Interpretable Artificial Intelligence for Topological Photonics

Ali Ghorashi ★, Sachin Vaidya ★, Ziming Liu, Charlotte Loh, Thomas Christensen, Max Tegmark, Marin Soljačić · 2025

Read Interpretable Artificial Intelligence for Topological Photonics by Ali Ghorashi ★, Sachin Vaidya ★, Ziming Liu, Charlotte Loh, Thomas Christensen, Max Tegmark, Marin Soljačić (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-2505-10485.txt.

Introduction

Ali Ghorashi and co-authors use machine learning to learn the band topology of photonic crystals as explicit formulas, in a 2025 paper. These crystals can guide and trap light in ways that resist disorder, but designing them for a chosen topology is hard. The map from a crystal's layout to its band class is strongly nonlinear.

A compact neural network of a special kind learned to classify the lowest bands of crystals with twofold symmetry. Because that network can be read as a formula, the team extracted plain algebraic rules that link the crystal's makeup to its band class. The rules kept the network's full accuracy, and they let the team design crystals with target bands directly.

Catalog record

CREATORAli Ghorashi ★, Sachin Vaidya ★, Ziming Liu, Charlotte Loh, Thomas Christensen, Max Tegmark, Marin Soljačić
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE21 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-10485

Browse the full library · About the Tower of Babel

Full text

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

Interpretable Artificial Intelligence for Topological Photonics Ali Ghorashi ★, Sachin Vaidya ★, Ziming Liu, Charlotte Loh, Thomas Christensen, Max Tegmark, Marin Soljačić Abstract Topological photonic crystals (PhCs) offer robust, disorder-resistant modes engendered by nontrivial band symmetries, but designing PhCs with prescribed topological band properties remains a challenge due to the complexity involved in mapping the continuous real-space design landscape of photonic crystals to the discrete output space of band topology. Here, we introduce a new approach to address this problem, employing Kolmogorov–Arnold networks (KANs) to predict and inversely design the band symmetries of two-dimensional PhCs with two-fold rotational ( C 2 C_{2} ) symmetry. We show that a single-hidden-layer KAN, trained on a dataset of C 2 C_{2} -symmetric unit cells, achieves 99% accuracy in classifying the symmetry eigenvalues of the lowest transverse-magnetic band. For interpretability, we use symbolic regression to extract eight algebraic formulas that characterize the band symmetry classes in terms of the Fourier components of the dielectric function. These formulas not only retain the full predictive power of the network but also enable deterministic inverse design. Applying them, we generate 2,000 photonic crystals with target band symmetries, achieving over 92% accuracy even for high-contrast, experimentally realizable structures beyond the training domain. Topological photonic crystals (PhCs) have emerged as promising platforms for achieving robust, disorder-resistant light transport, supporting unidirectional edge states, defect-localized modes, and scatter-free interfaces that are resilient to fabrication imperfections and environmental perturbations Wang et al. 2009 ; Rechtsman et al. 2013 ; Skirlo et al. 2015 ; Schulz et al. 2021 . These features make PhCs attractive for applications in waveguiding, lasing, optical isolation, and integrated photonic devices Lu et al. 2014 ; Ozawa et al. 2019 ; Jalali Mehrabad et al. 2023 ; Smirnova et al. 2020 ; Ota et al. 2020 ; Schulz et al. 2021 . Following early photonic realizations of topological degeneracies and Chern phases Lu et al. 2013 ; Haldane and Raghu 2008 ; Wang et al. 2009 ; Rechtsman et al. 2013 ; Vaidya et al. 2020 , the field has expanded to include higher-order He et al. 2020 ; Xie et al. 2018 , non-Hermitian Parto et al. 2020 ; Pontula et al. 2024 , quantum Jalali Mehrabad et al. 2023 ; Sloan et al. 2025 , and nonlinear Smirnova et al. 2020 topological phenomena. Large-scale symmetry-based surveys of two-dimensional PhCs have previously mapped regions of design space conducive to Dirac points, corner states, Chern bands, and fragile topological phases, revealing a rich and diverse topological landscape Ghorashi et al. 2024 ; Vaidya et al. 2023 ; Bradlyn et al. 2017 ; Po et al. 2018 ; Song et al. 2020 ; Xie et al. 2018 ; He et al. 2020 . Despite this progress, inverse design of PhCs with prescribed topological bands remains a fundamental challenge Kim et al. 2023 ; Christiansen and Sigmund 2021 . This difficulty arises primarily from the absence of general principles linking real-space geometry to band topology—a challenge analogous to predicting material properties from chemical composition and crystal structure. Conventional photonic inverse design methods often rely on heuristic parameter sweeps, random searches, or black-box optimization, offering limited physical insight and no guarantees of global optimality. Here, we pursue two complementary objectives for enabling interpretable design of topological photonic bands: (1) to develop a machine learning model that yields interpretable phenomenological laws mapping photonic crystal geometries directly to their band topology, and (2) to leverage this model for the inverse design of PhCs with deterministic topological character. We focus on the simplest nontrivial plane group, p 2, in which the C_{2} eigenvalues at high-symmetry points fully determine the topological classification of the bands. To balance interpretability and accuracy, we employ the recently introduced Kolmogorov–Arnold networks (KANs) Liu et al. 2024 with a single hidden layer, trained on a large dataset of C_{2} -symmetric PhC unit cells spanning continuous variations in geometry and dielectric contrast. We augment this dataset using insights from topological quantum chemistry Kruthoff et al. 2017 ; Bradlyn et al. 2017 ; Po et al. 2017 to increase the data available for training. Leveraging the symbolic regression capabilities of KANs, we extract compact algebraic expressions that predict the topological classes of the lowest band with 99% accuracy. And, using these formulas, we create an objective function on which we perform gradient descent, enabling deterministic inverse design for arbitrary target topological classes, achieving over 92% design accuracy even in high-contrast regimes. Our approach bridges high-accuracy prediction and physical interpretability, making it deployable for both high-throughput discovery of topological PhCs and as a tool for understanding the underlying structure–property relationships in photonic topology. Artificial intelligence for photonic topology. To date, machine learning has achieved significant progress in photonics and the study of topological materials Peurifoy et al. 2018 ; Claussen et al. 2020 . In this section, we briefly review some of the major advances to contextualize our work. Most efforts to predict photonic crystal band properties have focused on band dispersions, using images of the periodic dielectric function as inputs. Motivated by the success of convolutional neural networks (CNNs) in image recognition Rawat and Wang 2017 ; Guo et al. 2017 , CNNs have been effectively applied to predict band structures of two-dimensional photonic crystals Christensen et al. 2020 . In contrast, our approach—detailed in the following sections—uses the Fourier components of the dielectric function as input parameters instead of image representations. This choice is motivated by two key factors. First, it significantly enhances interpretability: a model based on a small number of Fourier components is far easier to analyze than one which uses 1,024 image pixels (e.g., a 32\times 32 input grid). Second, from a physical standpoint, the perturbation theory suggests that the lowest photonic bands are governed primarily by the lowest Fourier components Kittel and McEuen 2018 ; Kaxiras and Joannopoulos 2019 , indicating that the primary physical information necessary for the analysis of both band dispersion and band symmetry is contained in this representation, despite its dimensional sparsity. In the context of topological materials, CNNs have also been employed to predict topological invariants, such as winding numbers in one-dimensional models Zhang et al. 2018 , with high accuracy. Notably, Zhang et al Zhang et al. 2018 showed through a subsequent analysis of the trained networks’ weights that they had implicitly learned meaningful physical laws. While instructive, this approach relies on interpreting the network’s weights, which is not always feasible—especially when the underlying physics is unknown. Our work instead prioritizes interpretability through symbolic regression, which offers a more direct and generally applicable path to analytical understanding. In terms of inverse design, various neural network architectures have been proposed to engineer photonic and topological properties. For photonic crystals, generative adversarial networks (GANs) have been successfully employed to optimize bandgaps Christensen et al. 2020 , while other photonics applications have included neural network-driven design of metasurface-enabled invisibility cloaks Zhen et al. 2021 . In the topological domain, inverse design has been used to tailor edge state dispersions, for instance in the Aubry–André model Pilozzi et al. 2018 . A central challenge in inverse design is the many-to-one problem Deng et al. 2024 , where multiple input structures yield the same output properties.

Read the complete text