Source-linked AI summary
Global optimization of dielectric metasurfaces using a physics-driven neural network
Jiaqi Jiang, Jonathan A. Fan
TL;DR
The paper addresses computationally efficient design of high-performance metasurfaces despite immense training datasets and complex geometry–optical-response relationships. It introduces conditional topology optimization networks tied to device-efficiency improvement, with many generated devices outperforming or closely matching adjoint-based optimization, while short wavelengths and small deflection angles remain challenging.
Problem
Computationally efficient design of high-performance devices is difficult because geometry–optical-response relationships are complex and training datasets can be immense.
Method
The approach uses topology optimization networks whose training is directly tied to enhancing device efficiency.
Results
75% of devices from the conditional GLOnet have efficiencies higher than those from adjoint-based optimization, while 92% have efficiencies higher than or within 5% those from adjoint-based optimization.
Takeaways & Limitations
Conditional GLOnets are an effective and computationally efficient approach for metasurface optimization.
Takeaways & Limitations
The method does not optimally perform in certain regimes, such as short wavelengths and small deflection angles.
Abstract
from arXiv · showhide
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.
Introduction
Metasurface inverse design can produce high-performance devices but is computationally difficult to scale. The paper introduces GLOnets, which integrate adjoint calculations into conditional generative neural networks to search and refine device distributions without pre-optimized training data.
- Inverse-design methods can produce high-performance metasurfaces, but their computational cost makes large device ensembles and area devices difficult to scale.
- Existing machine-learning photonics approaches typically train on device geometries paired with optical properties, then generate new designs at low computational cost.Creating the training dataset can itself require tens to hundreds of thousands of devices, while the geometries may occupy a high-dimensional space.
- GLOnets incorporate adjoint variable calculations directly into a conditional generative neural network to generate topology-optimized devices across operating parameters.
- The network initially samples broadly across the design space, then shifts and refines its device distribution toward a cluster of high-efficiency designs.
- Physics-based electromagnetic gradients drive backpropagation, allowing the network to learn geometry–response relationships directly from simulations without a training set of known devices.
- GLOnets perform a global search for optimal devices, but non-convex optimization prevents guaranteeing that the final generated devices are globally optimal.
Methods
Conditional GLOnets generate ensembles of metagratings conditioned on wavelength and angle, then optimize their distribution using physics-based efficiency gradients. The approach surveys broad design regions while incorporating fabrication constraints and supports simultaneous optimization across operating conditions.
- Conditional generation: The generator maps noise vectors and operating conditions (λ, θ) to refractive-index device profiles, producing different device instances from different noise inputs.The output ensemble is denoted {n|λ, θ}.
- Initialization: Initialization makes the generated ensemble approximately follow the noise distribution, allowing it to span the full device design space before training.The procedure combines small random network weights with an identity shortcut that adds z to the final deconvolution output.
- Physics-driven training: Training samples batches of devices, wavelengths, and outgoing angles, then uses forward efficiencies and adjoint-derived efficiency gradients to update network weights.The objective maximizes the probability of generating high-efficiency devices across the target parameter range.
- Design constraints: A tunable loss balances efficiency enhancement against binary silicon-or-air designs, while differentiable output operations and Gaussian filtering support backpropagation and fabrication practicality.The Gaussian filter removes small pixel-level features, whereas the binarization term favors |n(m)| = 1 and can limit efficiency.
- Optimization strategy: Conditional GLOnets optimize an entire distribution of devices, unlike adjoint-based topology optimization, which optimizes a single device locally.Adjoint optimization iteratively adjusts an initial dielectric distribution toward a local maximum, with performance depending strongly on initialization.
Results and discussion
Conditional GLOnet devices perform comparably to or better than adjoint-optimized devices across wavelength–angle conditions, while using substantially less computation. Training concentrates generated designs toward high-efficiency regions, and subsequent boundary optimization usually yields only modest gains.
- The best conditional GLOnet devices compare well with or outperform the best devices from adjoint-based optimization across the tested wavelength and angle pairs.
- Conditional GLOnet outputs cluster near the high-efficiency end and often have similar geometries, unlike the highly variable efficiencies and layouts from adjoint optimization.
- Initially, generated devices broadly span the design space with mostly low-to-modest efficiencies; training clusters them and shifts the histogram toward high efficiency.At 1000 iterations, efficiencies are very high and strongly skewed toward high values.
- Conditional GLOnet uses 10x less computational cost than the benchmark adjoint-based topology-optimization calculations when simultaneously optimizing devices across wavelengths and angles.The authors note that scaling simulations and batch sizes across more computing nodes can further enhance efficacy and power.
- Only 4% of devices gain more than 5% efficiency from boundary optimization, indicating that conditional GLOnet devices are already at or near local optima.The refinement fixes binary refractive indices and considers gradients only at silicon–air boundaries.
Conclusions
Conditional GLOnets provide a computationally efficient global topology optimizer that generates and jointly improves device distributions across operating parameters. For metagratings, their best devices compare well with adjoint-based optimization, while the framework is intended to extend to other gradient-improvable physical systems.
- Conclusions: Conditional GLOnets enable global topology optimization by initially spanning the design space and optimizing device distributions.The generated distribution is collectively biased toward high-efficiency regions during training.
- Conclusions: The best conditional-GLOnet devices compare well with the best devices produced by adjoint-based topology optimization.The comparison concerns metagrating devices across wavelength and deflection-angle conditions.
- Conclusions: Conditioning GLOnets on a continuum of operating parameters enables simultaneous optimization of device ensembles and reduces overall computational cost.
- Conclusions: Future work will extend conditional GLOnets to other metasurface systems, including aperiodic broadband devices.
- Conclusions: The approach can be adapted to other photonic devices and physical systems whose performance can be improved using gradients.The loss and gradient definition can be tailored to the specific optimization target, including weighted gradients for broadband devices.
Supporting Information Available
The supporting information provides materials on GLOnet architecture and training, loss-function formulation, computational-cost comparisons, and supplementary figures.
- Supporting Information Available: Supporting information includes the GLOnet network architecture and training process.
- Supporting Information Available: Supporting information includes the formulation of the conditional GLOnet loss function.
- Supporting Information Available: Supporting information includes a computational-cost comparison between conditional GLOnet and adjoint-based topology optimization.
- Supporting Information Available: Supporting information includes Figures S1 to S4.