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

Richard Tran, Janice Lan, Muhammed Shuaibi, Siddharth Goyal, Brandon M. Wood, Abhishek Das, Javier Heras-Domingo, Adeesh Kolluru, Ammar Rizvi, Nima Shoghi, Anuroop Sriram, Zachary W. Ulissi, C. Lawrence Zitnick

TL;DR

This paper introduces the Open Catalyst 2022 (OC22) dataset, a large-scale collection of DFT calculations for oxide electrocatalysts, addressing a critical data gap for these important materials. The dataset enables new "total energy" prediction tasks, moving beyond adsorption energies to better model complex oxide systems and accelerate the discovery of catalysts for applications like water splitting.

What Problem Does This Paper Address?

Developing efficient and cost-effective catalysts for renewable energy technologies, such as water splitting for hydrogen production, is crucial for a sustainable future. Inorganic oxides are promising electrocatalysts, but their complex nature—including diverse crystal structures, surface reconstructions, and magnetic properties—makes them challenging to model. Existing machine learning datasets have largely focused on simpler metal surfaces and adsorption energies, leaving a significant data gap for oxides. This lack of data hinders the development of generalizable machine learning models capable of accurately predicting the behavior of oxide electrocatalysts, slowing down materials discovery.

Methods

The authors introduce the OC22 dataset, comprising 62,331 DFT relaxations (approximately 9.8 million single-point calculations) for a diverse range of oxide materials, coverages, and adsorbates. A key innovation is the definition of "total energy" tasks (S2EF-Total and IS2RE-Total) that predict the absolute DFT energy and forces, rather than just adsorption energies. This allows for a more comprehensive understanding of material properties. They also evaluate several state-of-the-art graph neural network (GNN) models, including GemNet-OC, on these new tasks. The paper explores various training strategies: OC22-only training, joint training with the existing OC20 dataset, and fine-tuning pretrained models.

Datasets

  • Open Catalyst 2022 (OC22) Dataset: A new dataset containing 62,331 DFT relaxations (~9.8 million single-point calculations) of oxide materials, adsorbates, and coverages. It includes unary and binary oxides, with a focus on chemical diversity and various surface terminations and vacancy defects.
  • Open Catalyst 2020 (OC20) Dataset: Used in conjunction with OC22 for joint training and fine-tuning experiments. OC20 primarily focused on metal surfaces and adsorption energies.

Key Insights and Ideas

  • Total Energy is Key for Oxides: Moving beyond adsorption energies to predict total energies is crucial for accurately modeling complex oxide systems, which exhibit phenomena like long-range electrostatic and magnetic interactions that are not fully captured by adsorption-centric tasks.
  • Data Diversity is Essential: The OC22 dataset's emphasis on chemical diversity, surface terminations, and vacancy defects is vital for building generalizable models that can perform well on unseen oxide compositions.
  • Joint Training Boosts Performance: Combining the OC20 and OC22 datasets significantly improves model performance, demonstrating the value of leveraging diverse data sources, even if they differ in DFT theory or material class.
  • Fine-tuning Offers a Compromise: Fine-tuning pretrained models (e.g., from OC20) on OC22 can achieve substantial performance gains, offering a more computationally efficient path to high accuracy compared to training on the full combined dataset.

Key Results

  • GemNet-OC Achieves State-of-the-Art: GemNet-OC, a graph neural network model, demonstrates top performance across various tasks.
  • Significant Improvements with Joint Training: Joint training OC20 and OC22 led to a ~19% improvement in total energy predictions and a ~9% improvement in force predictions.
  • Fine-tuning Benefits: Fine-tuning OC20 pretrained models on OC22 resulted in a ~36% improvement in energy predictions for the general S2EF-Total task.
  • Literature Validation: The models successfully captured literature adsorption energies and important OER scaling relationships, demonstrating practical utility.
  • OOD Performance is Challenging: Out-of-domain (OOD) predictions for OC22 were substantially higher than in-domain (ID) predictions, highlighting the difficulty of generalizing to unseen material compositions and interactions.

Limitations and Future Work

  • Computational Cost: Generating large, high-fidelity DFT datasets for complex materials like oxides remains computationally expensive.
  • Long-Range Interactions: Capturing intricate long-range electrostatic and magnetic interactions in oxide surfaces is a significant challenge for current GNN models. Future work should explore models that can better handle these interactions.
  • Higher-Level Theory: The dataset highlights the need for models that can handle varying levels of theory and materials, potentially requiring more sophisticated approaches than standard GGA functionals.
  • Magnetic and Charge Effects: Incorporating magnetic configurations and charge balancing effects is crucial for accurate oxide modeling, and future work should explore how to integrate this physics into ML models.
  • Solvation Effects: The current dataset does not explicitly model solvation effects, which can significantly impact catalytic properties. Future research could incorporate these effects.
  • Experimental Validation: While models show promise, direct experimental validation is essential to confirm their real-world applicability.

Summary updated

Abstract

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.
BibTeX
@inproceedings{tran2023open,
  author = {Richard Tran and Janice Lan and Muhammed Shuaibi and Siddharth Goyal and Brandon M. Wood and Abhishek Das and Javier Heras-Domingo and Adeesh Kolluru and Ammar Rizvi and Nima Shoghi and Anuroop Sriram and Zachary W. Ulissi and C. Lawrence Zitnick},
  title = {{The Open Catalyst 2022 (OC22) Dataset and Challenges for Oxide Electrocatalysis}},
  booktitle = {ACS},
  year = {2023},
  doi = {10.1021/acscatal.2c05426},
}
Figure