Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Machine Learning-Driven Insights into Excitonic Effects in 2D Materials

Ahsan Javed, Sajid Ali

TL;DR

The paper tackles the computational bottleneck of determining exciton binding energies in 2D materials by developing a machine learning framework trained on C2DB descriptors to predict EBE and G0W0 band gaps from inexpensive features. Random Forest models deliver strong predictive performance (G0W0 gap R^2 ~ 0.98; EBE R^2 ~ 0.84), with SHAP analysis highlighting the PBE band gap and layer thickness as key drivers. A Bayesian optimization loop using Gaussian process regression and Expected Improvement accelerates the search, efficiently identifying high-EBE monolayers from ~4001 candidates and listing notable candidates in both ACl2-type and TMDC families. The approach offers a rapid, scalable screening pathway for 2D optoelectronic materials and could be extended to three-dimensional systems, enabling broader data-driven materials discovery.

Abstract

Understanding excitonic effects in two-dimensional (2D) materials is critical for advancing their potential in next-generation electronic and photonic devices. In this study, we introduce a machine learning (ML)-based framework to predict exciton binding energies in 2D materials, offering a computationally efficient alternative to traditional methods such as many-body perturbation theory (GW) and the Bethe-Salpeter equation. Leveraging data from the Computational 2D Materials Database (C2DB), our ML models establish connections between cheaply available material descriptors and complex excitonic properties, significantly accelerating the screening process for materials with pronounced excitonic effects. Additionally, Bayesian optimization with Gaussian process regression was employed to efficiently filter materials with largest exciton binding energies, further enhancing the discovery process. Although developed for 2D systems, this approach is versatile and can be extended to three-dimensional materials, broadening its applicability in materials discovery.

Machine Learning-Driven Insights into Excitonic Effects in 2D Materials

TL;DR

The paper tackles the computational bottleneck of determining exciton binding energies in 2D materials by developing a machine learning framework trained on C2DB descriptors to predict EBE and G0W0 band gaps from inexpensive features. Random Forest models deliver strong predictive performance (G0W0 gap R^2 ~ 0.98; EBE R^2 ~ 0.84), with SHAP analysis highlighting the PBE band gap and layer thickness as key drivers. A Bayesian optimization loop using Gaussian process regression and Expected Improvement accelerates the search, efficiently identifying high-EBE monolayers from ~4001 candidates and listing notable candidates in both ACl2-type and TMDC families. The approach offers a rapid, scalable screening pathway for 2D optoelectronic materials and could be extended to three-dimensional systems, enabling broader data-driven materials discovery.

Abstract

Understanding excitonic effects in two-dimensional (2D) materials is critical for advancing their potential in next-generation electronic and photonic devices. In this study, we introduce a machine learning (ML)-based framework to predict exciton binding energies in 2D materials, offering a computationally efficient alternative to traditional methods such as many-body perturbation theory (GW) and the Bethe-Salpeter equation. Leveraging data from the Computational 2D Materials Database (C2DB), our ML models establish connections between cheaply available material descriptors and complex excitonic properties, significantly accelerating the screening process for materials with pronounced excitonic effects. Additionally, Bayesian optimization with Gaussian process regression was employed to efficiently filter materials with largest exciton binding energies, further enhancing the discovery process. Although developed for 2D systems, this approach is versatile and can be extended to three-dimensional materials, broadening its applicability in materials discovery.
Paper Structure (14 sections, 3 equations, 6 figures, 2 tables)

This paper contains 14 sections, 3 equations, 6 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. Dataset
  4. Features selection
  5. Algorithm selection
  6. Workflow
  7. Results and discussion
  8. Predicting Quasi-particle energies
  9. Exciton binding energy
  10. SHAP Analysis
  11. Bayesian Optimization
  12. Conclusion
  13. CRediT author statement
  14. Competing interests

Figures (6)

  • Figure 1: Random forest-based machine learning model for predicting G$_0$W$_0$ band gaps in 2D materials.
  • Figure 2: Distribution of prediction errors for G$_0$W$_0$ band gaps using the Random Forest model, showing a peak at zero, indicating high accuracy with minimal deviation between predicted and actual values.
  • Figure 3: Exciton Binding Energy predicted using the RF model, illustrating the relationship between machine learning predictions and C2DB values, with a focus on capturing the accuracy and reliability of excitonic effects in 2D materials.
  • Figure 4: Radar plots comparing the performance of different ML algorithms in predicting exciton binding energy. The area under each curve represents the MAE and RMSE, with larger areas indicating higher error. The RF model shows the smallest area, indicating the lowest prediction error.
  • Figure 5: SHAP (SHapley Additive exPlanations) summary plot for the RF model predicting EBE.
  • ...and 1 more figures