predict experimental values for CO2 and N2 adsorption.
Lev: The best performing model for CO2 adsorption among all ODAC25 models was identified as UMA-Medium one point one, which gives us a concrete example of what's working well <ref:2508.03162#pg1>.
Kai: Beyond just the CO2 prediction, the eSEN models met or outperformed every model tested in a benchmark study for predicting the overall adsorption energy in deformable MOFs, achieving zero point zero one electron volts of adsorption energy at that benchmark.
Mira: That result is compelling because it shows that these MLFFs are not just good at predicting single interactions; they can handle the complexity of the entire framework's flexibility during adsorption.
Lev: I mean, when they used models trained on ODAC25, it allows all the MOF atoms to move during relaxation with the adsorbate molecules, which solves a problem where rigid framework assumptions often failed to identify viable materials before.
Kai: So what they’re showing us here is that by including framework flexibility in the training data and the models themselves, we can get much more reliable predictions for how these materials will actually perform under real-world conditions.
Mira: It really underscores how important it is to move away from just rigid assumptions when we are trying to find practical materials for something like direct air capture.
Lev: And this whole effort, using the ODAC25 dataset, is designed to empower the community by making both the dataset and models publicly available so other researchers can use them.
Kai: So, in a nutshell, ODAC25 is a massive step forward in computational sorbent discovery for direct air capture because it's much bigger and more chemically diverse than before.
Mira: The improved machine learning force fields demonstrate better performance on both the older ODAC23 test sets and these new functionalized MOFs, which really shows the impact of this expanded data.
Lev: By making the dataset, models, and tools publicly available, they’re aiming to help accelerate the development of scalable and energy-efficient sorbents for carbon dioxide removal in a way that's accessible to everyone.
Kai: That’s what it is; ODAC25 is a significant step forward in computational sorbent discovery for direct air capture, and it’s setting a new standard for how we approach this kind of material science problem.
The paper's summary: Kai: So, basically, this paper is about building a massive open dataset for gas adsorption in metal organic frameworks that can actually move or deform when you put stuff on them.
Mira: Right, it’s about taking the knowledge from earlier work and making it much bigger and more diverse for direct air capture applications.
Kai: They’ve taken the old data set, ODAC23, and they expanded it into ODAC25. We're talking nearly sixty million DFT calculations now for CO two hydrogen two oxygen and nitrogen two adsorption across fifteen thousand different metal organic frameworks.
Mira: That expansion is huge because it means we’re not just looking at carbon dioxide and water vapor anymore; we’ve got nitrogen and oxygen included, which is what you actually get in the air.
Kai: And they didn't just add molecules; they added structural diversity too, including functionalized frameworks with metal sites on both ends of the linker and even some frameworks that were built from scratch synthetically.
Mira: That means the AI isn't just learning about simple structures; it’s learning about materials that are chemically tweaked in different ways, which helps us find novel sorbents.
Kai: They also included configurations from Grand Canonical Monte Carlo simulations, which are these high energy placements that aren't easily found in regular experiments.
Mira: So the dataset is really covering a much wider range of chemical and structural possibilities for how these materials interact with air molecules under real conditions.
Kai: And they didn't just stop at the data; they trained new machine learning force fields, like EquiformerV2 and eSEN, on this ODAC25 data.
Mira: The way they trained them is important because they balanced the energy losses and the force losses against each other using the total DFT energy as their main target.
Kai: The results show these models perform better than before on both those older test sets and these new functionalized MOFs, with lower errors in both energy and force predictions.
Mira: They even used a physical way to check accuracy by calculating Henry’s coefficients using the Widom insertion method against experimental values for CO two and nitrogen two.
Kai: And the best performing model they found for CO two adsorption was UMA Medium one point one, which gives us a specific example of what’s working well.
Mira: But the real strength here is that these eSEN models meet or even beat every other model tested in a benchmark study for predicting energy in deformable metal organic frameworks.
Kai: That means they can handle the flexibility of the framework during adsorption, which is something rigid framework assumptions often failed to do before.
Mira: So, this isn't just about having more data points; it's about building models that can accurately predict how these materials will actually perform when they are flexible and interacting with real gas mixtures.
Kai: This whole effort, making the dataset and the models public, is designed to give other researchers a tool to speed up the discovery of sorbents for carbon dioxide removal.
Mira: It’s a big step toward making this computational discovery process more accessible to everyone working on materials science problems.
Kai: So they've given us a much bigger playground, and now we have better tools to explore it, which leads us into how these models can predict selectivity when dealing with those complex mixtures.
The paper's improvements: Kai: So, we’re looking at how they suggest improving the ODAC25 dataset itself now that they have these MLFFs trained on it.
Mira: Right, they’re talking about making the data even better by adding more diversity in a specific way to help the AI learn faster.
Kai: They are suggesting incorporating those high energy multi component adsorption configurations from GCMC simulations into the training set more heavily.
Mira: That makes sense because those are these extreme situations that we need to model for realistic DAC applications, and it broadens the chemical space the AI can cover.
Kai: They also mentioned adding more structural diversity by including functionalized MOFs with both linker and open metal site functionalization in a way that’s better suited for modeling adsorption.
Mira: That means they want to ensure the models don't just learn from one type of structure; they need to see the whole spectrum of how different chemical groups affect performance.
Kai: They are also focusing on how we can improve the prediction of selectivity using methods like Widom insertion, making sure those calculations are more representative across a wider range of pressures.
Mira: It’s about moving past just getting a single good answer for CO two and figuring out how the sorbent will behave under varied conditions, which is key for real-world use.
Kai: They also pointed out that they need to keep focusing on the physics of framework flexibility, making sure the models are really learning how all those MOF atoms move during adsorption.
Mira: That ties back to what we talked about earlier; if the model doesn't handle the movement correctly, all that data is just noise for real applications.
Kai: They essentially suggest a continuous loop where better data generation feeds better models, and then better models help us generate even better configurations.
Mira: It’s like they’re telling us we need to keep pushing the boundaries on what goes into these datasets if we want the AI to truly find those highly efficient materials for DAC.
Kai: So, this isn't just about the initial calculation; it's about designing a better data pipeline that captures more complex physics and structural interactions.
Mira: And that’s exactly where we need to focus next: how do these improved MLFFs actually translate into predicting selectivity ratios accurately when you have a real gas mixture present?
Conclusion: Kai: So to wrap up, this paper presents ODAC25, an open dataset of gas adsorption in deformable metal organic frameworks for direct air capture that’s significantly bigger and more diverse than what we had before.
Mira: Right, it’s a big step forward because it gives the community a massive resource to train machine learning models on for finding better sorbents.
Kai: They show that these updated machine learning force fields actually perform better on both old and new test sets, which means we have more reliable predictions now.
Mira: And they've shown that these models can handle the flexibility of the framework during adsorption, which is something rigid assumptions just couldn't do before.
Kai: So, this work really helps us move from just theoretical predictions to having tools that can actually model how these materials behave in a complex real-world environment.
Mira: It opens up a lot of possibilities for screening novel materials quickly instead of doing slow and expensive experimental work for every single candidate.
Kai: Lev, what does this mean when you think about running these kinds of force fields on actual hardware for error correction purposes?
Lev: Well, the models are getting better at predicting the energy and forces, but scaling those complex simulations up to a full quantum error correction routine still lands on a pretty heavy computational overhead.
Kai: That’s true; the fidelity is there in theory, but moving from a single point calculation to real hardware implementation adds another layer of complexity for testing.
Mira: But the point here is that having a high quality initial dataset and force field means we can spend less time on low-fidelity simulations and more time focusing on the actual physical chemistry of adsorption.
Lev: For us, it’s about building better benchmarks; if you can't reliably predict the energy of a flexible MOF, you can't design a reliable quantum system to test it against.
Kai: So, this paper with ODAC25 is essentially providing the necessary foundation—the data and the models—to accelerate that whole process for carbon capture materials.
Mira: It’s a really important step toward making sure that when we eventually find these sorbents, they are truly optimized for the real chemistry of humid air.
Lev: Exactly; it moves us past just theoretical constructs and gives us something concrete to work with that mimics the complexity of a real system.
Kai: So, ODAC25 is now available for everyone to use, which is a huge thing for pushing this field forward.
Mira: It’s a significant contribution because it makes the discovery process more scalable and less reliant on just traditional experimental synthesis routes.
Lev: We look forward to seeing how these improved force fields handle multi component mixtures in the next phase of testing.
More episodes
- 2610.32678-Probing the limits of materials design via a comprehensive crystal database of 14.6 million structures with 140 properties
- 2610.33595-Native-oxide-driven layer asymmetry, two-step switching and exchange bias in Co/Ir/Co trilayers
- 2503.19572-Spin models from nonlinear cellular automata
- 2606.23350-Irreversibility Enhances Quantum-Enhanced Markov-Chain Monte Carlo
- 2608.02744-Two-Parameter Ansatz for the Violation of Eigenstate Thermalization
- 2210.00806-Low- and high-energy localization landscapes for tight-binding Hamiltonians in two-dimensional lattices
- 2409.09625-Selective switch for band-edge alignments in pentagonal CdSeTe monolayer: Atom-valley locking
- 2606.20168-Norms, overlaps and Yangian descendants for the Haldane-Shastry spin chain
- 2512.19503-Strain-induced splitting of the CCDW-NCCDW phase transition in 1T-TaS 2
- 2307.12638-Strong localization of microwaves beyond two dimensions in aperiodic Vogel spirals