HiLAB: A Hybrid Inverse-Design Framework

Area of Science:

• Nanophotonics
• Computational electromagnetics
• Materials science

Background:

• Inverse design of nanophotonic structures is crucial for advanced optical devices.
• Conventional methods like topological optimization (TO) can be computationally expensive and prone to local optima.
• Designing multifunctional nanophotonic devices with specific optical properties remains a challenge.

Purpose of the Study:

• To introduce HiLAB (Hybrid inverse-design with Latent-space learning, Adjoint-based partial optimizations, and Bayesian optimization), a new paradigm for efficient nanophotonic inverse design.
• To address the challenge of multifunctional device design by generating diverse freeform configurations at reduced simulation costs.
• To accelerate the discovery of fabrication-friendly nanophotonic devices.

Main Methods:

• Combining early-terminated topological optimization (TO) with a Vision Transformer-based variational autoencoder (VAE) and Bayesian optimization.
• Compressing initial structures into a compact latent space using the VAE for co-optimization of geometry and physical hyperparameters.
• Utilizing shortened adjoint-driven TO runs with randomized physical parameters to generate robust initial structures.

Main Results:

• HiLAB systematically explores near-global optima with significantly fewer electromagnetic simulations compared to conventional TO pipelines.
• The trained VAE can be reused for alternative objectives or constraints by adjusting the acquisition function.
• Demonstrated efficacy by designing an achromatic beam deflector with balanced diffraction efficiencies and mitigated chromatic aberrations.

Conclusions:

• HiLAB provides a flexible and efficient platform for robust, multi-parameter nanophotonic designs.
• The method accelerates the discovery process for next-generation nanophotonic devices.
• HiLAB offers a significant reduction in computational cost (an order of magnitude fewer simulations) for complex nanophotonic designs.