The Open Catalyst 2022 (OC22) Dataset and Challenges for Oxide Electrocatalysts

Motivation

• OER is especially challenging and energy intensive, with few stable, active, and affordable catalysts under acidic conditions; multi-component oxides are being pursued to improve activity, stability, and cost.1 2,3Computational chemistry can help screen, discover, and understand these materials.Source: Introduction · p. 2
• Oxide electrocatalyst design is challenging because surfaces undergo reconstruction, leaching, doping, defects, and varied terminations, while correlated electrons and long-range interactions make calculations difficult.These configurational and computational complexities make oxide datasets smaller and more expensive than metal-system datasets.Source: Introduction · p. 2

Dataset and tasks

• OC22 introduces an oxide-focused dataset for OER and electrocatalysis that complements OC20, covering surface terminations, adsorbate coverage, substitutions, vacancies, and other complex configurations.The dataset is designed to enable more generalizable machine-learning models for catalysis.Source: Introduction · p. 3
• OC22 expands beyond adsorption energies by making DFT total energy a target, enabling applications that require a more general property prediction task.The work also introduces challenges for predicting total energy and force, relaxed total energy, and relaxed structures from initial structures.Source: Introduction · pp. 3–4

Transfer learning and benchmarks

• The work tests whether OC20 can improve OC22 through transfer learning or joint training, building on prior transfer-learning successes in catalysis and related scientific domains.28 29 30–32It trains leading GNNs and provides train/validation/test splits, baseline models, data loaders, and training scripts for the proposed tasks.Source: Introduction · pp. 3–4

The OC22 Dataset

• OC22 contains 19,142 slabs, 43,189 adsorbate+slabs, and 9,854,504 single-point calculations generated through bulk selection, surface construction, structure generation, and relaxation.Forces and energies were partitioned into train, validation, and test splits to support generalized model development.Source: The OC22 Dataset · p. 4
• The dataset prioritizes chemical, surface-termination, and adsorbate diversity over stability, so some structures may be electrochemically unstable or less relevant to particular reaction pathways.Only oxides composed of 26 of the 51 considered elements were previously shown to be relatively stable under aqueous conditions.Source: The OC22 Dataset · p. 4, Bulk selection · pp. 4–5

Bulk selection

• Bulk selection began with 4,728 unary and binary oxides, retaining up to five low-energy-above-hull structures per chemical system and including varied band gaps plus 173 rutile structures.The selection used Materials Project entries, 51 metals, and structures with fewer than 150 atoms.Source: Bulk selection · p. 4

Surface selection

• Surface construction randomly sampled 4,286 bulk oxides, enumerated terminations with Miller index ≤ 3, and introduced symmetric oxygen vacancies to model potential reactive sites without dipole artifacts.Slabs were limited to fewer than 250 atoms and replicated to at least 8 Å depth and cross-sectional width.Source: Surface selection · p. 5

Initial Structure Generation

• Adsorbate generation sampled nine species, including OER intermediates and additional H*, O*, N*, C*, and CO*, with binding allowed at surface oxygen, under-coordinated metals, or vacancies.Placement strategies also incorporated oxide-specific intermediate formation, including dimers, CO2*, and OOH* through reactions with surface oxygen.Source: Initial Structure Generation · pp. 5–6

Structure Relaxation

• Relaxation used PBE-GGA, Hubbard U corrections for selected transition-metal oxides, spin-polarized calculations, and relaxation of all slab and adsorbate atoms.One magnetic polymorph was considered per slab, initialized as ferromagnetic or nonmagnetic according to suggested metal moments; calculations used VASP.Source: Structure Relaxation · pp. 6–8

Task modifications

• OC22 replaces adsorption-energy targets with DFT total-energy targets for S2EF and IS2RE because unconstrained surfaces make OC20’s clean-surface reference ill-posed and noisy.IS2RS remains unchanged.Source: Tasks · p. 8

Task definitions

• S2EF-Total predicts DFT total energy and per-atom forces, while IS2RE-Total predicts relaxed DFT total energy; both support isolated surfaces and adsorbate–surface structures.All structures may contain a surface with an adsorbate or an isolated slab, and targets are computed with DFT.Source: Tasks · pp. 8–9

Physical interpretation

• Total-energy predictions let users reference energies flexibly, including adsorption energies from separate adsorbate–slab and clean-surface predictions.This is particularly important for identifying stable OER surface coverages or terminations.Source: Tasks · p. 9

Task scope

• S2EF-Total is the most general task and closest to a DFT surrogate, enabling surface-stability studies and use of surface trajectories previously unusable with OC20 references.These capabilities support catalyst-discovery workflows beyond adsorption-energy screening.Source: Tasks · p. 9

Graph construction

• Graphs represent atoms as nodes and interactions as edges, incorporate periodic boundary conditions, and use cutoff radii for computational tractability.Unless otherwise noted, edges use a 6 Å cutoff and at most 50 neighbors per atom, while GemNet-OC uses interaction-specific cutoffs.Source: Baseline GNN Models · pp. 9–10

Benchmark scope

• S2EF-Total benchmarks seven models, while IS2RS uses SpinConv, GemNet-dT, and GemNet-OC, and IS2RE-Total uses SchNet, PaiNN, DimeNet++, and GemNet-dT.Top-performing S2EF-Total models are also evaluated for IS2RE-Total through iterative relaxations.Source: Baseline GNN Models · p. 9

Model architectures

• The baselines span continuous edge filters, directional message passing, direct force prediction, equivariance, symmetric message passing, and efficient quadruplet interactions.GemNet-OC extends GemNet-dT and is described as the state-of-the-art model across OC20 tasks.Source: Baseline GNN Models · pp. 9–10

Training objectives

• S2EF-Total models balance energy and force losses, using p = 2 for GemNet-dT and GemNet-OC and p = 1 for other models, whereas IS2RE-Total evaluates energy only.Baseline S2EF-Total training sets λE = 1 and λF proportional to N^2 to preserve size invariance.Source: Baseline GNN Models · p. 10

Evaluation Metrics

• OC22 evaluates total-energy and force tasks with Energy MAE, Force MAE, Force cosine, and threshold metrics, while IS2RE-Total and IS2RS use energy, structure-distance, and force-threshold metrics.The metrics follow OC20, but OC22 uses DFT total energies rather than adsorption energies as ground-truth targets; models are also expected to be substantially faster than DFT.Source: Evaluation Metrics · pp. 10–11

ML Experiments

• The experiments compare OC22-only training with joint OC20+OC22 training and OC20-pretrained fine-tuning on OC22, including multiple OC20 sizes and OC22 fractions for S2EF-Total.Joint training uses DFT total-energy targets for OC20, whereas fine-tuning updates all weights from public OC20 adsorption-energy checkpoints.Source: ML Experiments · p. 11, Figure 5 · p. 11

Results

• GemNet-OC performs best across S2EF-Total metrics, significantly outperforming other models on energy-based metrics, plausibly through its 12 Å cutoff and explicit quadruplet interactions.S2EF-Total predicts DFT total energy and per-atom forces; GemNet-dT remains competitive on force metrics.Source: Results · p. 12
• OC22 out-of-domain performance is substantially worse than in-domain performance, indicating that generalizing total energies across unseen material-species combinations is particularly challenging.The authors relate this difficulty to total-energy prediction being more complex than referenced adsorption-energy prediction and report poor analogous OC20 results in the SI.Source: Results · pp. 12–13
• Joint training improves GemNet-OC energy and force predictions, while fine-tuning generally improves GemNet-OC and GemNet-dT energy MAE; GemNet-OC fine-tuned on larger capacity improves in-domain energy and force predictions but has little or negative out-of-domain impact.Increasing OC20 data helps GemNet-OC more consistently than PaiNN or SpinConv, and fine-tuning requires careful learning-rate and hyperparameter selection.Source: Results · pp. 13–14
• For IS2RE-Total, GemNet-OC reaches EwT values of 3.89% with relaxation and 3.66% directly; fine-tuning helps relaxation, joint training hurts, and relaxation generalizes better out of domain.Relaxation-based methods are 200–300x slower than direct prediction, and all approaches retain substantial room for improvement.Source: Results · pp. 16–17
• For IS2RS, GemNet-OC outperforms other models across metrics, while joint training and fine-tuning improve DFT force metrics; GemNet-OC-Large achieves the best force metrics, but practical accuracy remains insufficient.Distance-based ADwT correlates poorly with practical DFT metrics, whereas FbT and AFbT indicate substantial remaining error.Source: Results · p. 17, Table 6 · p. 17
• Joint OC20+OC22 training improves energy MAE by approximately 20%, while OC22 predictions show at least 75% linear correlation with literature adsorption energies, with most within 0.6 eV.The fitted OH*/OOH* scaling slope is 0.73 and intercept 3.44, consistent within 0.05 eV with literature values of 0.73 and 3.49. Total-energy adsorption predictions remain far from chemical accuracy of approximately 0.1 eV.Source: Results · pp. 17–19

Discussion

• 0.24 eV ID and 0.94 eV OOD energy MAEs on S2EF-Total, rising to 1.12 eV ID and 1.85 eV OOD for relaxed total energies, remain insufficient for OER overpotential prediction.The models may nevertheless help with more limited tasks, including accelerating oxide calculations through online fine-tuning.Source: Discussion · p. 20
• OC22’s total-energy tasks move beyond adsorption-energy prediction toward general DFT surrogates capable of modeling surface stability and other surface properties.This broader formulation is especially important for oxide catalysis, where comparing surface termination stability is necessary.Source: Discussion · p. 20

Future directions

• Long-range electrostatic and magnetic interactions challenge local GNN assumptions because oxide semiconductors contain variable partial charges and spin interactions with effects extending over long distances.More message-passing steps or larger local cutoffs, alongside other long-range-interaction methods, may improve OC22 predictions.Source: Future directions · p. 21
• Future models should incorporate higher-level theory, magnetic moments, charge or doping descriptors, and Fermi-level effects to capture oxide-specific energetic, structural, and magnetic behavior.Spin configurations can produce distinct relaxation trajectories and oxidation states, while electron-count changes can cause remote structural reconstructions through self-compensation.Source: Future directions · p. 21
• OC22 models could enable low-cost adsorption, reaction, surface-energy, phase-diagram, Pourbaix, and crystal-shape calculations, while future validation must account for solvation, evolving catalyst structures, and experimental conditions.Explicit oxygen vacancies may support studies of dissolution-sensitive oxide catalysis and vacancy-mediated mechanisms such as Mars–van Krevelen reactions; experiments and auxiliary methods can strengthen validation.Source: Future directions · pp. 22–23
• Joint OC20–OC22 training unexpectedly improves energy prediction and yields the most accurate models, while limited fine-tuning does not substantially improve the accuracy–cost Pareto front.These results support learning shared representations from diverse datasets and motivate broader dataset combination and improved fine-tuning strategies.Source: Future directions · p. 22

Supporting Information Available

• The supporting information contains extended benchmark results, alternative reference schemes, adsorption-energy analyses, validation studies, baseline training details, Gibbs adsorption-energy corrections, and Hubbard U corrections.It also reports performance across OC22 adsorbate-plus-slabs and slabs evaluated independently.Source: Supporting Information Available · p. 23
• The full dataset is available through the Open Catalyst Project in ASE trajectory and model-ready LMDB formats, with baseline models, dataloaders, and trainers provided in an open-source repository.Resources are provided at http://opencatalystproject.org and https://github.com/Open-Catalyst-Project/ocp.Source: Supporting Information Available · p. 23

• OOD metrics are significantly worse than ID metrics for S2EF-Total, mirroring the corresponding OC22 trend.Source: OC20 S2EF-Total and IS2RE-Total results · p. 32
• Total-energy prediction is more challenging because models must capture subsurface, surface, and adsorbate interactions, whereas adsorption references subtract slab energy and isolate the adsorbate–surface interface.Source: OC20 S2EF-Total and IS2RE-Total results · p. 32
• S2EF-Total performs worse than S2EF across all models and validation splits, while IS2RE-Total likewise performs worse than IS2RE.Source: Table 9 · p. 32, OC20 S2EF-Total and IS2RE-Total results · p. 32, Table 10 · p. 33

Method

• Linear referencing fits per-element energies with linear regression and uses the fitted coefficients to transform OC22 total-energy training targets.For system i, the model fits per-element coefficients from element counts and DFT total energies before referencing targets using those coefficients.Source: Alternative reference scheme · p. 34

Results

• 4.5%, 49.9%, and 24.5% ID energy improvements were achieved with linear referencing for GemNet-OC, GemNet-dT, and SpinConv, respectively.Table 11 compares each model with and without the proposed normalization scheme.Source: Alternative reference scheme · p. 34, Table 11 · p. 34
• Energy improvements were similar on OOD splits except for GemNet-OC, while force metrics changed little across models.The approach was omitted from the main paper to encourage alternative energy-normalization strategies for large, diverse datasets such as OC22 and OC20.Source: Alternative reference scheme · p. 34

Prediction routes

• Mixed-ML and Full-ML calculate adsorption energies from total-energy models, whereas the direct route predicts Ead itself from adsorbate, slab, and gas-phase DFT references.The direct target is Ead = Esys − Eslab − Egas; total-energy models use the two approaches defined in Main Equations 4 and 5.Source: Use of total energy models to predict… · p. 35

Motivation

• Changing the reference enables adsorption-energy predictions across surface coverages and supports additional properties such as surface energies.These benefits motivate total-energy prediction despite total energies lacking physical meaning without a proper reference.Source: Use of total energy models to predict… · p. 35

Reference consistency

• Allowing all slab atoms to relax can make clean-slab references inconsistent with relaxed adsorbate+slabs, especially because OC22 pairs were generated from unrelaxed slabs and relaxed in parallel.An accurate adsorption energy requires identical or similar relaxed slab references; OC22 showed substantially greater slab movement than OC20.Source: Use of total energy models to predict… · p. 35, Figure 10 · p. 35

Validation protocol

• Validation reran DFT for approximately 700 systems using serial slab-then-adsorbate relaxation, but ML predictions used DFT initial structures, making the comparison easier and conditioning interpretation of Main Table 8.Subsurface atoms were fixed, and some clean slabs came from training data to obtain a meaningful sample size.Source: Use of total energy models to predict… · pp. 35–36

Atomic drift with unconstrained relaxation

• 0.220 ˚A per-atom slab displacement in OC22 is comparable to OC20’s 0.231 ˚A despite cumulative drift of 14.83 ˚A versus 3.80 ˚A.All OC22 atoms were unconstrained and slabs were not optimized before adsorbate placement; the small per-atom drift indicates limited reconstruction for most slabs.Source: Atomic drift with unconstrained… · p. 37

OC22 adsorbate+slab and slab only performance

• Training on both OC22 subsplits achieves the best adsorbate+slab and slab performance, with each subset improving the other.Adsorbate+slab performance is better than slab-only performance, potentially because of dataset size and broader configurational sampling.Source: OC22 adsorbate+slab and slab only… · p. 38, Table 12 · p. 38

Training and hyperparameters

• S2EF-Total experiments used fixed model architectures, while hyperparameters varied across learning rate, optimizer, and scheduler settings.The reported configurations are made available in the Open Catalyst Project repository.Source: Training and hyperparameters · p. 38
• S2EF-Total OC22-only models used atomwise energy-and-force loss weighting and a stepwise learning-rate schedule, which improved force metrics.Joint training used the original OC20 loss and reduce-on-plateau scheduling, whereas fine-tuning used the OC20 loss because atomwise loss overfit forces quickly.Source: Training and hyperparameters · p. 38
• No energy normalization was used for OC22 training because it reduced performance across the board.Joint experiments involving OC20 used DFT total energies rather than adsorption energies.Source: Training and hyperparameters · p. 38
• Fine-tuning learning rates were reduced by 5–10× for stable training, and all models were optimized with AMSGrad.Models used 4–64 GPUs with 32GB NVIDIA Volta cards; top GemNet-OC joint and fine-tuning hyperparameters are provided in Table 13.Source: Training and hyperparameters · p. 38, Table 13 · p. 39

S2EF-Total, IS2RE-Total, IS2RS validation results

• Validation results cover default, joint-training, and fine-tuning strategies for S2EF-Total, IS2RE-Total, and IS2RS.Tables 14–16 report validation performance for total-energy prediction, relaxed-energy prediction, and relaxed-structure prediction, respectively.Source: S2EF-Total, IS2RE-Total, IS2RS… · p. 40, Table 14 · p. 40, Table 15 · p. 41

Additional DFT settings

• All structure relaxations used VASP with PAW, spin-polarized PBE-GGA, 500 eV plane-wave cutoffs, and convergence thresholds of 1 × 10−4 eV for energies and 0.05 eV Å−1 for forces.Γ-centered k-point meshes were used for bulk and slab calculations.Source: Additional DFT settings · p. 42
• Adsorbates were placed on one slab surface, so an electrostatic potential correction was added to compensate for the resulting nonphysical dipole moment.This correction prevents diverging total DFT energies caused by uneven surface charges.Source: Additional DFT settings · p. 42
• Adsorbate gas-phase references used linear combinations of N2, H2O, CO, and H2, explicitly avoiding O2 because GGA overestimates its O-O bond strength.The per-atom adsorbate energies are listed in Table 17.Source: Additional DFT settings · p. 42, Table 17 · p. 42

Hubbard U corrections

• Hubbard U corrections adapted from Materials Project values accounted for strong electron correlations that GGA cannot accurately represent in selected transition-metal oxides.Those U values were fitted to experimental formation enthalpies and are listed in Table 18.Source: Hubbard U corrections · p. 42, Table 18 · p. 43

Chemical Diversity of Open Catalyst 2022

• OC22 considered 51 elements, yielding 1,275 possible unary and binary oxides and 4,728 candidate bulk oxide structures from Materials Project queries.The element set included transition metals, alkali and alkaline-earth metals, selected lanthanides, and p-block metals and metalloids.Source: Chemical Diversity of Open Catalyst 2022 · p. 44, Figure 11 · p. 44
• Only 4,286 of 4,728 bulk oxides were sampled, with roughly 50 to 100 slabs and adsorbate+slabs per chemical system demonstrating broadly even chemical-space coverage.Some systems were unavailable in Materials Project or too large for slabs under 150 atoms.Source: Chemical Diversity of Open Catalyst 2022 · p. 44, Figure 12 · p. 45
• Li-O, Sb-Cr-O, and Ag-O systems were randomly oversampled, each exceeding 250 final slab or adsorbate+slab entries.The heat map distinguishes unary compositions, binary compositions, unavailable systems, and systems not randomly sampled.Source: Chemical Diversity of Open Catalyst 2022 · p. 44, Figure 12 · p. 45

Gibbs free energy calculations

• Gibbs free energy combines DFT energy, zero-point energy, heat-capacity, temperature, and entropy terms to determine ΔGad and OER overpotential.These quantities support accurate reaction pathways and microkinetic models for catalytic activity and selectivity.Source: Gibbs free energy calculations · p. 45
• Because OC22 datasets omit vibrational and entropic effects, the study did not train predictive models directly for ΔG.Instead, systematic corrections provide a lower-cost route from predicted adsorption energies to Gibbs adsorption energies.Source: Gibbs free energy calculations · pp. 45–46
• Corrections were calibrated from 15–20 relaxed adsorbate–slab systems per adsorbate using vibrational calculations at 298.15 K and 1 atm.The slab was confined while adsorbate sites relaxed, and Vasp Gibbs114 was used to obtain Gibbs free energies.Source: Gibbs free energy calculations · p. 46
• A slope-1 linear fit between Ead and ΔGad for O*, OH*, and OOH* supports obtaining Gibbs adsorption energies by adding an adsorbate-specific intercept correction.Figure 13 provides the fitted slope, while the intercepts define ΔGcorr for correcting predicted adsorption energies without additional intensive calculations.Source: Gibbs free energy calculations · p. 46, Figure 13 · p. 47
• All ΔGO*corr quantities agree with Gunasooriya and Nørskov 62 within 0.1 eV, whereas ΔGOOH*corr shows a substantial literature deviation.The OOH* discrepancy is associated with a 25.89 J K−1 mol−1 difference in H2O(g) entropy, which changes ΔGcorr by 0.06 eV.Source: Gibbs free energy calculations · p. 46

Trends and literature validation

• 83% linear correlation and 0.57 MAE were obtained for predicted O*, H*, and OH* adsorption energies versus literature values for perovskite surfaces.Predictions used the S2EF-total task with the GemNet-OC OC20+OC22 model and Full-ML scheme, excluding adsorbate–slab combinations present in training.Source: Trends and literature validation · p. 47, Figure 14 · p. 48
• Literature comparisons covered different functionals, Hubbard U corrections, magnetic configurations, chemical spaces, materials, and rutile as well as perovskite structures.Approximately 180 slabs shared the same clean slabs and adsorbate positions across all nine investigated adsorbates, with one-third of materials in each domain split.Source: Trends and literature validation · pp. 47–48
• Scaling slopes agreed with Gunasooriya and Nørskov within 0.15 eV, while the predicted intercept underestimated the literature value by 0.7 eV.The comparison used correction energies for ∆G_OH* versus ∆G_O* and related literature scaling data.Source: Trends and literature validation · p. 48, Figure 14 · p. 48
• R^2 exceeded 0.6 for monatomic-adsorbate scaling relationships except E_H* versus E_C*, which reached 0.55.Moderate scaling occurred for O*–OH*, H*–OH*, and OH*–OOH*, whereas other relationships were weak or absent.Source: Trends and literature validation · p. 48, Table 20 · p. 49

Changelog

• The changelog reports updated dataset counts, minor text and figure revisions, and publication of the document in ACS Catalysis.These changes were recorded across versions 2 and 3 after the initial release.Source: Changelog · p. 50
• Later revisions added conventional-DFT adsorption energies for a small validation subset and benchmarked total-energy-model predictions against that DFT set and literature data.The revisions also added correction energies for ∆G_O*, ∆G_OH*, and ∆G_OOH* and validated predicted scaling relationships against literature.Source: Changelog · p. 50
• The revisions expanded discussion of long-range, magnetic, charge, solvation, and experimental-outlook effects in oxide slabs.These additions accompanied the literature-validation and scaling analyses.Source: Changelog · p. 50