Schedule June 2-3

École d'été - COPL
Summer School

Sophie Larochelle, Professor, Director, COPL Mathieu Juan, Professor, Université de Sherbrooke

École d'été - COPL
Summer School

Yuri Grinberg - Senior Research Officer, Data Science for Complex Systems, Digital Technologies Research Center, National Research Council Canada Abstract Machine learning and deep learning in particular have been under a spotlight for a number of years due to their ability to effectively identify complex patterns in data and leverage them in multitude of ways across domains and applications. In this seminar, we will look at different problems in integrated photonics and the way they leverage diverse machine learning problem formulations. Finding commonalities among redundant solutions in a large design space, speeding up the design optimization process, and infusing fabrication considerations into this process are among the use cases that will be discussed. Finally, we will wrap up with a few key points that should be considered in the early stages of research to create impactful and practical demonstrations of machine learning in the field. Biography Dr. Yuri Grinberg is a Senior Research Officer within the Digital Technologies Research Center. He obtained his PhD in Computer Science from McGill University in 2014 and was an NSERC Postdoctoral fellow in Ottawa Hospital Research Institute before joining NRC. His expertise is applied and theoretical machine learning and reinforcement learning. In the last several years he has lead the development of AI techniques for the design of photonic components, leveraging a variety of computational methods borrowed from optimization and machine learning fields. In particular, his work focuses on development of methods to improve efficiency, reduce the footprint, and improve fabrication tolerance of photonic components without requiring major human efforts. He has co-authored over 50 peer-reviewed publications.

École d'été - COPL
Summer School

Elena Redchenko, Postdoctoral Researcher, TU Wien Abstract Quantum solids made of noble gases (e.g. Ar, Ne) or molecules (e.g. p-H2, N2) provide a platform for studying complex quantum many-body systems in a hybrid architecture with superconducting circuits. The noble gas crystals offer a soft, inert, predominantly spin-0 host matrix for the atomic impurities, while the alkali atoms have addressable hyperfine transitions in the GHz regime. We deposit these doped quantum solids inside the cryostat directly atop superconducting circuit, however, the final properties of each crystal depends on multiple growing parameters. In my talk, I will present the quantum solids hybrid platform and demonstrate how we use machine learning to optimize the crystal deposition and to reach the strong coupling regime between the impurity ensemble and the superconducting resonator at mK temperatures for efficient coherent information exchange. At the end of my talk, I will discuss how this unique architecture can be used for exploring fundamental quantum effects and new technologies. Biography Elena is a leading postdoc of the hybrid quantum systems team in the group of Prof. Joerg Schmiedmayer at the Technical University of Vienna. She obtained her PhD from the Institute of Science and Technology Austria, where she studied the collective behaviour of superconducting qubit ensembles in the group of Prof. Johannes Fink.

École d'été - COPL
Summer School

Panel discussion to inspire upcoming graduates about the world of work in photonics and optics “skills, expectations and advice for a smooth transition”. Session moderator - Mathieu Juan, Professor, Université de Sherbrooke Angeles Camacho - Coractive & Optica Ambassador Devika Nair - Ansys Lumerical Adrien Longa - DistriQ

École d'été - COPL
Summer School

École d'été - COPL
Summer School

Marouchka-M. Brisebois, programs and communications manager, COPL Vincent Queneville-Guay, PhD Student, INRS-ÉMT

École d'été - COPL
Summer School

Rejeena Radhika Sebastian, PhD student, ÉTS Montréal Research Supervisor: Prof. Francois Blanchard Abstract: Although noise may appear random, it often follows specific statistical distribution functions. This project aims to uncover the underlying distribution functions of noise in terahertz (THz) images and establish mathematical correlations between them to enhance image denoising. THz imaging has demonstrated remarkable potential in fields such as material characterization, security screening, and biomedical imaging, but noise remains a significant barrier to achieving high-resolution and reliable images. This work introduces the first statistical treatment of noise in THz imaging using deep neural networks (DNNs). We employed Noise2Clean (N2C) and Noise2Noise (N2N) models to identify and model various noise distributions, including additive noises such as Gaussian, impulse, and Poisson, as well as multiplicative noises like Bernoulli, gamma, and exponential. The noise distributions were derived by iterating through multiple statistical functions, with experimental methods defining the boundaries of these functions. The resulting noise models were used for image denoising, achieving substantial improvements in image quality, with peak signal-to-noise ratio (PSNR) enhancements of 8 dB in the N2C model and 6 dB in the N2N model.

École d'été - COPL
Summer School
L'intelligence artificielle
Artificial Intelligence

Yingming Lai, PhD Student, INRS Program: Doctorat en sciences de l'énergie et des matériauxResearch Supervisor: Jinyang Liang Abstract: Deep learning techniques are pivotal in advancing computational ultrafast imaging systems, which combine cutting-edge hardware and software to capture transient phenomena on ultrashort timescales. While deep-learning algorithms enhance reconstruction performance, most of them are applied as replacements for analytical-modeling-based image reconstruction, addressing only one part of the imaging pipeline. To address this limitation, we develop an end-to-end convolutional neural network—termed deep high-dimensional adaptive network (D-HAN)—to provide comprehensive and multi-faceted supervision across the entire imaging system. Acting as a central “brain”, D-HAN optimizes hardware configurations and software processes, bridging gaps between system design and reconstruction. When applied to a compressed ultrafast photography system, D-HAN overcomes the limitations of conventional random coded apertures and idealized linear shearing models by optimizing mask patterns and sensing actual shearing operations. When integrated into a compressed ultrafast tomographic imaging system, D-HAN provides feedback on spatiotemporal projection behaviors, enabling precise alignment of data acquisition with computational models for accurate tomographic reconstructions. With the aid of D-HAN, both modalities achieve enhancements in imaging quality and reconstruction speed. We envision that D-HAN’s deep-learning-driven framework will benefit computational ultrafast imaging by unifying system design, image reconstruction, and performance evaluation.

École d'été - COPL
Summer School
L'intelligence artificielle
Artificial Intelligence

Corentin Soubeiran, PhD student, Université Laval Research Supervisor: Pierre Marquet Abstract: We developed a physic aware deep learning approach to enhance digital holographic microscopy (DHM) images by overcoming their resolution and noise limitations in large fields of view (FOV). Using polychromatic DHM as a reference, we trained a single-image super-resolution model to denoise and refine low-resolution DHM images. This hardware-free approach enables resolution of fine cellular structures like neuronal extensions across large FOVs at full sensor speed. Our method effectively bridges the microscopy trade-off between optical resolution and FOV in quantitative phase imaging.

École d'été - COPL
Summer School
L'intelligence artificielle
Artificial Intelligence

Marjolaine Malgéri, Master student, Polytechnique, Montréal Programme: Maîtrise recherche en Génie Biomédical, Polytechnique Montréal Directeur de recherche: Michel MeunierTitre original: Analyse d’images assistée par Intelligence Artificielle de biopsies de cancer marquées par des nanoparticules plasmoniques Résumé: En 2024, le cancer du sein a été le deuxième cancer le plus diagnostiqué au Canada, entraînant de grands efforts de recherche aboutissant à l’émergence de nouveaux traitements, comme le Trastuzumab Deruxtecan (T-DXd). Pour identifier efficacement les patients qui pourraient être éligibles à ces traitements, il est nécessaire de proposer des méthodes particulièrement spécifiques et sensibles pour quantifier précisément l’expression protéique des tumeurs. Le laboratoire du Professeur Meunier, spécialisé dans l’utilisation de nanoparticules (NPs) plasmoniques dans des contextes biologiques, propose d’utiliser des NPs d’or biofonctionnalisées pour cibler spécifiquement la protéine HER2 (protéine particulièrement importante dans la classification des cancers du sein). La technologie développée est appelée Imunoplasmonique (IP). Les propriétés de diffusion optique des NPs les rendent visibles comme des points brillants en microscopie optique, ce qui permet leur quantification. L’analyse d’image des échantillons marqués par IP nécessite une stratégie robuste de détection des NPs ainsi que des structures biologiques auxquelles elles sont attachées. Pour cela, des algorithmes U-Net ont été entrainés à réaliser des tâches de segmentation sémantique. Les résultats de ces segmentations sont ensuite utilisés pour développer des métriques, dans un premier temps pour vérifier la qualité du marquage IP, et dans un second temps pour évaluer l’expression protéique des lames histologiques marquées par IP. Finalement, l’accès aux données d’une biobanque permettront de relier les résultats des analyses IP avec la réponse au traitement de patientes traitées avec T-DXd, et ainsi de montrer l’efficacité de la méthode IP pour la quantification de protéines.

École d'été - COPL
Summer School
L'intelligence artificielle
Artificial Intelligence

Alexis Horik, Master student, Université LavalDirecteurs de recherche : Louis Archambault et Simon Thibault Titre d'origine: Reconstruction 3D par apprentissage profond pour la dosimétrie 3D par scintillation en temps-réel. Résumé: Objectif: Développer un nouveau type de dosimètre 3D en temps réel pour les contrôles qualité (CQ) des accélérateurs linéaires utilisés en radiothérapie externe. Méthodes: Un montage expérimental a été conçu en utilisant un cube scintillant de 1000 cm3, deux miroirs plans et une caméra CCD. Les miroirs étaient positionnés pour capturer trois vues orthogonales du volume dans une seule image. Puisque la scintillation est proportionnelle à la dose déposée, l’émission lumineuse autour du volume reflète la distribution de dose en 3D. Un modèle d’apprentissage profond a été construit en exploitant les avantages du mécanisme d’attention des transformeurs et le biais spatial induit par les réseaux neuronaux convolutifs (CNN). Le réseau de neurones peut donc prendre en entrée une image 2D et, après avoir appliqué le mécanisme d’attention, un encodage positionnel permet de remodeler en 3D la sortie du transformeur. Ensuite, en utilisant des couches de convolution transposées, le CNN 3D peut amener le volume à la résolution désirée. Pour l’entraînement du modèle, des techniques de régularisation telles que l’abandon (de l’anglais dropout) et la dégradation des pondérations (de l’anglais weight decay) ont été implémentées. Une recherche d’hyperparamètres utilisant une fonction de perte combinant plusieurs métriques telles que l’indice gamma, l’erreur quadratique moyenne (MSE) et l’indice de similarité structurelle (SSIM) a également été menée. Puisque seulement 250 données expérimentales ont été récoltées, 30 000 données synthétiques ont été générées avec Python à l’aide de techniques de tracé de rayon. Ainsi, un pré-entraînement et un ajustement fin du modèle ont été effectués en utilisant respectivement les données synthétiques et les données expérimentales.

École d'été - COPL
Summer School
L'intelligence artificielle
Artificial Intelligence

Hamed Rabbani, PhD student, Université LavalResearch Supervisor: Leslie A. Rusch Abstract: Physics-based models for designing erbium-doped fiber amplifiers (EDFAs) and erbium ytterbium-doped fiber amplifiers (EYDFAs) are computationally expensive and struggle to accurately predict behavior in the extended L-band. We train fast and accurate neural network (NN) models using experimental data for both prediction and optimization of single-stage and double-stage amplifiers. For single-stage amplifiers, independent NN models for gain and noise figure enable precise modeling of both core- and cladding-pumped configurations. Compared to physics-based model, the NN approach significantly reduces absolute error while being 400 times faster. Extending this framework to two-stage amplifiers, the model captures the complexities of booster-stage behavior, including non-uniform input signals and amplified spontaneous emission (ASE). A large experimental dataset ensures reliable training, resulting in an average absolute error below 0.27 dB for gain and 0.15 dB for noise figure. Beyond prediction, the NN model serves as a powerful optimization tool. We optimize a mid-stage filter in double-stage EDFA amplifier using particle swarm optimization (PSO) technique to balance gain flattening and noise figure degradation. Additionally, we explore effectiveness of a hybrid-pumped EYDFAs. We examine several pump configurations and find good trade-offs in performance and power when employing a hybrid pumping scheme that combines forward 915 nm cladding-pumping with backward 1480 nm core-pumping. We train a fast and accurate NN model using experimental data to examine the effect on performance (gain, noise figure and total electrical power consumption) as we

École d'été - COPL
Summer School
L'intelligence artificielle
Artificial Intelligence

Abhiroop Bhattacharya, PhD student, ETS, MontréalResearch Supervisor: Sylvain G. Cloutier Abstract : Defect engineering in crystals can help tailoring their structural and physical properties. Rapid estimation of the properties for a given lattice structure and defect configuration is an important part of the design process, yet it presents a considerable challenge due to the high computational cost of Density Functional Theory (DFT) calculations. As an alternative to costly DFT calculations, we explore self-supervised learning methods leveraging large amounts of unlabeled data to learn expressive and meaningful representations. To benchmark our model’s performances, we train and evaluate it using a publicly-available 2D dataset. We report accuracies on par with state-of-the-art supervised learning methods to avoid compute-heavy DFT simulations. We can then exploit the learned embedding as input for a linear or nearest neighbor model to predict multiple material properties, oQering versatility and scalability. Most importantly, our framework is especially well-suited for use with specialized curated datasets containing limited data points. As such, this work provides a valuable new tool for future materials research and defect engineering.

École d'été - COPL
Summer School
L'intelligence artificielle
Artificial Intelligence

École d'été - COPL
Summer School