Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Accelerated prediction of dielectric functions in solar cell materials with graph neural networks

Caden Ginter, Kamal Choudhary, Subhasish Mandal

TL;DR

This work demonstrates an ALIGNN-based framework to predict the complex dielectric function $D(\omega)$ directly from crystal structures using TB-mBJ data, enabling rapid SLME screening across millions of materials. Two independent models predict the real and imaginary parts of $D(\omega)$, achieving robust accuracy (MAE:MAD ~ 0.36–0.55 and CSD ~ 0.43–0.55) and yielding ML-augmented SLME predictions with $\text{MAE} \approx 1.95\%$ and $R^2 \approx 0.84$. Application to the Alexandria database reveals clear elemental and perovskite-related trends, notably that vanadium-based perovskites exhibit a substantially higher fraction of high-SLME compounds, underscoring their promise for optoelectronic applications. The approach enables scalable, high-throughput discovery of dielectric-function–driven materials and provides design guidance for next-generation solar absorbers and related devices.

Abstract

We present an atomistic line graph neural network (ALIGNN) model for predicting dielectric functions directly from crystal structures. Trained on $\sim$7000 dielectric functions from the JARVIS-DFT database computed with a meta-GGA exchange-correlation functional, the model accurately reproduces spectral features, including peak intensities and overall line shapes, while enabling efficient high-throughput screening. Applied to the recently developed Alexandria materials database, containing over four hundred thousand insulating materials, we uncover a clear elemental trend, with vanadium emerging as a strong indicator of materials with high-spectroscopic limited maximum efficiency (SLME). In particular, vanadium-based perovskite materials show a substantially higher fraction of high-SLME compounds compared to the database average, underscoring their promise for optoelectronic applications.

Accelerated prediction of dielectric functions in solar cell materials with graph neural networks

TL;DR

This work demonstrates an ALIGNN-based framework to predict the complex dielectric function directly from crystal structures using TB-mBJ data, enabling rapid SLME screening across millions of materials. Two independent models predict the real and imaginary parts of , achieving robust accuracy (MAE:MAD ~ 0.36–0.55 and CSD ~ 0.43–0.55) and yielding ML-augmented SLME predictions with and . Application to the Alexandria database reveals clear elemental and perovskite-related trends, notably that vanadium-based perovskites exhibit a substantially higher fraction of high-SLME compounds, underscoring their promise for optoelectronic applications. The approach enables scalable, high-throughput discovery of dielectric-function–driven materials and provides design guidance for next-generation solar absorbers and related devices.

Abstract

We present an atomistic line graph neural network (ALIGNN) model for predicting dielectric functions directly from crystal structures. Trained on 7000 dielectric functions from the JARVIS-DFT database computed with a meta-GGA exchange-correlation functional, the model accurately reproduces spectral features, including peak intensities and overall line shapes, while enabling efficient high-throughput screening. Applied to the recently developed Alexandria materials database, containing over four hundred thousand insulating materials, we uncover a clear elemental trend, with vanadium emerging as a strong indicator of materials with high-spectroscopic limited maximum efficiency (SLME). In particular, vanadium-based perovskite materials show a substantially higher fraction of high-SLME compounds compared to the database average, underscoring their promise for optoelectronic applications.

Paper Structure

This paper contains 14 sections, 7 equations, 9 figures.

Table of Contents

  1. Introduction
  2. Methods
  3. Machine Learning Models
  4. Data Acquisition
  5. Model Training and Usage
  6. Computations of SLME
  7. Model Evaluation
  8. Alexandria Database Analysis
  9. Results & Discussions
  10. Model Evaluation
  11. Analysis on Alexandria Database
  12. Conclusions
  13. Acknowledgments
  14. References

Figures (9)

  • Figure 1: Flow chart demonstrating the proposed technique, where software and data transfer components are represented in solid and dashed boxes respectively.
  • Figure 2: Distribution of MAE:MAD scores across the test set for the proposed model in obtaining (a) imaginary part ($\mathfrak I [D(\omega)]$) & (b) real part ($\mathfrak R [D(\omega)]$) of the dielectric function; quartiles are indicated as shaded regions.
  • Figure 3: 5 best (top) and 5 worst (bottom) examples from the performance of $\mathfrak I [D(\omega)]$ model. Red and black represent predicted and reference data respectively.
  • Figure 4: 5 best (top) and 5 worst (bottom) examples from the performance of $\mathfrak R [D(\omega)]$ model. Red and black represent predicted and reference data respectively.
  • Figure 5: Distribution of the absolute errors in Spectroscopic Limited Maximum Efficiency (SLME) across the test set with MAE of 1.95%.
  • ...and 4 more figures