Embedding-driven physics informed neural network for predicting optical constants across materials.

Area of Science:

• Materials Science
• Computational Physics
• Optics

Background:

• Predicting optical constants (refractive index n(λ) and extinction coefficient k(λ)) is crucial for material characterization.
• Traditional methods can be time-consuming and require extensive experimental data.
• Developing accurate and efficient predictive models is an ongoing challenge.

Purpose of the Study:

• Introduce a deep learning framework for predicting wavelength-dependent optical constants.
• Encode material-specific information using learnable embeddings.
• Explore a physics-informed extension incorporating reflectance loss.

Main Methods:

• Developed an embedding-driven deep learning architecture.
• Utilized discrete material identifiers and normalized wavelengths as input.
• Implemented a differentiable reflectance loss based on the Fresnel equation for physics-informed training.

Main Results:

• The model accurately predicts optical constants across various material classes.
• Learned embeddings effectively capture essential dispersion characteristics.
• Physics-informed reflectance loss showed minimal impact on predictive accuracy for most materials.
• Lower accuracy observed for structurally diverse oxide materials.

Conclusions:

• The proposed deep learning framework offers a scalable approach for predicting optical constants.
• The embedding-driven architecture sufficiently captures material optical behavior.
• The framework has potential for integration into optical simulations, materials screening, and inverse design.