Adapting OC20-trained EquiformerV2 Models for High-Entropy Materials (2403.09811v1)

Abstract: Computational high-throughput studies, especially in research on high-entropy materials and catalysts, are hampered by high-dimensional composition spaces and myriad structural microstates. They present bottlenecks to the conventional use of density functional theory calculations, and consequently, the use of machine-learned potentials is becoming increasingly prevalent in atomic structure simulations. In this communication, we show the results of adjusting and fine-tuning the pretrained EquiformerV2 model from the Open Catalyst Project to infer adsorption energies of *OH and *O on the out-of-domain high-entropy alloy Ag-Ir-Pd-Pt-Ru. By applying an energy filter based on the local environment of the binding site the zero-shot inference is markedly improved and through few-shot fine-tuning the model yields state-of-the-art accuracy. It is also found that EquiformerV2, assuming the role of general machine learning potential, is able to inform a smaller, more focused direct inference model. This knowledge distillation setup boosts performance on complex binding sites. Collectively, this shows that foundational knowledge learned from ordered intermetallic structures, can be extrapolated to the highly disordered structures of solid-solutions. With the vastly accelerated computational throughput of these models, hitherto infeasible research in the high-entropy material space is now readily accessible.

• The paper introduces an 'Atoms of Interest' energy correction method that improves adsorption energy predictions, achieving a MAE of 0.030 eV for *OH.
• The research demonstrates the effective use of both zero-shot and few-shot learning by fine-tuning large and small EquiformerV2 models, reducing computational costs.
• The findings enable high-throughput screening of catalytic properties in high-entropy alloys, offering a scalable alternative to computationally intensive DFT methods.

Adapting EquiformerV2 Models for High-Entropy Materials Analysis

Introduction

The exploration of high-entropy alloys (HEAs) and materials (HEMs) presents significant challenges due to their complex microstates and vast compositional spaces. Despite the proven utility of density functional theory (DFT) in investigating catalytic materials, its application to HEAs and HEMs is computationally prohibitive. The adaptation of machine-learned potentials, particularly the EquiformerV2 model trained on the Open Catalyst Project's dataset (OC20), offers a promising avenue for accelerating computational research in this domain. This paper outlines improvements to the EquiformerV2 model's ability to predict adsorption energies on high-entropy materials, highlighting its capacity for zero-shot and few-shot learning, and demonstrating its potential utility in high-throughput studies of HEAs and HEMs.

Methods

The research employs the EquiformerV2 model, assessing its performance in predicting adsorption energies of *OH and *O on an Ag-Ir-Pd-Pt-Ru high-entropy alloy, a relevant system for the oxygen reduction reaction (ORR). The model's efficacy is tested in two configurations: a larger model with 153 million parameters and a smaller one with 31 million parameters. The models were fine-tuned and assessed using a dataset of DFT-simulated slabs, refined with ASE and GPAW tools. A notable aspect of this paper is the introduction of an "Atoms of Interest" energy correction method to refine the model’s predictive accuracy by focusing on energetically significant atomic interactions.

Results and Discussion

The application of the "Atoms of Interest" energy correction method resulted in a marked improvement in the model's prediction accuracy, particularly for *OH adsorption energies, achieving a mean absolute error (MAE) of 0.030 eV. Moreover, the paper established the feasibility of fine-tuning the EquiformerV2 model with minimal training data, significantly enhancing its predictive capability for both *OH and *O adsorption energies. Additionally, the fine-tuned models facilitated the development of an informed direct inference model through knowledge distillation, further refining the accuracy of complex binding site predictions. These adaptations underscore the model's versatility and its potential as a foundational tool in exploring the catalytic properties of HEAs and HEMs.

Conclusion

Through methodical adjustments and fine-tuning, the EquiformerV2 model demonstrated substantial accuracy improvements in predicting adsorption energies on high-entropy materials. This paper not only advances the utility of machine-learned models in material science but also opens new pathways for the computational exploration of HEAs and HEMs. It posits the EquiformerV2 model as a potent tool for rapidly screening potential catalyst materials, ultimately facilitating research that was previously hindered by computational limitations.

Future Prospects

The research underscores the potential for continuous improvement in the application of machine-learned models to catalysis and material science. Future work may explore the extent to which these models can be generalized across a wider range of materials, further reducing reliance on computationally intensive methods like DFT. Additionally, the efficacious implementation of knowledge distillation techniques invites further investigation into streamlined model training processes, potentially accelerating the discovery of novel catalytic materials.

Acknowledgements and Data Availability

The paper benefited from insightful feedback and support from a broad spectrum of researchers within the field. The models and methods discussed are accessible via the Open Catalyst Project's GitHub repository, ensuring transparency and facilitating further research in this dynamic area of paper.