Active learning for photonic crystals

Area of Science:

• Photonics
• Machine Learning
• Materials Science

Background:

• Photonic crystals require extensive simulations for band gap prediction.
• Current methods are computationally expensive, hindering rapid design.
• Active learning can optimize data selection for faster model training.

Purpose of the Study:

• To accelerate photonic band gap prediction using active learning.
• To integrate analytic Bayesian neural networks with uncertainty-driven sampling.
• To enable efficient surrogate modeling for photonic crystal design.

Main Methods:

• Employed analytic approximate Bayesian last layer neural networks (LL-BNNs).
• Utilized uncertainty estimates correlated with predictive error for sample selection.
• Implemented an active learning strategy prioritizing informative simulations.

Main Results:

• Achieved up to a 2.7× reduction in required training data for band gap prediction.
• Maintained predictive accuracy compared to random sampling baselines.
• Demonstrated efficiency gains by focusing on high-uncertainty regions.

Conclusions:

• Analytic LL-BNN based active learning accelerates photonic crystal design workflows.
• The approach offers a general framework for data-efficient regression in scientific machine learning.
• Enables rapid and scalable surrogate modeling for complex photonic structures.