Learning incoherent light emission steering from metasurfaces using generative models

Prasad P. Iyer, Saaketh Desai, Sadhvikas Addamane, Rémi Dingreville, Igal BrenerWACV 2023 · DOI

Introduction

Steering light is a fundamental capability in optics, typically achieved with coherent sources like lasers through phase interference. However, steering incoherent light—such as that from LEDs or photoluminescence (PL)—is significantly more challenging due to randomized wavefronts. Researchers at Sandia National Laboratories have addressed this by combining generative AI with nanophotonics to discover "non-intuitive" optical patterns that steer incoherent light with unprecedented efficiency.

Experimental setup and schematic of photoluminescence steering through structured optical pumping.

The Challenge of Incoherent Steering

Dielectric metasurfaces, composed of sub-wavelength nanoresonator arrays, can steer incoherent light when "pumped" with specific spatial patterns. The primary hurdle is the astronomical search space for these patterns, exceeding $2^{212,000}$ possible configurations. Traditional approaches relied on human intuition, such as simple sawtooth gratings, which likely miss more optimal, complex patterns hidden in the combinatorial space.

A Closed-Loop Machine Learning Framework

To navigate this space, the team developed a framework integrating Variational Autoencoders (VAEs) and active learning (AL) directly with optical hardware:

  1. Generative Modeling: A VAE was trained on 50,000 synthetic 1D profiles, including standard gratings and Bezier curves. This compressed complex patterns into a manageable low-dimensional latent space.
  2. Active Learning: A Gaussian Process (GP) surrogate model predicted pattern performance in the latent space. Acquisition functions like Maximum Expected Improvement (MEI) balanced exploration of new patterns with exploitation of known high-performers.
  3. Experimental Feedback: The ML-selected patterns were projected via a Spatial Light Modulator (SLM) onto the metasurface. A detector measured the resulting steering, and the data was fed back to refine the model.

Results and Key Insights

The ML framework successfully identified patterns that outperformed traditional designs. Notably, it discovered "perturbed" sawtooth patterns—gratings with high-frequency variations that human designers might dismiss as noise but were critical for efficiency.

Applied SLM grating patterns and the resulting photoluminescence steering heatmap.

Key achievements include:

  • Efficiency Gains: ML-generated patterns improved steering efficiency by 5.6x to 8.7x over state-of-the-art sawtooth designs at specific angles.
  • Discovery Speed: The active learning scheme found optimal patterns using only 10% of the iterations required for brute-force methods.
  • Dynamic Range: The system demonstrated continuous steering across an 80° field of view.

Conclusion and Future Directions

While the current study focused on 1D patterns, extending the framework to 2D spatial patterns could offer even greater control. The researchers also aim to use the VAE's latent space to better understand the underlying light-matter interaction physics, moving beyond the "black box" nature of the current ML models. This fusion of generative AI and experimental physics paves the way for highly efficient, tunable incoherent light sources.

Reference: Iyer, P. P., Desai, S., et al. "Learning incoherent light emission steering from metasurfaces using generative models." Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2023.

Summary updated

BibTeX
@inproceedings{iyer2023learning,
  author = {Prasad P. Iyer and Saaketh Desai and Sadhvikas Addamane and Rémi Dingreville and Igal Brener},
  title = {{Learning incoherent light emission steering from metasurfaces using generative models}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2023},
  doi = {10.1109/wacv56688.2023.00376},
}
Figure