Réunion

Présentations invitées

Philipp del Hougne, CNRS, Université de Rennes 1, Physics-based digital twins for reconfigurable wave systems

Optimizating a reconfigurable wave system (e.g., reconfigurable intelligent surface, dynamic metasurface antenna, physical neural network) for a desired functionality is generally challenging because the effects of the tunable degrees of freedom are intertwined. In fact, the stronger the coupling between the tunable entities is, the more control reconfigurable wave systems offer over their transfer function. In this talk, I will discuss our recent efforts to formulate and calibrate physics-based models of reconfigurable wave systems. I will present closed-form and gradient-descent-based methods to calibrate the parameters of such models so that they can serve as digital twins for a priori unknown experimental reconfigurable wave systems, including a calibration method compatible with intensity-only detection. I will carefully explore the role of ambiguities in these physics-based digital twins, clarifying in what scenarios ambiguities need to be lifted and how that can be achieved. By comparison with physics-agnostic digital twins, I will highlight the favorable inductive bias of physics-based models. Once calibrated, these physics-based digital twins enable efficient in-software optimization of the reconfigurable wave system’s configuration for a desired functionality without further access to the experimental system. I will share insights into efficient optimization algorithms, notably for systems with few-bit programmable degrees of freedom.

Fanny Lehmann, ETH AI Center, Zurich, Des opérateurs de neurones aux modèles de fondation : Scientific Machine Learning pour les phénomènes de propagation en sciences de la Terre

Le Scientific Machine Learning a émergé comme un domaine à la frontière entre apprentissage automatique et simulation numérique pour construire des méta-modèles innovants. Inspirés par des principes physiques mais également capables de traiter une grande quantité de données d’entraînement, ces méta-modèles sont particulièrement pertinents pour résoudre des problèmes complexes de propagation d’ondes.

Dans cet exposé, je présenterai les opérateurs de neurones de Fourier appliqués à la propagation des ondes sismiques en trois dimensions. La rapidité du méta-modèle rend possible des approches bayésiennes inatteignables avec des solveurs numériques et j’illustrerai une application concrète pour la quantification d’incertitudes sur le séisme du Teil (Ardèche, 2019).

Dans les domaines où de grandes quantités de données sont disponibles, les méta-modèles sont étendus aux modèles de fondation pour capturer les propriétés intrinsèques du système physique. Je montrerai les bénéfices de ces modèles pour transférer l'information d'un problème physique à l'autre et ainsi diminuer les coûts de l'entraînement.

Raphaël Pestourie, Georgia Tech, USA ML-enhancement of simulation and optimization in electromagnetism,

Full-wave simulations of large-scale electromagnetic devices—spanning thousands of wavelengths while featuring sub-wavelength geometrical details—pose significant computational challenges. These simulations are critical in the design and optimization of metamaterials, where resolving fine-scale features is essential to predict macroscopic behavior accurately. In this talk, we present a specific use case where surrogate models are used to accelerate full-wave simulations of optical metasurfaces, thereby enabling iterative optimization loops that would otherwise be prohibitively expensive. We present scientific machine learning approaches for training these surrogate models to increase both the speed of evaluation and the data efficiency. Beyond surrogate modeling, we also demonstrate how machine learning can enhance optimization tasks independently of the underlying solver. In particular, we highlight methods for learning representations of the design space and guiding the search toward optimal configurations, thus improving the overall efficiency of the design process.

Pieter Wiecha, LAAS-CNRS, Toulouse, Surrogate models for nano-optics and photonics

Proxy models are becoming increasingly important for data-driven modelling in nano-optics and photonics. Dependent on the complexity of a system, simulations of nano-photonic devices can be very expensive. Data-based surrogate models are therefore very interesting for acceleration of slow simulations in time-critical applications. In particular, the rapid solution of ill-posed inverse problems can be tackled with the help of proxy models. Another strength of proxy models is the possibility of gradient calculations via automatic differentiation. In this context, I will discuss the potential of proxy models for optimization problems as well as the use of physics informed deep learning methods for improving nano-photonics proxy models.

Julien Mairal, INRIA Grenoble, Physical Models and Machine Learning for Scientific Imaging

Abstract: Deep learning has transformed image processing, often surpassing traditional methods that rely on precise modeling of optical image formation. In this presentation, we explore the synergy between model-based and learning-based approaches, highlighting the potential of hybrid methods for scientific imaging—an area where interpretability, robustness to real-world degradations, and the scarcity of ground truth data pose major challenges. We will showcase several applications, from high-contrast direct imaging for exoplanet detection to molecular-scale microscopy, illustrating how these hybrid strategies can push the boundaries of what is achievable in complex imaging scenarios.

Bio: Julien Mairal is a research director at Inria Grenoble, where he leads the Thoth research team. He joined Inria in 2012 following a postdoctoral position in the statistics department at UC Berkeley. He holds a Ph.D. from École Normale Supérieure de Cachan. His research spans machine learning, computer vision, mathematical optimization, and statistical image and signal processing. He has received several distinctions, including the Cor Baayen Award in 2013, the IEEE PAMI Young Researcher Award in 2017, and an ICML Test-of-Time Award in 2019. He was also awarded an ERC Starting Grant in 2016 and a Consolidator Grant in 2023.

Présentations orales

Differentiable ray-tracer and information gain used for coded hyperspectral systems co-design

Auteur·rice : Léo Paillet • LAAS-CNRS, IRAP • lpaillet@laas.fr

Co‑auteurs : Léo Paillet, Antoine Rouxel, Hervé Carfantan, Simon Lacroix, Antoine Monmayrant

Résumé :

We developed a differentiable ray-tracing optics simulator based on PyTorch and a framework previously designed for systems with axial symmetry. We extended this framework to non-symmetric systems and a wider variety of optical elements and systems. We believe that such a tool is particularly well suited to a co-design approach. We use this simulator to model a Coded Aperture Snapshot Spectral Imager (CASSI) for coded hyperspectral imagery. These systems employ a coded aperture (mask) to modulate the 3D hyperspectral scene, resulting in a 2D acquisition. We then adapted existing state-of-the-art unrolling reconstruction algorithms to solve the inverse problem of recovering scenes from coded acquisitions. We demonstrated that CASSI systems using the same mask but inducing different distortions can reconstruct scenes with similar performance, provided that the distortions are correctly modeled and accounted for. Therefore, we can say that the encoded information is constant regardless of distortions but depends on the mask. Thanks to the differentiability of our simulator, all its parameters—especially the mask—can be optimized with respect to a downstream loss function, leading to co-design. As reconstruction times are prohibitive in our high-dimensional use case, we attempt to bypass reconstruction by optimizing the mask to maximize the information gain that quantifies the information of the scene encoded in the acquisition. We hypothesize that an increase in information gain will lead to better processing performance, with a significant reduction in optimization time.

Self-Organized Freeform Waveguiding

Fadhila Chehami • Laboratoire XLIM, Université de Limoges • fadhila.chehami@unilim.fr

Co‑auteurs : Cyril Decroze, David R. Smith, Thomas Fromentèze

Résumé :

Nature offers remarkable examples of complex photonic architectures such as those responsible for the iridescent colors of butterfly wings that emerge spontaneously during growth, well before any centralized control takes place. Arising from local rules, these structures exhibit advanced optical functionalities, such as photonic band gaps, without relying on in-situ optimization or top-down design. Inspired by biological morphogenesis, we introduce an optimization-free approach for the automated generation of self-organized freeform waveguides that adapt to complex propagation paths. Our method relies on local reaction-diffusion dynamics to produce robust, spatially distributed structures. In contrast to conventional waveguides based on periodic media, which impose strong geometric constraints and require extensive fine-tuning, the proposed structures support nontrivial geometries while maintaining photonic band gap behavior. We experimentally demonstrate that these self-organized waveguides achieve superior transmission efficiency along complex paths. This optimization-free strategy enables the automated design of advanced electromagnetic components with intrinsic adaptability and resilience.

Single-shot quantitative polarimetric wavefront imaging

Auteur·rice : Grégoire Lelu • SPPIN – CNRS UMR 8003 • gregoire.lelu@u-paris.fr

Co‑auteurs : Baptiste Blochet, Grégoire Lelu, Miguel A. Alonso, Marc Guillon

Résumé :

Phase and polarimetric imaging are separately used in many application fields in microscopy. Polarized-resolved imaging can provide information about structural anisotropy at the molecular scale and is typically used for gems characterization in geology, healthy tissues identification in cancerology or fluorophore orientation determination in super-resolution microscopy. Phase imaging gives access to the height profile of reflecting surface samples and the optical path difference through samples. In cell biology, the optical path difference has been shown to be proportional to the dry mass and thus allows monitoring the metabolism of cells.

Here, we introduce a polarimetric wavefront imager that provides both the phase and the full-Stokes polarimetric images in a single acquisition of a multiplexed image. The device, composed of a birefringent Hartmann mask placed in the close vicinity of a camera, is first calibrated by recording images obtained when illuminating with beams of known polarization state. Then a single intensity image of a sample is recorded and demultiplexed thanks to a numerical inversion step to retrieve the full-Stokes polarization information and the optical path difference.

Presentation based on the work in the paper: Baptiste Blochet, Grégoire Lelu, Miguel A. Alonso, and Marc Guillon, “Quantitative polarimetric wavefront imaging,” Optica 12, 907–913 (2025).

Design and Morphogenetic Synthesis of Metasurface Antennas

Auteur·rice : Chidinma Nnekwu Uche • University of Limoges, CNRS, XLIM, UMR 7252, France • chidinma_nnekwu.uche@unilim.fr

Co‑auteurs : Nasibeh Parsaei, Cyril Decroze, Thomas Fromentèze

Résumé :

This contribution explores a morphogenetic approach for the design of self-organized metasurface antennas, offering high gain and wide bandwidth for applications such as satellite communications. Drawing inspiration from Alan Turing's foundational work on morphogenesis, we employ a reaction-diffusion model to synthesize and structure meta-atoms without relying on gradient-based optimization. This method enables the self-organization of anisotropic subwavelength elements, allowing for precise electromagnetic control at the macroscopic scale. Furthermore, the integration of controlled disorder into the meta-atom arrangement enhances the antenna’s homogenization and radiation performance. This work offers a promising route to simplified, scalable, and high performance metasurface antenna design.

Étalonnage en ligne de capteurs pour l'estimation collaborative de paramètres de mesure

Auteur·rice : Imane Meddour • CEA – List, Université Paris‑Saclay • imane.meddour@cea.fr

Co‑auteurs : Imane Meddour; Andréa Macario Barros; Cédric Gouy‑Pailler

Résumé :

L'étalonnage joue un rôle fondamental dans la qualité des mesures fournies par les capteurs. Cependant, dans des environnements réels et non contrôlés, les méthodes géométriques classiques montrent rapidement leurs limites face aux dérives, incertitudes et conditions variables. Dans ce travail, nous proposons une méthode d’étalonnage en ligne, basée sur l’apprentissage profond, capable d’estimer et de corriger les paramètres de mesure sans calibration préalable en laboratoire. Cette approche exploite la collaboration entre capteurs hétérogènes, combinant plusieurs modalités pour une acquisition multi‑capteurs robuste.

Elle vise à renforcer la précision et l’adaptabilité du système dans des conditions réelles, tout en s’intégrant efficacement dans des architectures embarquées à ressources limitées. Cette solution facilite le déploiement à grande échelle de systèmes intelligents et distribués.

An Equivariant Self-Supervised VAE for Uncertainty Quantification in Bayesian Imaging Problems

Auteur·rice : Bernardin Tamo Amougou • MAP5, Université Paris Cité • bernardin.tamo-amougou@etu.u-paris.fr

Co‑auteurs : Andrés Almansa; Marcelo Pereyra

Résumé :

Imaging problems are often ill‑posed, leading to significant uncertainty about their solutions. Accurately quantifying this uncertainty is critical for rigorous interpretation of experimental results and for reliably using reconstructed images as quantitative evidence. Unfortunately, existing Bayesian imaging methods cannot quantify uncertainty in a manner robust to experiment replications.

We introduce a new uncertainty‑quantification methodology based on an equivariant formulation of variational auto‑encoders that leverages the symmetries and invariance properties commonly encountered in imaging. The approach is fully self‑supervised and can be trained from observed data alone, enabling accurate data‑driven uncertainty quantification even when no ground‑truth data are available. Numerical experiments demonstrate the advantages of the proposed method over alternatives.

Apprentissage pour l’identification et la correction adaptative de la PSF en microscopie

Auteur·rice : Florian Sarron • IRIT / CBI, Université de Toulouse • florian.sarron@irit.fr

Co‑auteurs : Florian Sarron; Minh Hai Nguyen; Paul Escande; Pierre Weiss

Résumé :

L'identification précise de la fonction d'étalement du point (PSF) est essentielle pour améliorer la qualité des reconstructions en microscopie à fluorescence. Nous présentons ici une approche reposant sur l’apprentissage statistique pour estimer automatiquement la PSF à partir des images acquises, sans recourir à des billes fluorescentes ou à des procédures d’étalonnage spécifiques. Cette estimation automatique permet non seulement d'améliorer la qualité de la reconstruction numérique, mais s’inscrit dans une perspective plus large de co‑conception, où l’identification de la PSF pourrait, à terme, piloter dynamiquement des éléments adaptatifs du système (miroir déformable, position d’éléments optiques), afin de corriger en temps réel les dégradations instrumentales.

Cette contribution mettra l’accent sur la méthodologie d’identification, les résultats obtenus et la quantification des performances sur des données de simulations et des données réelles où la PSF induite est connue.

Inverse problem approaches for the reconstruction of mouse brain tissues with multi-beam

scanning transmission electron microscopy, Dylan Brault1, Maria Kormacheva1, Arent Kievits1, Tatiana Latychevskaia1,2, Adrian Wanner1

1PSI Center for Life Sciences, 5232 Villigen PSI, Switzerland

2Department of Physics, University of Zurich, Winterthurerstrasse 190, 8057 Zurich, Switzerland

Existing volume electron microscopy (vEM) imaging methodologies [1,2] depend on the laborious and error-prone process of sectioning and accumulating thousands of ultrathin serial sections (30–40 nm). Artifacts arising from ultrathin serial sectioning, such as section loss, cracks, and folds, account for over 70% of errors in contemporary automated neuron segmentation algorithms. To address these artifacts, a novel vEM method was developed, involving the preparation of semi-thin sections, approximately 250 nm thick, followed by iterative thinning of each section using a broad ion beam (BIB). Following each milling iteration, a projection image of the remaining semi-thin section is acquired using multi-beam scanning transmission electron microscopy (mSTEM) [3]. With this technique, the microscope does not directly record the sample's 3D structure because the projection image of the lower layers superimposes on the images of the top layers of the volume. These projection artifacts consequently introduce errors into the segmentation of neurons in the imaged 3D volume. It is therefore necessary to use advanced reconstruction algorithms to recover a highresolution 3D volume that untangles the contribution of each layer in the signal.

High-resolution reconstruction that aligns with physics is achieved within the inverse problems approach framework by accurately modeling image formation and leveraging prior sample knowledge with regularization techniques. This study presents an application of this inverse problem framework for the reconstruction of mSTEM data of mouse brain tissues acquired with our mSTEM microscope.

[1] Eberle, A. L. et al. High-resolution, high-throughput imaging with a multibeam scanning electron microscope. J. Microsc. 259, 114–120 (2015)

[2] Maniates-Selvin, J. T. et al. Reconstruction of motor control circuits in adult Drosophila using automated transmission electron microscopy. bioRxiv 2020.01.10.902478 (2020)

[3] Ren, Y. & Kruit, P. Transmission electron imaging in the Delft multibeam scanning electron microscope 1. J. Vac. Sci. Technol. B, Nanotechnol. Microelectron.: Mater., Process., Meas., Phenom.

34, 06KF02 (2016).

Session poster

Auteur·rice : Grégoire Lelu • SPPIN – CNRS UMR 8003

Co‑auteurs : Baptiste Blochet, Grégoire Lelu, Miguel A. Alonso, Marc Guillon

Résumé : Phase and polarimetric imaging are separately used in many application fields in microscopy. Polarized-resolved imaging can provide information about structural anisotropy at the molecular scale and is typically used for gems characterization in geology, healthy tissues identification in cancerology or fluorophore orientation determination in super-resolution microscopy. Phase imaging gives access to the height profile of reflecting surface samples and the optical path difference through samples. In cell biology, the optical path difference has been shown to be proportional to the dry mass and thus allows monitoring the metabolism of cells. Here, we introduce a polarimetric wavefront imager that provides both the phase and the full-Stokes polarimetric images in a single acquisition of a multiplexed image. The device, composed of a birefringent Hartmann mask placed in the close vicinity of a camera, is first calibrated by recording images obtained when illuminating with beams of known polarization state. Then a single intensity image of a sample is recorded and demultiplexed thanks to a numerical inversion step to retrieve the full-Stokes polarization information and the optical path difference. Presentation based on the work in the paper: Baptiste Blochet, Grégoire Lelu, Miguel A. Alonso, and Marc Guillon, “Quantitative polarimetric wavefront imaging,” Optica 12, 907– 913 (2025).

Wen Ruan, Laboratoire Charles Fabry

Thomas Lebrat, Photonics Bretagne

Oihan Joyot, IRIT, Toulouse