PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Richard Tran, Janice Lan, Muhammed Shuaibi, Brandon M. Wood, Siddharth Goyal, Abhishek Das, Javier Heras-Domingo, Adeesh Kolluru, Ammar Rizvi, Nima Shoghi, Anuroop Sriram, Felix Therrien, Jehad Abed, Oleksandr Voznyy, Edward H. Sargent, Zachary Ulissi, C. Lawrence Zitnick

arXiv:2206.08917v3cond-mat.mtrl-scics.LGphysics.comp-ph

TL;DR

Oxide electrocatalysts lack sufficient training data for generalizable machine-learning models, despite their importance in oxygen evolution and related applications. The paper introduces OC22 with generalized total-energy tasks and evaluates transfer between OC22 and OC20, finding improved energy performance through joint training and fine-tuning.

  • Problem

    Oxide electrocatalysts lack sufficient training data, while their complex surfaces create challenges for catalyst design and generalizable machine-learning methods.

  • Method

    The paper presents the OC22 oxide-electrocatalyst dataset, accompanying tasks, graph-neural-network baselines, and generalized DFT total-energy targets spanning complex oxide configurations.

  • Results

    Joint training on OC20 and OC22 significantly improves energy MAE, while fine-tuning transfers knowledge between the chemically dissimilar datasets.

  • Takeaways & Limitations

    Generalized total-energy tasks support property prediction beyond adsorption energies and can improve catalyst-screening throughput when ML replaces adsorbate+slab energy calculations.

  • Takeaways & Limitations

    Bulk-oxide selection prioritized chemical diversity over stability, and many selected materials are not electrochemically stable prerequisites for viable electrocatalysts.

Abstract

from arXiv · show

The development of machine learning models for electrocatalysts requires a broad set of training data to enable their use across a wide variety of materials. One class of materials that currently lacks sufficient training data is oxides, which are critical for the development of OER catalysts. To address this, we developed the OC22 dataset, consisting of 62,331 DFT relaxations (~9,854,504 single point calculations) across a range of oxide materials, coverages, and adsorbates. We define generalized total energy tasks that enable property prediction beyond adsorption energies; we test baseline performance of several graph neural networks; and we provide pre-defined dataset splits to establish clear benchmarks for future efforts. In the most general task, GemNet-OC sees a ~36% improvement in energy predictions when combining the chemically dissimilar OC20 and OC22 datasets via fine-tuning. Similarly, we achieved a ~19% improvement in total energy predictions on OC20 and a ~9% improvement in force predictions in OC22 when using joint training. We demonstrate the practical utility of a top performing model by capturing literature adsorption energies and important OER scaling relationships. We expect OC22 to provide an important benchmark for models seeking to incorporate intricate long-range electrostatic and magnetic interactions in oxide surfaces. Dataset and baseline models are open sourced, and a public leaderboard is available to encourage continued community developments on the total energy tasks and data.

Introduction

Oxide electrocatalysts are important for energy applications but difficult to model because their surfaces and electronic structures are complex. OC22 addresses limited oxide training data with generalized tasks, benchmarks, and dataset combinations for machine-learning catalysis.

  • Motivation: Oxides are abundant electrocatalysts for water splitting and other energy applications, but oxide surfaces complicate catalyst design.Water splitting produces H2 and O2, while computational screening can help discover multi-component oxide OER catalysts.
  • Challenges: Oxide modeling must account for polymorphs, reconstruction, leaching, doping, defects, surface terminations, competing mechanisms, and strong electron correlation.Large systems and long-range electrostatic or magnetic interactions can also slow convergence.
  • Data gap: OC20 excluded metal oxides because of these complexities, leaving oxide materials underrepresented in catalyst-learning datasets.OC20 included approximately 250 million single-point calculations but did not include metal oxide materials.
  • Contribution: OC22 presents an oxide-focused dataset for OER and broader oxide electrocatalysis, with accompanying tasks and graph-neural-network baseline models.The dataset is intended to complement OC20 and support more generalizable machine-learning models for catalysis.
  • Contribution: OC22 expands beyond adsorption-energy prediction by including DFT total energy targets for complex oxide surfaces and configurations.The dataset covers varying terminations, adsorbate-plus-slab configurations, coverage, substitutions, and vacancies.
  • Evaluation: The study evaluates total-energy, force, relaxed-energy, and relaxed-structure prediction tasks using composition-based train, validation, and test splits.The splits target prediction for unseen crystal compositions, and the work explores transfer learning and joint training between OC20 and OC22.

The OC22 Dataset

OC22 was constructed as a chemically diverse oxide dataset spanning slabs, adsorbate-plus-slab systems, surface defects, coverages, and oxide-relevant adsorbates. Its curation prioritizes broad chemical coverage, while explicitly acknowledging stability and representation limitations.

  • Dataset scope: 9,854,504 single-point calculations were generated from 19,142 slabs and 43,189 adsorbate-plus-slab systems.The calculations provide forces and energies partitioned into training, validation, and test splits.
  • Bulk selection: 4,728 unary and binary metal oxides from the Materials Project formed the bulk-oxide starting set, supplemented by 173 unary and binary rutile structures.The selection considered multiple low-energy-above-hull structures for each chemical system.
  • Bulk selection: The bulk selection prioritized chemical diversity over stability, so many selected materials are not electrochemically stable.Pourbaix analysis indicated relative aqueous stability for oxides containing only 26 of the 51 considered elements.
  • Bulk selection: The dataset uses Materials Project entries consistently, despite uneven representation across chemical systems and coverage gaps in other oxide databases.Some systems have over 300 reported entries while 200 chemical systems have none in Materials Project.
  • Surface construction: Slabs were sampled with Miller indices up to 3, replicated to at least 8 Å depth and width, and decorated symmetrically with randomly sampled oxygen vacancies.Symmetric vacancy removal avoids non-physical dipole moments that can cause diverging DFT energies.
  • Adsorbate construction: Nine adsorbates were sampled across OER intermediates and additional oxide chemistry, with binding allowed at surface oxygen, under-coordinated metals, or oxygen vacancies.Adsorbates could occur multiple times per surface, subject to site availability and minimum-separation constraints.
  • Adsorbate construction: Placement strategies include oxygen-dimer and surface-reaction configurations, enabling exploration of intermediate reactions specific to oxides.Examples include formation of CO2* and OOH* through binding to surface oxygen.
  • Computational settings: OC22 uses spin-polarized PBE-GGA calculations and differs from OC20 in computational settings, including exchange-correlation treatment.The dataset also accounts for strong electron correlations in some transition-metal oxides with a Hubbard U correction.

Tasks

OC22 reframes catalyst-learning tasks around DFT total energies because unconstrained oxide surfaces make OC20-style adsorption-energy references ill-posed. These tasks support prediction for isolated surfaces and adsorbate-plus-slab systems.

  • Motivation: Total-energy models can cover surface selection, termination stability, coverage, and adsorption stages in the OER workflow.The workflow first selects a bulk structure, enumerates terminations, identifies a stable termination, places an adsorbate, and relaxes the combined system.
  • Task design: OC22 includes isolated surfaces and surface-plus-adsorbate structures with DFT-computed ground-truth targets.Each surface uses a periodic unit cell with at least 12 Å of vacuum.
  • Tasks: S2EF, IS2RE, and IS2RS respectively predict energy and forces, relaxed energy, and relaxed structure from an input structure.OC22 retains these established task formats while introducing total-energy variants.
  • Task design: OC22 modifies OC20’s S2EF and IS2RE tasks because unconstrained atoms invalidate the comparable relaxed-slab reference assumed for adsorption energies.Using the OC20 reference would produce an incorrect, ill-posed, noisy target.
  • Tasks: S2EF-Total predicts DFT total energy and per-atom forces from a structure, whereas S2EF predicts referenced adsorption energy and forces.The tasks differ only in the energy target.
  • Tasks: IS2RE-Total predicts relaxed DFT total energy from an initial structure, extending IS2RE’s relaxed-energy prediction beyond adsorption-energy references.Total-energy predictions can later be referenced flexibly to compute physically relevant properties.

Baseline GNN Models

OC22 benchmarks diverse graph neural networks, including equivariant and direct-force architectures, on total-energy and relaxation tasks. The evaluation uses composition-based splits and balanced energy-force losses.

  • Architecture: GNNs represent atoms as graph nodes and atomic interactions as edges, learning representations through message passing.This distinguishes them from descriptor-based models that use hand-crafted atomic-environment representations.
  • Model coverage: OC22 evaluates GNNs that performed well on OC20 or molecular datasets, with GemNet-OC among the benchmarked models.S2EF-Total includes SchNet, DimeNet++, ForceNet, SpinConv, PaiNN, GemNet-dT, and GemNet-OC.
  • Evaluation: S2EF-Total benchmarks seven model families, while IS2RS and IS2RE-Total use smaller sets focused on selected architectures.Top-performing S2EF-Total models were also evaluated for IS2RE-Total through iterative relaxation.
  • Architecture: The benchmark includes direct-force models, equivariant models, angular-interaction models, and models with symmetric or higher-order message passing.PaiNN uses spherical harmonics through l = 1, while GemNet-OC captures quadruplet interactions.
  • Evaluation: The dataset splits distinguish in-distribution compositions from out-of-distribution unseen catalyst compositions across slab and adsorbate-plus-slab systems.Validation and test splits are similar in size and use exclusive compositions.
  • Training objective: Models balance energy and force prediction losses using empirical weights, with p = 1 for most S2EF-Total models and p = 2 for GemNet-dT and GemNet-OC.IS2RE-Total evaluates only the energy term by setting λF = 0.

Evaluation Metrics

OC22 evaluates generalized total-energy and force prediction, structure relaxation, transfer-learning strategies, and adsorption-energy utility across oxide electrocatalyst tasks. GemNet-OC generally performs strongly, while out-of-domain total-energy prediction and practical relaxation accuracy remain challenging.

  • Training strategies: OC22-only, joint training, and fine-tuning compare alternative uses of OC20 and OC22, with fine-tuning starting from pretrained OC20 models.For S2EF-Total, joint-training experiments also vary the amount of OC20 data.
  • S2EF-Total: GemNet-OC performs best across S2EF-Total energy and force metrics, plausibly benefiting from its 12 Å receptive field and explicit quadruplet interactions.The authors suggest these design choices better capture long-range interactions.
  • Generalization: OOD total-energy metrics are substantially worse than ID metrics, indicating that generalization to unseen material-species combinations is more difficult than referenced adsorption-energy prediction.The paper frames adsorption energy as a correction-like target that can simplify the learning problem.
  • Training strategies: Fine-tuned GemNet-OC consistently outperforms its OC22-only variant across ID data fractions, while performance continues improving with additional data on OOD cases.The ID gains show diminishing returns near 50% of the data, whereas OOD performance continues to improve with data size.
  • IS2RE-Total: Fine-tuning improves relaxation-based IS2RE-Total performance, but the best relaxation and direct GemNet-OC approaches achieve EwT values of 3.89% and 3.66%, respectively.Joint training consistently hurts direct-approach performance, while relaxation gives considerably better OOD metrics.
  • IS2RS: GemNet-OC outperforms other models across IS2RS metrics, although force-based and practical-accuracy results still indicate substantial room for improvement.Joint training and fine-tuning improve DFT force-based metrics over OC22-only, with fine-tuned GemNet-OC-Large achieving the best force metrics.
  • Cross-dataset training: Joint OC20+OC22 training improves OC20 energy MAE by approximately 20%, while force metrics are unchanged or slightly worse.The improvement occurs despite OC22 being much smaller and oxide-specific relative to OC20.
  • Adsorption-energy utility: Total-energy models produce higher adsorption-energy error than direct adsorption-energy models, reflecting the greater challenge and broader scope of total-energy prediction.Mixed-ML predicts adsorbate+slab energy, whereas Full-ML predicts both slab and adsorbate+slab energies.

Discussion

OC22 extends catalyst modeling toward generalized potentials for oxide systems, while exposing challenges from long-range, magnetic, charge, solvation, and multi-level-theory effects. Baseline models show useful but insufficient performance for experimentally relevant OER properties.

  • Limitations: OC22 highlights challenges in capturing long-range interactions, modeling properties beyond adsorption energy, and fitting multiple datasets and theory levels.These challenges arise alongside variation across adsorbates and materials and difficulties selecting calculations that improve generalizability.
  • Results: 0.24 eV and 0.94 eV energy MAEs were achieved for ID and OOD S2EF-Total predictions using transfer learning from OC20.The same model produced 1.12 eV and 1.85 eV MAEs for ID and OOD relaxed total-energy predictions.
  • General-purpose potentials: Total-energy tasks target general-purpose potentials that can predict surface energy and other properties beyond adsorption energy.This broader scope is especially relevant when assessing the stability of competing oxide surface terminations.
  • Long-range interactions: Oxide semiconductor partial charges create long-range electrostatic interactions that existing GNNs may struggle to capture.Many OC22 oxides lack the electronic screening characteristic of metal surfaces.
  • Future directions: Magnetic configurations, charge balancing, oxygen vacancies, and solvation remain important directions for improving oxide-surface modeling.The paper notes that magnetic polymorphs can alter energetic, structural, magnetic, and chemisorption trajectories, while solvation was not directly modeled.
  • Training and applications: Joint OC20–OC22 training produces unexpected energy improvements and suggests broader dataset combinations may benefit model scaling.The authors also identify adsorption, surface, and reaction energies as quantities accessible from predicted slab and adsorbate-plus-slab total energies.
  • Applications and outlook: OC22 is intended to support oxide catalyst discovery across reaction families, including systems containing oxygen defects and vacancies.The authors also propose complementing idealized DFT models with experimental validation and auxiliary modeling data.

Supporting Information Available

The supporting information provides additional analyses, methodological details, and dataset resources, while the full dataset and baseline implementation are publicly available.

  • Supporting information: The supporting information includes additional total-energy results, reference-scheme comparisons, adsorption-energy analyses, validation results, corrections, and Hubbard U details.It also documents adsorbate-plus-slab and slab performance, training, and hyperparameters.
  • Open resources: The full OC22 dataset is available in ASE trajectory and model-ready LMDB formats, with baseline code, dataloaders, and trainers in the open-source repository.The paper provides both the dataset website and repository links.

OC20 S2EF-Total and IS2RE-Total results

OC20 total-energy variants are harder than their referenced counterparts, with especially degraded performance for out-of-distribution cases. This reflects the greater difficulty of predicting total system energies rather than adsorption-reference energies.

  • S2EF-Total: S2EF-Total performs considerably worse than S2EF across all models and splits on OC20.Results are reported on the OC20 validation splits.
  • IS2RE-Total: IS2RE-Total similarly performs worse than IS2RE across all models and splits.The comparison uses OC20 performance on initial-structure-to-relaxed-energy tasks.
  • ID versus OOD: OOD metrics are significantly worse than ID metrics for OC20 total-energy tasks.The paper attributes the overall difficulty to predicting total energies rather than referenced counterparts.

Alternative reference scheme

The paper introduces linear referencing as an energy-normalization scheme for OC22 total-energy targets. It improves ID energy metrics across baseline models, while forces change little.

  • Linear referencing: Linear referencing fits per-element energies and uses them to normalize OC22 DFT total-energy targets before training.The fitted coefficients are combined with each system’s element counts.
  • Results: 49.9%, 24.5%, and 4.5% ID energy improvements were observed for GemNet-dT, SpinConv, and GemNet-OC, respectively.These improvements compare S2EF-Total performance with and without linear referencing.
  • Results: OOD energy metrics also improve for all models except GemNet-OC under the linear reference.Force metrics show little change across models.
  • Interpretation: The authors omitted linear referencing from the main paper to encourage exploration of alternative normalization strategies for large, diverse datasets.The scheme is presented as one approach rather than a definitive solution.

Use of total energy models to predict adsorption energies

OC22 uses total-energy models to support adsorption-energy prediction with references designed for broader oxide-catalysis properties, including varying surface coverages. Because unconstrained slab relaxation can produce inconsistent references, the authors validate these approaches against additional DFT calculations.

  • Reference choice: Total-energy predictions require a proper reference, but can support properties beyond adsorption energies.The paper contrasts total-energy prediction with direct adsorption-energy prediction and identifies adsorption energy as one of several accessible properties.
  • Prediction approaches: Direct-ML predicts adsorption energy from an adsorption-energy target, Ead = Esys − Eslab − Egas, using adsorbate+slab, clean-slab, and gas references.
  • Prediction approaches: Mixed-ML and Full-ML use total-energy predictions to calculate adsorption energies through two alternative reference schemes.
  • Reference choice: OC22 changes the referencing scheme to span different surface coverages and enable properties such as surface energies.
  • Validation: The authors validate Mixed-ML and Full-ML predictions on approximately 700 validation systems with newly rerun DFT adsorption-energy references.Some clean slabs came from the training set, and the validation calculations fixed subsurface atoms while using serial relaxations.

Atomic drift with unconstrained relaxation

Unconstrained OC22 slab relaxations produce much larger cumulative atomic drift than OC20, although average per-atom displacement remains comparable. Model performance also differs between adsorbate+slab and slab subsets, while the benchmark tables report several total-energy and relaxation tasks.

  • Atomic drift: 14.83 Å versus 3.80 Å: OC22 and OC20 show these average cumulative atomic drifts, respectively.The larger OC22 drift follows from unconstrained atoms and slabs not being optimized before adsorbate placement.
  • Atomic drift: 0.220 Å versus 0.231 Å: OC22 and OC20 have comparable average per-atom slab displacements despite different cumulative drift.Most slab atoms are bulk-like, so their individual movements contribute substantially to cumulative drift.
  • Atomic drift: Full relaxation generally does not cause significant reconstruction for most slabs, based on the small OC22 per-atom displacement.The authors connect this displacement to preservation of bulk-like central-slab properties within 0.220 Å.
  • Subset performance: Adsorbate+slab models perform considerably better than slab-only models in isolated subset evaluations.The authors suggest dataset size and the broader configurational space sampled by adsorbate+slab relaxations as possible reasons.
  • Benchmark tasks: Tables 14–16 evaluate S2EF-Total, IS2RE-Total, and IS2RS under default, joint-training, and fine-tuning strategies.The reported validation settings include model hyperparameters and baseline comparisons.

Additional DFT settings

OC22 calculations use spin-polarized VASP with PAW and PBE-GGA, together with electrostatic and Hubbard-U treatments relevant to oxide systems. The dataset covers unary and binary oxides assembled from selected elements and Materials Project structures.

  • Electronic-structure settings: All relaxations use VASP with the projector augmented wave approach and PBE-GGA exchange-correlation.
  • Electronic-structure settings: Spin polarization accounts for significant spin states in metal oxides, with plane-wave kinetic-energy cutoffs of 500 eV.
  • Electrostatic treatment: A dipole correction is applied because adsorbates on one slab surface create uneven surface charges and nonphysical dipole moments.
  • Reference energies: Adsorbate gas references use linear combinations of N2, H2O, CO, and H2, explicitly avoiding O2 because GGA overestimates its bond strength.
  • Electronic-structure settings: Hubbard U corrections account for strong electron correlations that GGA cannot properly represent in certain transition-metal oxides.
  • Dataset construction: The dataset samples unary and binary oxides from 51 elements and queries Materials Project structures with the five lowest energies above hull and fewer than 150 atoms.The element set includes transition, alkali, alkaline-earth, lanthanide, p-block metals and metalloids, with Tc excluded for radioactivity.

Gibbs free energy calculations

The paper estimates Gibbs adsorption energies by combining DFT adsorption energies with vibrational, zero-point, heat-capacity, and entropy contributions. For O*, OH*, and OOH*, the resulting Gibbs corrections show near-unity linear relationships with adsorption energies, while vibrational and entropic effects are absent from the OC22 predictive datasets.

  • Purpose and formulation: Gibbs adsorption energies support reaction-pathway construction, microkinetic modeling, and OER overpotential assessment.
  • Purpose and formulation: The Gibbs free-energy expression includes DFT system energy, zero-point-energy change, heat-capacity integration, temperature, and entropy change.
  • Scope: OC22 datasets omit vibrational and entropic effects and therefore do not include predictive models for Gibbs free energy.
  • Correction procedure: At specified atmospheric conditions, Gibbs adsorption energy can be estimated by adding an adsorbate-specific correction to DFT adsorption energy.
  • Correction procedure: 15–20 relaxed adsorbate+slabs per adsorbate are sampled for vibrational calculations at 298.15 K and 1 atm to determine the correction.
  • Results: A near-unity slope and non-zero intercept relate Gibbs and DFT adsorption energies for O*, OH*, and OOH*; the intercept provides the correction.
  • Results: The OOH* correction underestimates the literature value, with a 25.89 J K−1 mol−1 entropy difference for gaseous water changing the correction by 0.06 eV.

Trends and literature validation

The study evaluates total-energy-model predictions against literature adsorption energies and adsorbate scaling relationships, using out-of-domain oxide surfaces and comparisons with prior relationships. The reported predictions broadly track literature values, while scaling strength varies across adsorbate pairs.

  • Literature validation: The model was benchmarked against literature adsorption energies across differing functionals, Hubbard U corrections, magnetic configurations, chemical spaces, and materials.Training-set overlap was avoided for evaluated adsorbate–slab combinations.
  • Scaling relationships: The predicted ΔGOH*–ΔGO* scaling relationship had a slope consistent with the Gunasooriya–Nørskov relationship within 0.15 eV.The comparison used correction energies from Table 19 and included the literature relationship and datapoints.
  • Literature validation: Predicted O*, H*, and OH* adsorption energies for perovskite surfaces were compared with literature values using parity and mean-absolute-error reference lines.The literature values were computed with the rPBE functional.
  • Scaling relationships: R2 exceeded 0.6 for monatomic-adsorbate scaling relationships except EH* versus EC*, which had R2 = 0.55.Moderate linear scaling was also observed for EO* versus EOH*, EH* versus EOH*, and EOH* versus EOOH*; other relationships were weak or absent.
  • Scaling relationships: The best-performing scaling-relationship results were obtained with the GemNet-OC OC20+OC22 total-energy model.Table 20 presents the slopes, intercepts, and R2 values for the OC22 adsorbates.
Loading 2206.08917v3…