Source-linked AI summary
Deep neural networks for the evaluation and design of photonic devices
Jiaqi Jiang, Mingkun Chen, Jonathan A. Fan
TL;DR
Photonics simulation and inverse design involve nonlinear relationships, expensive forward simulations, and difficult non-convex searches. This Review explains discriminative and generative deep neural networks for surrogate electromagnetic modeling and device optimization. It reports that trained networks can provide reasonably accurate surrogate models for limited-complexity nanophotonic systems, while also emphasizing their computational cost and accuracy limits.
Problem
Photonic forward simulations can be resource-intensive, while inverse design must search a vast non-convex space for devices meeting desired objectives and constraints.
Method
The Review synthesizes machine-learning principles and photonics research on discriminative surrogate models, generative device distributions, and neural-network-based inverse design.
Results
Trained discriminative networks serve as reasonably accurate surrogate models for nanophotonic systems with limited complexity, while generative networks support device-distribution modeling and inverse-design strategies.
Takeaways & Limitations
Deep learning offers complementary tools for accelerating electromagnetic evaluation and exploring photonic inverse-design spaces, rather than replacing conventional simulators.
Takeaways & Limitations
Neural networks require training sets ranging from thousands to millions of devices and cannot guarantee accuracy when exact physics calculations are required.
Abstract
from arXiv · showhide
The data sciences revolution is poised to transform the way photonic systems are simulated and designed. Photonics are in many ways an ideal substrate for machine learning: the objective of much of computational electromagnetics is the capture of non-linear relationships in high dimensional spaces, which is the core strength of neural networks. Additionally, the mainstream availability of Maxwell solvers makes the training and evaluation of neural networks broadly accessible and tailorable to specific problems. In this Review, we will show how deep neural networks, configured as discriminative networks, can learn from training sets and operate as high-speed surrogate electromagnetic solvers. We will also examine how deep generative networks can learn geometric features in device distributions and even be configured to serve as robust global optimizers. Fundamental data sciences concepts framed within the context of photonics will also be discussed, including the network training process, delineation of different network classes and architectures, and dimensionality reduction.
1 Introduction
Photonics combines rich geometry–response relationships with difficult forward and inverse problems. The Review presents deep neural networks as accessible tools for surrogate electromagnetic modeling and device optimization.
- Forward analysis computes electromagnetic response from structure, whereas inverse design seeks structures meeting desired responses and constraints.Forward simulation is established but can be computationally expensive; inverse design is difficult because its solution space is non-convex.
- Deep neural networks model highly nonlinear relationships and can address both photonic forward and inverse problems.The Review examines their use as high-speed surrogate Maxwell solvers and device optimizers.
- Deep learning is motivated by proven performance on complex phenomena, broad accessibility, available electromagnetic simulators, and scalable computational resources.Open-source software, simulation tools, distributed computing, and specialized hardware support training and evaluation.
- The Review covers discriminative and generative networks, surrogate modeling, population-based inverse design, global optimization, and research practices.
2 Principles of deep neural networks
Deep neural networks learn nonlinear mappings through layered neuron architectures, with network classes and data representations selected for photonic tasks. Discriminative models approximate response functions, while generative models learn device-layout distributions.
- Neural networks stack nonlinear processing layers to capture increasingly abstract features and fit complex input–output relationships.Training iteratively adjusts network weights to reduce a loss measured against ground-truth training values.
- Photonic data use physical variables x, such as geometry and materials, and physical responses y, such as spectra and performance characteristics.For electromagnetics, a given x maps to a single y.
- Discriminative networks learn y = f(x), supporting regression and classification and serving as surrogate models for forward electromagnetic problems.
- Generative networks use latent variables z and conditioning labels θ to produce device distributions P(x|θ) matching the training distribution.Unlike discriminative networks’ single-valued mappings, generative networks produce one-to-many outputs and learn layout-distribution properties.
- Architectures are tailored to data structures: convolutional networks process spatial image features, while other structures motivate graph or sequential models.Freeform devices can be represented as images and processed with convolutional layers, sometimes combined with fully connected layers for discrete outputs.
- Mini-batch gradient descent balances computational cost with an approximation to the full-training-set gradient.It randomly samples more than one but fewer than all training examples per iteration.
3 Surrogate modeling and inverse design with discriminative models
Discriminative neural networks model electromagnetic responses as fast surrogate solvers across microwave, guided-wave, and nanophotonic systems, then support several inverse-design strategies. Their usefulness is balanced by costly training, imperfect accuracy, non-unique inverse mappings, and the need for dimensionality reduction as design complexity grows.
- Surrogate modeling: Neural-network modeling progressed from microwave devices in the 1990s to guided-wave photonics and, more recently, nanostructured optical media.Applications include lumped components, waveguides, photonic crystals, scatterers, metasurfaces, and metamaterials.
- Surrogate modeling: Fully connected, convolutional, and graph neural networks map device descriptions to scattering spectra, internal polarization fields, and microwave transmission characteristics.Inputs can be shell thicknesses, voxelized geometries, or graph representations containing resonator geometries and positions.
- Surrogate modeling: Trained discriminative networks provide reasonably accurate surrogates for systems with limited complexity, but some cases remain poorly modeled and training requires tens of thousands of full-wave simulations.Once trained, networks can calculate responses orders of magnitude faster than full-wave solvers.
- Inverse design: Backpropagation performs inverse design by fixing network weights and iteratively adjusting device geometry to reduce the response error relative to a desired target.Hybrid methods can combine global search, such as genetic algorithms, with local gradient-based refinement; inverse mappings may contain multiple branches.
- Inverse design: Direct inverse networks are difficult when multiple layouts produce the same optical response, because training examples from different branches can prevent convergence to any branch.Backpropagation can also fail through surrogate-model inaccuracies or undesired local optima, motivating larger training sets, varied initial layouts, or alternative optimizers.
- Dimensionality reduction: Dimensionality reduction addresses sampling demands by compressing design spaces through PCA, Fourier representations, or autoencoders before discriminative modeling.Autoencoders learn nonlinear low-dimensional latent representations while minimizing reconstruction error, unlike PCA's linear transformation.
4 Generative networks
Generative networks model distributions of photonic devices and support conditional generation, latent-space optimization, and global topology optimization. Unlike training-set-based methods limited by extrapolation, GLOnets search for globally optimal devices without a training set and can incorporate practical constraints.
- Generative model types: Generative networks learn geometric or statistical features of device distributions for photonic inverse design.VAEs sample explicit latent-variable distributions, whereas GANs learn implicit training-set distributions.
- Generative model types: GANs use adversarial training to make generated devices resemble training-set structures, with conditional discriminators enforcing target spectral responses.Topology-optimized training data restricts learning to high-performance freeform geometries rather than extraneous design-space regions.
- Adapting generative networks to photonic systems: Conditional generative networks interpolate device layouts between labeled training examples, such as operating wavelengths or deflection angles.GANs trained on selected wavelength–deflection pairs generated layouts across continuous parameter ranges, although early devices required optimization refinement.
- Global topology optimization networks: GLOnets reframe topology optimization as dataless generative-network training that shifts a device distribution toward the global optimum.Generated devices are evaluated with a Maxwell solver, and performance metrics and gradients are backpropagated to update network weights.
5 Future research directions and practices
Deep neural networks offer complementary strengths for photonic simulation and inverse design, but future progress depends on physics-integrated methods, faster simulators, streamlined training, and open community practices.
- Current capabilities: Discriminative networks can act as fast surrogate Maxwell solvers, while generative models support freeform inverse design through learned geometry and dataless training.These approaches are presented as complementary tools rather than replacements for conventional electromagnetic simulation.
- Scope and limitations: Conventional electromagnetic simulators remain necessary because neural networks require large training sets, cannot guarantee accuracy, and are unsuitable when exact physics calculations are required.For low-dimensional problems, classical statistical, machine-learning, and optimization tools may be equally effective without extensive hyperparameter tuning.
- Current capabilities: Neural networks operate orders of magnitude faster than conventional simulators and can efficiently model related device variants with a single conditional network.Their regression capabilities also extend to complex, high-dimensional systems, and inverse-design approaches can produce higher-performing devices.
- Future research directions: Future methods should integrate Maxwell-equation structure with machine learning and develop faster electromagnetic simulators for larger training sets and simulation batches.Application-specific solvers are anticipated to support ultra-fast learning workflows as device complexity increases.
- Future research directions: Training workflows should be streamlined through transfer learning and meta-learning, reducing repeated training from scratch and simplifying algorithm setup.Transferring weights from concentric-shell scatterers to dielectric-stack prediction improved the latter network’s training accuracy.
- Community practices: Openly shared algorithms, benchmark problems, training sets, and code could support broader progress in photonics machine learning.MetaNet is described as a repository containing over 100,000 freeform metagrating designs and inverse-design codes.