Shaping freeform nanophotonic devices with geometric neural parameterization

Tianxiang Dai, Yixuan Shao, Chenkai Mao, Yu Wu, Sara Azzouz, You Zhou, Jonathan A. Fan · 2025

Read Shaping freeform nanophotonic devices with geometric neural parameterization by Tianxiang Dai, Yixuan Shao, Chenkai Mao, Yu Wu, Sara Azzouz, You Zhou, Jonathan A. Fan (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-17499.txt.

Introduction

Tianxiang Dai and six co-authors present Neuroshaper, a way to design tiny light devices by letting a neural network hold the shape. Their paper appeared in two thousand twenty-five. Freeform nanophotonic parts can beat standard designs, but shaping them while honoring factory limits is hard. Each tweak must keep the part buildable.

In Neuroshaper, the layout is written as an analytic neural network rather than pixels. That over-rich form smooths the search space for the optimizer. Local rules can reshape the whole device in a controlled way, including its topology. The team shows the method on a spread of devices, in simulation and in built hardware. Shapes stayed within design and manufacturing constraints throughout. The work suggests neural shape holders can widen what optical designers dare to try.

Catalog record

CREATORTianxiang Dai, Yixuan Shao, Chenkai Mao, Yu Wu, Sara Azzouz, You Zhou, Jonathan A. Fan
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE60 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-17499

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Shaping freeform nanophotonic devices with geometric neural parameterization Tianxiang Dai, Yixuan Shao, Chenkai Mao, Yu Wu, Sara Azzouz, You Zhou, Jonathan A. Fan Abstract Nanophotonic freeform design has the potential to push the performance of optical components to new limits, but there remains a challenge to effectively perform optimization while reliably enforcing design and manufacturing constraints. We present Neuroshaper, a framework for freeform geometric parameterization in which nanophotonic device layouts are defined using an analytic neural network representation. Neuroshaper serves as a qualitatively new way to perform shape optimization by capturing multi-scalar, freeform geometries in an overparameterized representation scheme, enabling effective optimization in a smoothened, high dimensional geometric design space. We show that Neuroshaper can enforce constraints and topology manipulation in a manner where local constraints lead to global changes in device morphology. We further show numerically and experimentally that Neuroshaper can apply to a diversity of nanophotonic devices. The versatility and capabilities of Neuroshaper reflect the ability of neural representation to augment concepts in topological design. 2022 1 Introduction Free space and on-chip nanophotonic systems have emerged as revolutionary platforms for controlling, routing, filtering, and transducing photons using subwavelength-scale structured media [ 1 , 2 , 3 , 4 , 5 , 6 , 7 , 8 , 9 , 10 ] . The immense versatility of these platforms largely derives from the strong relationship between device geometry and optical response at subwavelength scales, which enables an exponentially large design space for optical manipulation. Traditionally, these structure-function relations have derived from the use of simple, physically intuitive geometries in the form of nanoscale waveguides and resonators, together with their tailored coupling [ 11 , 12 , 13 , 14 , 15 , 16 ] . Such simplifications have led to robust and scalable design rules for many classes of photonic systems, however, these concepts also pose fundamental limitations in performance and device footprint. To overcome these limits, a wide range of freeform inverse design algorithms have been proposed that utilize a much larger design space of possible device geometries to achieve an objective [ 17 , 18 , 19 , 20 ] . The most advanced and computationally efficient algorithms utilize gradients calculated using the adjoint variables method or autodifferentiation to perform iterative optimization [ 21 , 22 , 23 , 24 , 25 , 26 , 27 ] . A core feature of all inverse design methodologies is the specification of a geometric parameterization scheme that describes the device layout. The choice of parameterization scheme is critical to reliably enforce constraints to the device geometry, in a manner that enables reliable device manufacturability and electromagnetic mode engineering control, and to ensure proper convergence of the optimization algorithm. Constraints of interest range from feature size and curvature tolerances to specification of the layout topology and connectivity. There are currently three primary classes of parameterization schemes used in freeform inverse design. One is pixel-based representations [ 28 , 29 ] , in which the device is described as an ultra-fine grid of pixels, each with a distinctly tailored dielectric constant value. The second is level sets [ 30 , 31 ] , in which layouts are implicitly represented as iso-contours of scalar functions [ 31 ] . The third is explicit geometric parameterization [ 32 , 33 , 34 ] , in which simple analytic functions are defined to map latent variables to constrained geometric shapes. While these parametrization concepts have been utilized effectively in certain design limits, each presents distinct trade-offs between optimization capability and constraints compliance. Pixel-based methods are the most straightforward to adapt to gradient-based optimizers such as the adjoint variables method. However, the use of spatial filters [ 35 ] , density filters, and projection methods [ 36 ] to impose constraints lead to either poor fabrication compliance or the imposition of overly restrictive constraints. Traditional level set methods offer improved topology control, but they struggle with multi-scale features, require careful numerical schemes to maintain stability [ 37 ] , and are limited in performance by their reliance on local shape derivatives [ 38 ] . Explicit parameterization schemes have the benefit of hard coding strict design constraints in the parameterization functions themselves, but they are ultimately overly restrictive and suffer from limited topology control. Qualitatively new geometric parameterization concepts are needed to generally and robustly enforce multiple competing objectives and constraints in the device, while maintaining access to the large freeform design landscape. In this work, we present Neuroshaper, a neural network-based approach to freeform geometric representation that can streamline with existing gradient-based freeform optimizers. Neuroshaper is inspired by recent work in computational imaging and computer graphics on neural representations [ 39 , 40 ] and specifically the development of multi-resolution techniques for representing complex signals [ 41 ] . Unlike traditional deep learning models that are trained on large datasets, Neuroshaper functions as a differentiable geometric parameterization scheme; its neural network parameters directly define the device geometry and are optimized using physics-based gradients and constraint penalties. It addresses limitations with existing parameterization schemes in photonics in multiple ways. First, it takes advantage of the immense expressivity of neural networks paired with learnable coordinate encoding to capture a wide range of multi-scalar freeform shapes. Second, it is differentiable and uses network loss function engineering and backpropagation to robustly and flexibly enforce constraints. Third, by capturing geometric device parameters using large numbers of neural network weights, Neuroshaper serves as an overparameterized representation scheme that captures a smoothened, high-dimensional geometric design space. In this manner, gradient-based optimizers are able to more easily traverse the design space and detour around forbidden zones when constraints are encountered (Figure 1 a). An example of this optimization trajectory is included in the Supplementary Movie 1. Fourth, Neuroshaper captures a global representation of geometric shapes, such that local geometric constraints map onto global modifications to the device layout. With these collective advantages, devices designed using Neuroshaper exhibit better handling and enforcement of constraints, and also similar or better performance, compared to conventionally parameterized freeform devices. Figure 1: Neuroshaper overview. a) Visualization of gradient-based optimization in the high dimensional Neuroshaper parameterization space. The optimizer avoids forbidden zones imposed by constraints to reach a suitable local minima. b) Overview of the Neuroshaper framework for nanophotonic device design. The pipeline consists of: (1) Neural design representation combining sampling points with lookup tables and multilayer perceptrons to generate continuous level set functions; (2) Level set operations to obtain binary/grayscale patterns; (3) Multi-resolution conversion between high ( M_{1} ) and low ( M_{2} ) resolutions; (4) Design evaluation based on geometric constraints and electromagnetic performance; (5) Parameter gradient computation via backpropagation. The full framework for Neuroshaper-enabled freeform inverse design is illustrated in Figure 1 b. Freeform layouts are produced using a novel neural architecture that combines multi-resolution hash encoding with multilayer perceptrons to generate continuous level set functions. These functions undergo level set operations to produce binary or grayscale patterns suitable for both constraint evaluation and simulation. The ability for Neuroshaper to support multi-resolution representations enables high resolution representations to be used for geometric constraint evaluation ( M_{1} ) and low resolution versions for electromagnetic simulations ( M_{2} ). Global loss functions accounting for feature size, curvature,

Read the complete text