Source-linked AI summary
Freeform Diffractive Metagrating Design Based on Generative Adversarial Networks
Jiaqi Jiang, David Sell, Stephan Hoyer, Jason Hickey, Jianji Yang, Jonathan A. Fan
TL;DR
Designing many high-performance, topologically complex metagratings is computationally demanding with iterative optimization. This paper trains conditional generative neural networks on topology-optimized devices and combines their outputs with electromagnetic evaluation and topology refinement, producing devices with efficiencies up to 86% and lower computational-cost scaling than iterative-only design.
Problem
Iterative-only optimization does not scale efficiently to designing large numbers of topologically complex devices across high-dimensional parameter spaces.
Method
The approach trains conditional GANs on 600 topology-optimized metagrating images, evaluates generated designs electromagnetically, and topology-refines selected devices.
Results
The highest topology-refined device reaches 86% efficiency, while 80% of topology-refined results match or approach the best iteratively optimized efficiencies within 5%.
Takeaways & Limitations
Generative neural networks can facilitate computationally efficient design of high-performance, topologically complex metasurfaces when large families of designs are needed.
Takeaways & Limitations
Performance drops for shorter wavelengths and ultra-large deflection angles because the training set does not generalize the required distinctive topological features.
Abstract
from arXiv · showhide
A key challenge in metasurface design is the development of algorithms that can effectively and efficiently produce high performance devices. Design methods based on iterative optimization can push the performance limits of metasurfaces, but they require extensive computational resources that limit their implementation to small numbers of microscale devices. We show that generative neural networks can train from images of periodic, topology-optimized metagratings to produce high-efficiency, topologically complex devices operating over a broad range of deflection angles and wavelengths. Further iterative optimization of these designs yields devices with enhanced robustness and efficiencies, and these devices can be utilized as additional training data for network refinement. In this manner, generative networks can be trained, with a onetime computation cost, and used as a design tool to facilitate the production of near-optimal, topologically-complex device designs. We envision that such data-driven design methodologies can apply to other physical sciences domains that require the design of functional elements operating across a wide parameter space.
Results and discussion
The conditional GAN generates diverse, high-efficiency metagrating designs across broad wavelength and deflection-angle spaces, and topology refinement and retraining further improve device performance. This workflow provides a computationally efficient design strategy with a one-time training cost, while performance drops remain for shorter wavelengths and ultra-large deflection angles.
- GAN training set: A 600-image training set of topology-optimized metagratings was filtered to retain devices in the top 40th percentile of efficiency for GAN training.Unfiltered training data produced worse GAN performance, indicating that exclusively high-efficiency devices were needed for training.
- GAN-generated devices: 62% efficiency was achieved by the best devices among those generated by the conditional GAN, with some generated devices exceeding 60% efficiency.Efficiencies were evaluated using a rigorous coupled-wave analysis solver, and the generated devices showed a broad efficiency distribution.
- Topology refinement: Nearly all topology-refined metagratings spanning 600‒1300 nm wavelengths and 35‒75 degree deflection angles reached efficiencies near or over 80%.Efficiencies dropped for devices designed at shorter wavelengths and ultra-large deflection angles.
- Network retraining: Over 80% of devices in the parameter space improved in efficiency after retraining the GAN with additional data from generation and topology refinement.The retraining data provide a pathway to expand the conditional GAN’s efficacy with high computational efficiency.
- Computational efficiency: The trained GAN-based generator can serve as a computationally efficient design tool by treating initial training as a one-time offline cost before generating devices across the target parameter space.The workflow combines GAN generation with refinement to reduce the computational burden of designing metagratings compared with repeated iterative-only optimization.
Conclusions
Generative neural networks enable computationally efficient design of high-performance, topologically complex metasurfaces when generating large families of devices. The GAN-based approach can extend across more device parameters and support data-driven design of complex devices and functional elements in other physical fields.
- Conclusions: Generative neural networks facilitate computationally efficient design of high-performance, topologically complex metasurfaces for generating large families of devices.The approach is appropriate because device topology and optical response are strongly interdependent, particularly for high-performance devices.
- Conclusions: The GAN-based approach could generalize beyond wavelength and deflection angle to thickness, dielectric, polarization, phase response, and incidence angle.Multifunctional devices could use high-quality multifunctional datasets or multiple discriminators for pattern synthesis.
- Conclusions: Data-driven design processes could apply to complex nanophotonic devices including dielectric and plasmonic antennas and photonic crystals.The envisioned applications extend across device design and characterization.
- Conclusions: The methods could also design devices and structured materials in acoustics, mechanics, and heat transfer across broad parameter spaces.These fields share a need for functional elements operating across a broad parameter space.