Explainable Encoder-Prediction-Reconstruction Framework for the Prediction of Metasurface Absorption Spectra

Area of Science:

• Metasurface optics and nanophotonics.
• Computational electromagnetics and materials science.
• Artificial intelligence and machine learning applications in physics.

Background:

• Predicting metasurface absorption spectra from their structures is complex, involving intricate physics and computationally intensive simulations based on Maxwell's equations.
• Existing artificial intelligence (AI) models for spectral prediction often function as black boxes, lacking transparency and making it difficult to ascertain the physical basis of their predictions.
• The computational cost of traditional simulations hinders rapid design and development cycles for metasurfaces.

Purpose of the Study:

• To develop an explainable AI framework for predicting metasurface absorption spectra.
• To enhance understanding of the physical relationships between metasurface structures and their optical properties.
• To accelerate the design and optimization of metasurfaces by providing faster and more interpretable AI-driven predictions.

Main Methods:

• Introduction of the Explainable Encoder-Prediction-Reconstruction (EEPR) framework, which decomposes spectral prediction into feature extraction and spectra generation.
• Utilizing feature-level analysis to understand the model's internal workings and the physical correlations it learns.
• Quantitative evaluation of the EEPR framework's accuracy, speed, and generalization capabilities compared to traditional methods and other AI models.

Main Results:

• The EEPR framework achieved a 66.23% reduction in average Mean Square Error (MSE) compared to mainstream networks, with an MSE of 2.843 × 10-4.
• The model demonstrated a significant speed improvement, operating approximately 500,000 times faster than Maxwell's equation simulations (3×10-3 seconds per sample).
• The framework provides feature-level explainability, offering insights into physical properties and enabling targeted adjustments to metasurface absorption spectra by modifying features.

Conclusions:

• The EEPR framework offers a transparent and efficient approach to predicting metasurface absorption spectra, overcoming the limitations of black-box AI models.
• Feature-level explainability empowers designers with a deeper understanding of structure-property relationships, fostering trust and facilitating refined design of metasurfaces.
• This work paves the way for accelerated discovery and optimization of functional metasurfaces through interpretable AI.