An open dataset of gas adsorption in deformable metal–organic frameworks for direct air capture
Anuroop Sriram, Logan M. Brabson, Xiaohan Yu, Sihoon Choi, Kareem Abdelmaqsoud, Elias Moubarak, Pim de Haan, Sindy Löwe, Johann Brehmer, Muhammed Shuaibi, John R. Kitchin, Max Welling, C. Lawrence Zitnick, Zachary Ulissi, Andrew J. Medford, and David S. Sholl, “An open dataset of gas adsorption in deformable metal–organic frameworks for direct air capture”, Nature Chemical Engineering (2026). https://doi.org/10.1038/s44286-026-00445-5
This summary was written by a language model from the arXiv preprint (“The Open DAC 2025 Dataset for Sorbent Discovery in Direct Air Capture”) and has not been checked line by line. The published version may differ; rely on the published paper for details.
Question
Direct air capture (DAC) needs sorbents that take up CO2 at about 400 ppm from humid air and can be regenerated cheaply. Metal–organic frameworks (MOFs) are highly tunable candidates, and computational screening can narrow the search, but only if the underlying calculations are accurate and cover the relevant chemistry. The earlier Open DAC 2023 (ODAC23) dataset covered only CO2 and H2O, did not include functionalized frameworks, and referenced adsorption energies to a relaxed empty MOF that was not always the lowest-energy state of the framework. The authors set out to build a larger, more accurate and more diverse dataset (ODAC25) that treats framework deformation consistently, and to train machine-learned interatomic potentials (MLIPs) on it (Sriram et al. 2026).
Approach
All calculations are spin-polarized PBE+D3. ODAC23’s single k-point was replaced with a grid scaled to the cell size, and the energies of each relaxation trajectory were corrected for k-point error using the first and last frames. To account for flexibility, the empty MOF was re-relaxed starting from every converged MOF+adsorbate geometry, and the lowest-energy empty framework found was used as the reference for adsorption energies; the ODAC23 adsorption energies were recomputed the same way (Sriram et al. 2026).
The dataset was extended in four directions. N2 and O2 were added, with about 56,000 relaxations over about 6,400 pristine and defective ODAC23 MOFs. Linker amine functionalization and diamines grafted to open metal sites were generated, including starting geometries for ammonium carbamate and carbamic acid formation, some with co-adsorbed water. Grand canonical Monte Carlo (RASPA, UFF, 300 K, 5 to 50 kPa, pure and mixed CO2/H2O) supplied higher-energy, multi-adsorbate configurations evaluated as DFT single points. Finally, 460 synthetic MOFs were chosen by farthest point sampling from about 4,000 generative-model candidates that passed MOFChecker and geometric filters. Structures flagged by MOFChecker were kept in a “full” dataset, and a “filtered” version is also provided (Sriram et al. 2026).
eSEN models (trained on the full and filtered data) and UMA were compared against EquiformerV2 from ODAC23 and MACE variants on energies, forces and adsorption energies. Henry’s law coefficients were computed by Widom insertion in rigid frameworks and compared with experiments for UiO-66, HKUST-1 and MOF-5 (Sriram et al. 2026).
Key findings
ODAC25 contains nearly 60 million DFT calculations across about 15,000 MOFs, compared with 38 million across 8,400 MOFs in ODAC23, including more than 2.7 million GCMC-derived single points and 52,756 adsorbate placements in 3,920 functionalized MOFs. About 7% of ODAC23 calculations had k-point errors above 0.2 eV; the correction brought convergence errors to about 0.01 eV. Re-relaxing the empty framework shifted the median adsorption energy by more than 0.1 eV toward weaker binding in every split, showing that ignoring framework deformation systematically overstates binding. Of the single-adsorbate configurations, 63.5% fall in the physisorption range (−0.5 to 0 eV), 5.7% (8,320) are chemisorbed, and 30.8% are unfavorable (Sriram et al. 2026).
The eSEN model trained on the full dataset reached an adsorption energy MAE of 0.077 eV and a force MAE of 0.023 eV/Å, against 0.155 eV for the ODAC23 EquiformerV2 model. For Henry’s coefficients, the best MLIPs had errors (about 0.015 eV for CO2 and 0.014 eV for N2) comparable to the spread among experimental measurements and roughly half those of UFF. On a benchmark of 59 MOFs that deform on adsorption, the filtered eSEN model gave an adsorption energy MAE of 0.087 eV, but both eSEN models had trouble identifying which frameworks deform significantly (Sriram et al. 2026).
Limitations
PBE+D3 without a Hubbard U and without systematic spin-state enumeration is unreliable for some redox-active metal sites, so the authors advise caution with the O2 results. Only a few adsorbate placements were sampled per MOF in many cases, adsorbate placement did not target functional groups, and some functionalized and generated structures may not be synthesizable. Henry’s coefficients assume rigid frameworks. The dataset is released on Hugging Face (facebook/ODAC25) and the models through fairchem (Sriram et al. 2026).