An open dataset of gas adsorption in deformable metal-organic frameworks for direct air capture
Authors: 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, David S. Sholl
Organizations: FAIR at Meta · School of Chemical and Biomolecular Engineering, Georgia Institute of Technology · Department of Chemical Engineering, Carnegie Mellon University · CuspAI · University of Tennessee-Oak Ridge Innovation Institute, Oak Ridge National Laboratory
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.
Figures & tables
Figure 1: Magnitude and effects of MOF deformation on DFT adsorption energies. (a) Comparison of original DFT-relaxed empty MOF energies ( Eempty ) to the most favorable MOF energy found across all ODAC25 relaxations ( Eground ) for 3,592 pristine and 2,788 defective ODAC23 MOFs. (b) Comparison of DFT adsorption energies in rigid and flexible MOFs for 976 MOF+adsorbate configurations.
Figure 2: Summary of NX2 and OX2 adsorption statistics. (a) Distribution of adsorption energies for different adsorbates computed using kernel density estimation. (b) Percentage of MOFs that adsorb each adsorbate most strongly, determined by taking the strongest adsorption energy of each adsorbate across all sampled active sites in a given MOF framework.
Figure 3: Two approaches (OMS and linker functionalization) used to generate functionalized MOFs using N-Ethylethylenediamine (een) in IRMOF-74-III (CSD code RAVWAO) as an example. (a) OMS functionalization with a metal-oxygen bond. (b) OMS functionalization with a metal-nitrogen bond. (c) Linker functionalization.
Figure 4: Effect of functionaliation on DFT adsorption energies. (a) Difference in most favorable adsorption energies between functionalized MOFs and their corresponding base MOFs. (b) Kernel density estimation plots for all COX2 and HX2O adsorption energies in linker-functionalized (N = 147) and OMS-functionalized (N = 2,672) MOFs. There are 672 COX2 and 577 HX2O adsorption energies and 11,395 COX2 and 9,472 HX2O adsorption energies in linker- and open metal site-functionalized MOFs, respectively.
Model
Training Set
EMAE-Tot
EMAE
FMAE
EquiformerV2 large [ 19 ]
ODAC23
–
0.155
0.024
MACE [ 82 ]
MPtraj [ 83 ]
–
0.323
0.369
MACE-DAC [ 84 ]
MPtraj [ 83 ] +GoldDAC
–
0.222
0.289
UMA-Small 1.1 [ 85 ]
ODAC25 subset+others
0.282
0.110
0.040
eSEN (Full)
ODAC25-Full
0.195
0.077
0.023
eSEN (Filtered)
ODAC25-Filtered
0.268
0.080
0.033
Table 1: Overall performance of various models on the S2EF task across all adsorbates. Total Energy MAE (EMAE-Tot) and Adsorption Energy MAE (EMAE) are reported in eV, and Force MAE (FMAE) is reported in eV / Å . The best result in each column is shown in bold.
Model
Training Set
CO 2
N 2
UFF
-
0.0299
0.0190
MACE-MP-0b2 [ 82 ]
MPtraj [ 83 ]
0.0566
0.0448
MACE-DAC [ 84 ]
MPtraj [ 83 ] + GoldDAC
0.0513
0.0698
UMA-Small 1.1 [ 85 ]
ODAC25 subset + others
0.0274
0.0631
UMA-Medium 1.1 [ 85 ]
ODAC25 subset + others
0.0145
0.0286
eSEN
ODAC25-Full
0.0250
0.0433
Table 2: Errors of Henry coefficient prediction for different models on CO 2 and N 2 . We show the mean absolute error between predictions and experimental values in eV, which is proportional to the logarithm of the Henry’s constant ( 90 ) . The best result in each column is shown in bold.
Appendix figures & tables23 assets
Supplementary material from the paper’s appendix.
Appendix
Check
ODAC23 pristine [%]
ODAC23 defective [%]
ODAC25 functionalized [%]
All MOFs [%]
has_metal
0.02
0
0
<0.01
has_carbon
4.4
6.6
0
4.1
has_hydrogen
10.4
0
0
3.5
has_atomic_overlaps
0
0
0
0
has_overcoordinated_carbon
0.5
0.03
0
0.2
has_overcoordinated_nitrogen
0
0
0
0
Appendix
Table S1: Error rates from MOFChecker v0.9.6 checks split by MOF types
Figure S1: Approximating higher k-point calculations. (a) Distribution of energy differences from reference calculations with k-point density K=40 Å. (b) Mean absolute energy errors across relaxation frames before and after energy correction. The mean error reduces significantly after correction to ∼0.01 eV.
Parameter
Value
ENCUT
600.000000
POTIM
0.010000
SIGMA
0.200000
EDIFF
1.00e-05
EDIFFG
-5.00e-02
ALGO
NORMAL
Appendix
Table S2: Example INCAR file showing the VASP settings used in creating the ODAC25 dataset. All DFT calculations used VASP version 6.3 [ 115 , 116 , 117 ] with the PBE exchange-correlation functional and D3 van der Waals correction (IVDW = 12). VASP 5.4 PBE pseudopotentials were employed with a plane-wave energy cutoff of 600 eV and electronic convergence of 1e-5 eV. Structural relaxations used the conjugate gradient algorithm with forces converged to 0.05 eV / Å or a maximum of 2000 ionic steps. Bare MOFs were relaxations allowed relaxation of lattice parameters (ISIF = 3), while MOF+adsorbate systems used fixed lattice parameters (ISIF = 2) to maintain consistent unit cells for binding energy calculations. Gaussian smearing ( σ=0.2 eV) was applied with symmetry disabled to allow proper adsorbate-framework relaxation.
Figure S2: Distribution of COX2 and HX2O adsorption energies in pristine and defective ODAC23 MOFs featuring empty MOF re-relaxations and split by adsorbate type(s). Blue histograms are for adsorption energies presented in ODAC23 without MOF re-relaxations; orange histograms are adsorption energies using the global empty MOF reference energy from all ODAC25 configurations with a given MOF.
Figure S3: Water adsorption in the MOF with CSD code COJHIT when the MOF is assumed (a) rigid and (b) free to deform. Co, C, N, O, and H atoms are shown in blue, brown, silver, red, and white, respectively.
Figure S4: Distribution of NX2 and OX2 adsorption energies in pristine and defective MOFs split by adsorbate type.
Figure S5: Histogram of COX2 , HX2O , NX2 , and OX2 adsorption energies in ODAC23 MOFs split by adsorbate.
Figure S6: NX2 chemisorption in MOF with CSD code DEJRUH with Eads = –0.796 eV. The inset shows isosurfaces of zero net charge transfer during chemisorption computed using Bader charge analysis. Yellow and cyan volumes show electron acceptors and donors, respectively. Co, B, C, N, and H atoms are shown in blue, green, brown, silver, and white, respectively.
Table S3: Linkers used to generate functionalized MOFs via linker functionalization.
Diamine Name
Diamine Structure
Abbreviation
Classification
Ethylenediamine
en
1°/1°
N-Methylethylenediamine
nmen
1°/2°
N-Ethylethylenediamine
een
1°/2°
N-Isopropylethylenediamine
ipen
1°/2°
N,N-Dimethylethylenediamine
dmen
1°/3°
N,N-Diethylethylenediamine
deen
1°/3°
Appendix
Table S4: Diamine structures used for open metal site functionalization, spanning primary (1°), secondary (2°), and tertiary (3°) amine classifications with their chemical structures and abbreviations.
Figure S7: Distribution of adsorption energies in functionalized MOFs split by adsorbate type(s).
Figure S8: Histogram of COX2 and HX2O adsorption energies in functionalized MOFs split by adsorbate and functionalization type.
Figure S9: Kernel density estimation for COX2 and HX2O adsorption energies in functionalized MOFs split by diamine used for functionalization.
Figure S10: Minimum COX2 adsorption energies in MOFs functionalized with the specified diamine as a function of diamine concentration. Diamine codes correspond to table S4 .
Figure S11: Two-dimensional histogram showing the distribution of COX2 and HX2O molecule counts in GCMC-generated configurations.
Figure S12: Properties of synthetically generated MOFs (blue) compared to experimental structures (black). Left: density distribution. Middle: distribution of the PLD. Right: two-dimensional UMAP projection, where for legibility we only show 500 experimental MOFs uniformly sampled from the dataset.
Figure S13: Distribution of single-molecule adsorption energies in synthetic MOFs split by adsorbate type.
Hyperparameters
eSEN-ODAC25
Number of parameters
50M
Maximum number of neighbors
30, 300
Cutoff radius (Å)
6
Number of layers
10
Number of sphere channels
128
Number of edge channels
128
Appendix
Table S5: Hyperparameters and training details for the eSEN [ 33 ] model trained on the ODAC25 dataset. Comma-separated values indicate pre-training, post-training parameters. The eSEN models were trained in two stages: first, a direct model with a maximum of 30 neighbors, and, subsequently, an energy-conserving model with up to 300 neighbors was trained. We limit the fine-tuning stage to structures with less than 350 atoms to reduce the GPU memory usage. Detailed descriptions of the eSEN model and architecture details are described in Fu et al. [ 33 ] .
Model
COX2
HX2O
COX2 + HX2O
NX2
OX2
EMAE ↓
FMAE ↓
EMAE ↓
FMAE ↓
EMAE ↓
FMAE ↓
EMAE ↓
FMAE ↓
EMAE ↓
FMAE ↓
EquiformerV2 large [ 19 ]
0.117
0.025
0.232
0.025
0.129
0.024
0.219
0.036
0.242
0.048
MACE [ 82 ]
0.219
0.382
0.168
0.392
0.367
0.355
0.550
0.420
0.209
0.456
MACE-DAC [ 84 ]
0.106
0.301
0.149
0.317
0.252
0.272
0.380
0.322
0.284
0.368
UMA-Small 1.1 [ 85 ]
0.074
0.041
0.082
0.042
0.103
0.037
0.386
0.044
0.157
0.060
eSEN (Full)
0.061
0.024
0.075
0.022
0.081
0.024
0.081
0.028
0.068
0.044
Appendix
Table S6: Performance breakdown of various models on the S2EF task by individual adsorbate configurations. Adsorption Energy MAE (EMAE) is reported in eV and Force MAE (FMAE) is reported in eV / Å . The UMA model was trained on a subset of ODAC25, while the two eSEN models were trained on the full and filtered versions of ODAC25. The best result in each column is shown in bold font.
Figure S14: Log-log scatterplot of the Henry coefficient from experiment and from Widom insertion for the UFF baseline and different MLFFs, with the adsorbates COX2 and NX2 . We show the mean absolute error between predictions and experimental values in eV, which is proportional to the logarithm of the Henry’s constant ( 90 ) and the standard deviation of the predicted coefficient..
MOF
Adsorbate
Temp (K)
Henry coeff. (mol/kg/Pa)
Henry coeff. (eV)
CALF20
COX2
293
3.76e-04
0.199
CALF20
COX2
298
6.10e-04
0.190
CALF20
COX2
303
2.14e-04
0.221
CALF20
NX2
293
8.90e-06
0.294
CALF20
NX2
298
3.39e-06
0.323
CALF20
NX2
303
3.03e-06
0.332
Appendix
Table S7: Experimental Henry coefficients reported in mol/kg/Pa and eV.
Model
Adsorption energy
Interaction energy
MOF deformation energy
eSEN-ODAC25-Full
0.125
0.118
0.096
eSEN-ODAC25-Filtered
0.087
0.103
0.054
CHGNet [ 83 ]
0.124
0.163
0.065
EqV2-ODAC23 [ 19 ]
0.191
–
–
Appendix
Table S8: Mean absolute errors in eV for energies relative to DFT calculations from ref. [ 39 ] . CHGNet and EqV2 results are adapted from that work for comparison.
Figure S15: Confusion matrices for determining MOF deformation class using (a) eSEN-Full and (b) eSEN-Filtered models for the 59 MOF+adsorbate systems from ref. [ 39 ] . True positives and true negatives are located in the upper left and lower right of each matrix, respectively.
School of Science and Engineering (SSE), The Chinese University of Hong Kong, Shenzhen (CUHKSZ), Shenzhen, China. · School of Data Science (SDS), The Chinese University of Hong Kong, Shenzhen (CUHKSZ), Shenzhen, China.
Department of Chemical Engineering, Stanford University, Stanford, California 94305, United States · SUNCAT Center for Interface Science and Catalysis, SLAC National Accelerator Laboratory, Menlo Park, California 94025, United States
Department of Chemical and Biomolecular Engineering, Korea Advanced Institute of Science and Technology, Daejeon 34141, Republic of Korea. · Department of Materials Science and Engineering, Korea Advanced Institute of Science and Technology, Daejeon 34141, Republic of Korea.