Physics-informed Neural Network (PINN) to Predict Vibrational Stability of Inorganic Semiconductors

TL;DR

This paper introduces a physics-informed neural network that predicts vibrational stability of inorganic semiconductors, integrating fundamental physics constraints to improve accuracy in high-throughput materials screening.

Contribution

The study develops a PINN that incorporates Born stability criteria into its loss function, enhancing vibrational stability predictions over existing models.

Findings

Achieved an F1-score of 0.83 on stability classification.

Attained an AUC-ROC of 0.82 on benchmark dataset.

Outperformed existing models, especially in identifying unstable materials.

Abstract

We tackle the challenge of predicting vibrational stability in inorganic semiconductors for high-throughput screening, an essential attribute for evaluating synthesizability alongside thermodynamic stability, frequently missing in prominent materials databases. We create a physics-informed neural network (PINN) that incorporates the Born stability requirements directly into its loss function. This integration is a key learning constraint since it only allows the model to make predictions that do not violate fundamental physics. The model shows consistent and improved performance, having been trained on a dataset of 2112 inorganic materials with validated phonon spectra, and getting an F1-score of 0.83 for both stable and unstable classes. The model shows an AUC-ROC of 0.82 on a benchmark dataset of 1296 materials. Our PINN surpasses the best models in comparative tests, especially when it comes to accurately identifying unstable materials, which is crucial for a stability filter. This work offers a comprehensive screening tool for identifying materials and a methodology for incorporating domain knowledge to enhance predictive accuracy in materials informatics.