Identifying Crystal Structure of Boron Nitride from DFT Obtained XAS Spectra Using Machine Learning

The recent advancement of IoT devices has driven the rapid growth of high-capacity, high-speed, and wide-area information communications. 2D materials have attracted enormous attention due to their potentials for use in low-cost, stable communication devices^[1]. Among these materials, hexagonal boron nitride (h-BN) and crystalline boron nitride (BN) have attracted huge attention on its semiconducting nature. However, it is challenging to understand how structural changes influence physical properties. In particular, point defects and stacking have a significant influence on the electronic structure^[2]. Furthermore, interpreting spectral data requires expert’s domain knowledge, and manually analyzing large datasets is labor-intensive.<br/>In this study, we demonstrate a structure-to-property linkage using machine learning approach. We performed first-principles calculations of electronic structures and X-ray absorption spectroscopy (XAS) spectra of hexagonal (h-BN), cubic (c-BN), and wurtzite (w-BN), which are different phases of BN crystal phases. We applied interpretable machine learning to connect the information between lattice structure, electronic structure and XAS spectra.<br/>All computations were performed using PBE functional and the PAW exchange-correlation potential within the GGA by VASP. XAS spectra (B K-edge) were obtained by SCH^[3] (super-cell core-hole) method implemented in VASP. We considered different phases of BN, such as h-BN (P6₃/mmc) both in monolayer and bulk form, c-BN (F-43m), and w-BN (P6₃/mc) with their single B and N vacancy analogues for our study. All structures were energy minimized, before obtaining their corresponding XAS spectra. Principal Component Analysis (PCA), Multidimensional Scaling (MDS), and Uniform Manifold Approximation and Projection (UMAP) machine learning models were applied to embed the XAS spectral data into lower dimensions.<br/>The structural properties and XAS spectra of different phases of BN obtained using DFT are in reasonable agreement with experimental data^[4-6]. By applying various dimension reduction methods, we confirmed that UMAP effectively arranged the spectral data points according to their crystal structure. This is because UMAP can capture the overall shape of the spectrum and slight peak shifts, not just the points with large dispersion. It is also confirmed that UMAP could identify differences in the atomic point defects and the number of layers. In the reverse analysis of the UMAP results, we were able to identify the π* and σ* peaks for h-BN, and σ* peaks for c-BN and w-BN structures. Hence, the dimension reduction approach using UMAP can distinguish the XAS spectra of BN by nature of crystal structures.<br/><br/>Reference<br/>[1] Y.Lei et al. ACS Nanosci. 2, 6, 450–485 (2022)<br/>[2] H. Li et al. JJAP Conf. Proc. 9, 011104 (2023)<br/>[3] Karsai, F. et al. Phys. Rev. B 98, 235205 (2018)<br/>[4] J. Wang et al. Nanoscale, 2015, 7, 1718 (2015)<br/>[5] McDougall, N.L. et al. MRS Advances 2, 1545–1550 (2017)<br/>[6] Meng, Y. et al. Nature Materials 3, 111–114 (2004)<br/><br/>Acknowledgement<br/>RH and AV thank Institute of Molecular Science, Okazaki, Japan for supercomputing facilities received for all calculations (Project: 23-IMS-C137), and all authors thank CREST project for generous funding (JPMJCR21O4).