Machine-learning models for Raman spectra analysis of twisted bilayer graphene

TL;DR

This paper develops a machine learning approach to predict and interpret Raman spectra of twisted bilayer graphene, enabling rapid analysis of structural information from spectral data with high accuracy and robustness.

Contribution

It introduces a continuous machine learning model linking twist angle to Raman spectra, improving analysis speed and reducing human bias.

Findings

MLRs explain 98% of data variance

Spectral features near the G-band are most informative

Models are robust to noise and applicable to experimental data

Abstract

The vibrational properties of twisted bilayer graphene (tBLG) show complex features, due to the intricate energy landscape of its low-symmetry configurations. A machine learning-based approach is developed to provide a continuous model between the twist angle and the simulated Raman spectra of tBLGs. Extracting the structural information of the twist angle from Raman spectra corresponds to solving a complicated inverse problem. Once trained, the machine learning regressors (MLRs) quickly provide predictions without human bias and with an average of 98% of the data variance being explained by the model. The significant spectral features learned by MLRs are analyzed revealing the intensity profile near the calculated G-band to be the most important feature. The trained models are tested on noise-containing test data demonstrating their robustness. The transferability of the present models to experimental Raman spectra is discussed in the context of validation of the level of theory used for the construction of the analyzed database. This work serves as a proof of concept that machine-learning analysis is a potentially powerful tool for interpretation of Raman spectra of tBLG and other 2D materials.