Global optimization of dielectric metasurfaces using a physics-driven neural network (1906.04157v2)

Abstract: We present a global optimizer, based on a conditional generative neural network, which can output ensembles of highly efficient topology-optimized metasurfaces operating across a range of parameters. A key feature of the network is that it initially generates a distribution of devices that broadly samples the design space, and then shifts and refines this distribution towards favorable design space regions over the course of optimization. Training is performed by calculating the forward and adjoint electromagnetic simulations of outputted devices and using the subsequent efficiency gradients for backpropagation. With metagratings operating across a range of wavelengths and angles as a model system, we show that devices produced from the trained generative network have efficiencies comparable to or better than the best devices produced by adjoint-based topology optimization, while requiring less computational cost. Our reframing of adjoint-based optimization to the training of a generative neural network applies generally to physical systems that can utilize gradients to improve performance.

• The paper introduces GLOnets, a physics-driven neural network that reframes traditional topology optimization for high-efficiency metasurface design.
• It leverages adjoint electromagnetic simulations to guide device generation, achieving superior or comparable efficiency in 75% of cases.
• The method significantly reduces computational cost by concurrently optimizing multiple devices across varying operational conditions.

Overview of Global Optimization of Dielectric Metasurfaces Using a Physics-driven Neural Network

This paper addresses the global optimization of dielectric metasurfaces through a novel approach that incorporates a conditional generative neural network as a global optimizer. The methodology presented offers a new paradigm by integrating a physics-driven mechanism into neural networks for the design of high-efficiency metasurfaces across various operating conditions, with reduced computational demands compared to traditional optimization methods.

Metasurfaces, which are intricate subwavelength structures, have seen extensive application in fields such as imaging and sensing due to their ability to manipulate electromagnetic waves. However, designing these surfaces to achieve specific electromagnetic responses has been a computationally intensive task, traditionally tackled through techniques like adjoint-based topology optimization. While effective, these methods can be computationally prohibitive, especially when expanded to large ensembles or large-area devices.

Main Methodological Contributions

The authors introduced the Global Topology Optimization Networks (GLOnets), which reframes the conventional adjoint-based topology optimization as a global search mechanism within a neural network framework. The key features of the GLOnet approach include:

• Generative Neural Network: The network starts by generating a broad spectrum of device distributions within the design space. This initial distribution then evolves, powered by training with electromagnetic simulations, towards regions that offer high performance in terms of device efficiency.
• Physics-driven Training: Adjoint electromagnetic simulations provide efficiency gradients that guide the backpropagation of the network, aligning device generation with enhanced efficiency.
• Conditional Inputs: The GLOnet incorporated wavelength and deflection angle as conditional inputs to generate diverse devices designed for varying operational parameters simultaneously.

Performance Evaluation

The paper provides a comprehensive comparison of the GLOnet approach to traditional adjoint-based topology optimization, focusing on silicon metagratings as a model system. Remarkably, it was observed that the devices produced by the GLOnet achieved efficiencies that were comparable or superior to those optimized by adjoint methods. The statistical analysis demonstrated that GLOnet-generated devices possess advantageous efficiencies in 75% of the cases compared to their adjoint-optimized counterparts, highlighting the effectiveness of this novel optimization framework.

Computational Efficiency

One of the most significant advantages highlighted by the authors is the notable reduction in computational cost achieved using GLOnets. This is primarily due to the ability of the network to simultaneously optimize multiple devices for various operating conditions in a single training session, leveraging the cross-talk inherent in the neural network architecture. This attribute positions GLOnets as a viable and scalable alternative for large-scale metasurface design optimization.

Implications and Future Directions

The theoretical implications of this work suggest broader applicability of neural-network-driven optimization in different physical design systems, wherever performance gradients can enhance the optimization process. Practically, the methodology presents a path toward more computationally efficient design processes in electromagnetics and potentially other domains reliant on complex structural optimization.

Future research could expand the application of GLOnets to other types of metasurface structures, such as aperiodic broadband devices, where design complexity is even more pronounced. Adjustments to the neural network architecture and training process, as suggested by the authors, may further enhance the optimization outcomes, particularly in challenging wavelength-angle regimes.

In conclusion, this paper represents a significant contribution to the field of nanophotonic device design by proposing a cutting-edge and computationally efficient means of tackling complex optimization challenges through neural networks infused with physical simulation capabilities.