Discovery of Surface Reconstructions via Universal Machine Learning Interatomic Potentials

The atomistic structures of material surfaces play a critical role in heterogeneous catalysis, corrosion, and energy storage [1]. Atoms near the surface of bulk crystalline materials exhibit dangling bonds and a different coordination environment compared to their bulk counterparts. The surface orientation and the atomistic relaxation of the surface atoms determine the final structure of the stable surface and its chemical, mechanical, and electronic properties. In some cases, significant rearrangements of surface atoms, referred to as surface reconstructions, lead to even more stable surfaces than the relaxed, idealized Miller index surface [2]. Surface reconstruction plays an essential role in the identification of the most stable surface and is therefore affects various applications where surfaces and their interfaces play a critical role, such as batteries, fuel cells, and semiconductor devices [3]. It also impacts medical technologies by enhancing the corrosion and wear resistance properties of coated implants [4]. Density Functional Theory (DFT) is commonly utilized for the investigation of surface reconstructions. However, the vast space of possible surface atom arrangements that need to be considered for surface reconstruction discovery makes DFT too computationally expensive. The objective of this study is to overcome these limitations by utilizing universal machine learning interatomic potentials (UIP) in a Monte Carlo framework to facilitate the discovery of stable surface reconstructions. Recent advancements in UIPs have demonstrated exceptional performance in various computational materials science problems, including issues closely related to surface reconstruction such as catalysis, materials adsorption, and adhesion [5]. However, while UIPs trained on bulk structures, like MACE [6] and M3GNet [7], have shown to accurately predict bulk properties, they often struggle with surface properties such as the surface energy [8]. In our proposed framework, we employ a Monte Carlo algorithm to sample surface reconstructions efficiently [9] and evaluate the atomistically relaxed surface structure at each sampling step using the graph-based EquiformerV2 model that was trained on the OC20 dataset [10]. We show that the pre-trained EquiformerV2 potential performs better on surface energy predictions compared to MACE and M3GNet. Further, we benchmark our workflow’s ability to discover known surface reconstructions such as GaN (0001), Si(111) and SrTiO3 (001). By leveraging this advanced computational workflow, we aim to enhance the predictive accuracy and efficiency of surface reconstruction discovery, ultimately contributing to the development of materials with optimized surface properties.<br/><br/>References:<br/>[1] J. Chen <i>et al.</i>, <i>Electrochem. Energ. Rev.</i>, vol. 4, no. 3, pp. 566–600, Sep. 2021.<br/>[2] Y. Huang, W. Quan, H. Yao, R. Yang, Z. Hong, and Y. Lin, <i>Inorg. Chem. Front.</i>, vol. 10, no. 2, pp. 352–369, 2023.<br/>[3] J. Zhang, Z. Wu, and F. Polo-Garzon, <i>ACS Catal.</i>, vol. 13, no. 23, pp. 15393–15403, Dec. 2023.<br/>[4] S. Piscanec, <i>Acta Materialia</i>, vol. 52, no. 5, pp. 1237–1245, Mar. 2004.<br/>[5] J. Lan <i>et al.</i>, <i>npj Comput Mater</i>, vol. 9, no. 1, p. 172, Sep. 2023.<br/>[6] I. Batatia <i>et al.</i>, arXiv, Mar. 01, 2024. Accessed: Jun. 24, 2024.<br/>[7] C. Chen and S. P. Ong, <i>Nat Comput Sci</i>, vol. 2, no. 11, pp. 718–728, Nov. 2022.<br/>[8] B. Focassio, L. P. M. Freitas, and G. R. Schleder, arXiv, May 30, 2024. Accessed: Jun. 18, 2024.<br/>[9] X. Du <i>et al.</i>, <i>Nat Comput Sci</i>, vol. 3, no. 12, pp. 1034–1044, Dec. 2023.<br/>[10] L. Chanussot <i>et al.</i>, <i>ACS Catal.</i>, vol. 11, no. 10, pp. 6059–6072, May 2021.