HAL-Polytechnique
Not a member yet
51406 research outputs found
Sort by
Relationship between molecular structures and thermogravimetric properties of gallium–amidinate based compounds
No full textInternational audienceThe ability to predict the thermal properties of molecular compounds is essential for their successful integration into vapor-phase processes such as atomic layer deposition (ALD) and chemical vapor deposition (CVD), as well as in catalysis and materials synthesis. In this work, we present a systematic study of 24 gallium amidinate complexes, designed to explore the relationship between molecular structure and thermal behavior. The series encompasses a range of structural variations: different ligand substituents, molecular symmetry, and co-ligands. Structural characterization, including in some cases single-crystal X-ray diffraction, was combined with detailed thermal analysis using thermogravimetric analysis (TGA) and differential scanning calorimetry (DSC) under both atmospheric and reduced pressure conditions. The results reveal clear correlations between thermal properties and ligand architecture, with features such as alkyl chain type, methyl group presence, and overall symmetry playing key roles in determining volatility and stability. Importantly, both symmetric and dissymmetric complexes were found to possess the desired thermal characteristics for vapor-phase deposition processes. Beyond offering valuable design principles for gallium precursors, the dataset generated herein provides a foundation for improving predictive models—empirical and AI-driven alike—towards the rational development of next-generation functional molecular compounds
Fully explicit numerical scheme for linearized wave propagation in nearly-incompressible soft hyperelastic solids
No full textInternational audienceThe numerical approximation of wave propagation problems in nearly or pure incompressible solids faces several challenges such as locking and stability constraints. In this work we propose a stabilized Leapfrog scheme based on the use of Chebyshev polynomials to relax the stability condition, which is strongly limited by the enforcement of incompressibility. The scheme is fully explicit, second order accurate and energy-preserving. For the space discretization we use a mixed formulation with high-order spectral elements and mass-lumping. A strategy is proposed for an efficient and accurate computation of the pressure contribution with a new definition of the discrete Grad-div operator. Finally, we consider linear wave propagation problems in nearly-incompressible hyperelastic solids subject to static preload
Can parameterizations reproduce the gravity waves momentum fluxes and drag simulated by a global high-resolution model?
No full textInternational audienceWe compare the gravity wave (GW) parameterizations used in the IPSLCM6 climate model with the GWs resolved in the ICON global model with 5 km horizontal resolution. The parameterizations are run offline using ICON fields coarse-grained to a 100 km grid and compared to the GWs with smaller scales that are resolved in ICON. Overall, the drags are comparable, the momentum fluxes align well, and each GW parameterization (fronts, convection, and mountains) plays a role at geographical locations consistent with ICON. Among the differences, we find that in ICON, GWs are substantially attenuated aloft the subtropical jets; this is underestimated by the parameterizations. It could be corrected by tuning the characteristic phase speeds or the breaking criteria in the parameterizations. It also seems that ICON underestimates frontal waves in the mid-latitudes, that the parameterizations underestimate the convective waves in the tropics, and that the mountain waves are more alike
Total synthesis of photoswitchable latrunculin B enables reversible control of actin polymerization and cell migration
No full textInternational audienceThe actin cytoskeleton plays a central role in regulating essential cellular properties such as cell shape, contraction, division, or migration. The filament growth, controlled by subtle regulation mechanisms, can also be inhibited by small molecules preventing actin polymerization, like latrunculin B. This compound has been widely used in cell biology to inhibit cytokinesis or cell migration. To further these applications and provide a mean for the spatiotemporal control of cell migration, we pre-pared photoswitchable latrunculin B through total synthesis, allowing the incorporation of a light-responsive azobenzene moiety. One of these modified latrunculins displayed excellent photophysical properties, with good photostationary states (ca. 90/10 in the two reversible states), long half-lives and excellent fatigue resistance during photoswitching, performed at 370 nm and 440 nm. The Z-photoisomer induced a significant inhibition of the growth of single actin filaments in vitro, com-paratively to the E-isomer—a reversible behavior that was rationalized through molecular docking. Most importantly, this inhibition was photoinduced in migrating cells, stopping actin-dependent membrane dynamics in a reversible manner, and was induced locally in a single cell
Total Synthesis of Daphniphyllum Alkaloids: A unified fragment-based approach
No full textInternational audienc
Avancées dans l’apprentissage de la représentation graphique et leurs applications en biologie computationnelle
No full textGraphs are a natural and flexible framework for modeling relationships and interactions across entities in various domains, including social networks, chemistry, and biology. This has fueled significant interest in graph representation learning, which focuses on learning informative representations and making predictions for graph-structured data. However, the inherently non-Euclidean nature of graph data poses substantial challenges for traditional machine learning models. Graph Neural Networks (GNNs) address these challenges by leveraging the message-passing scheme to learn node and graph representations. Despite their success, existing GNNs face significant difficulties in fully exploiting the information in graph data. In this thesis, we aim to overcome several limitations of GNNs, including their inability to capture synergistic interactions among nodes, difficulties in leveraging unlabeled graph data effectively, and other issues, such as the lack of interpretability. Furthermore, we develop new graph representation learning techniques to solve critical problems in computational biology, including cancer gene prediction and protein function prediction.In the first part of the thesis, we show that GNNs compute messages from each neighboring node independently, which limits their ability to capture complex, synergistic interactions among the nodes.To overcome this limitation, we propose a new GNN layer, which uses a self-attention mechanism and a recurrent neural network as an aggregator, thus considering the dependencies between the nodes.Secondly, we improve the performance of Graph Attention Networks (GATs) by introducing a novel loss function designed to guide the model in learning higher attention scores between nodes belonging to the same class.While the above methods have demonstrated strong performance in supervised settings, their reliance on labeled data limits their effectiveness in domains where annotations are scarce or expensive to obtain. One such domain is 3D protein graphs, where vast amounts of structural data exist, but only a small fraction is labeled with functional annotations.To this end, we propose a new geometric self-supervised pretraining framework for 3D protein graphs that predicts the subgraph distances relative to the overall geometric centroid of the protein graph.Moreover, we analyze the structural properties encoded by GNNs and reveal a strong connection between node embeddings and the number of random walks originating from nodes.Our findings connect to the over-squashing phenomenon, in which distant node information is compressed and lost, as well as to the stability of GNNs.In the second half of the thesis, we focus on applications of graph representation learning to computational biology.Cancer gene prediction is a critical task in computational biology, as the identification of cancer-associated genes can reveal insights into the underlying mechanisms of cancer and help in identifying new therapeutic targets.To this end, we propose a new framework based on GNNs to identify cancer genes, which leverages multiomics data and multiple gene and protein interaction networks.Our approach achieves state-of-the-art performance in predicting cancer driver genes while being interpretable.Furthermore, we tackle the problem of protein function prediction with a novel approach that combines GNNs and Large Language Models. Our framework integrates protein sequence and structure information to predict protein function in a text format, in contrast with previous classification models.Les graphes constituent un cadre naturel et flexible pour modéliser les relations et interactions entre entités dans divers domaines, tels que les réseaux sociaux, la chimie et la biologie. Cela a suscité un vif intérêt pour l’apprentissage des représentations de graphes, qui vise à extraire des représentations informatives et à réaliser des prédictions sur des données structurées en graphes. Cependant, la nature intrinsèquement non euclidienne des graphes pose d’importants défis aux modèles d’apprentissage automatique classiques. Les réseaux de neurones pour graphes (GNN) relèvent ces défis en s’appuyant sur un mécanisme de passage de messages pour apprendre des représentations de nœuds et de graphes. Malgré leur succès, les GNN existants peinent à exploiter pleinement l’information contenue dans les graphes. Dans cette thèse, nous cherchons à surmonter plusieurs de leurs limites, telles que l’incapacité à saisir les interactions synergiques entre nœuds, les difficultés à exploiter efficacement les données non étiquetées, ainsi que d’autres problèmes comme le manque d’interprétabilité. Par ailleurs, nous développons de nouvelles techniques d’apprentissage des représentations de graphes pour résoudre des problématiques critiques en biologie computationnelle, notamment la prédiction de gènes liés au cancer et la prédiction de la fonction des protéines.Dans la première partie de la thèse, nous démontrons que les GNN calculent les messages de chaque nœud voisin de manière indépendante, limitant ainsi leur capacité à capturer des interactions complexes et synergiques entre nœuds. Pour y remédier, nous proposons une nouvelle couche de GNN qui utilise un mécanisme d’auto-attention et un réseau de neurones récurrent en tant qu’agrégateur, afin de prendre en compte les dépendances entre nœuds. Nous améliorons ensuite les performances des Graph Attention Networks (GAT) en introduisant une fonction de perte novatrice destinée à encourager le modèle à attribuer des scores d’attention plus élevés aux nœuds de la même classe.Bien que ces méthodes aient montré d’excellents résultats en contexte supervisé, leur dépendance aux données étiquetées limite leur efficacité dans des domaines où les annotations sont rares ou coûteuses. Un tel domaine est celui des graphes 3D de protéines, où de vastes quantités de données structurelles existent, alors qu’une seule fraction est annotée fonctionnellement. Pour ce faire, nous proposons un nouveau cadre de préentraînement auto-supervisé géométrique pour les graphes 3D de protéines, capable de prédire les distances entre sous-graphes par rapport au centroïde géométrique global du graphe protéique. De plus, nous analysons les propriétés structurelles encodées par les GNN et révélons un lien étroit entre les embeddings de nœuds et le nombre de marches aléatoires initiées à partir de ces nœuds, se connectant ainsi au phénomène de sur-compression où l’information des nœuds éloignés est comprimée et perdue, ainsi qu’à la stabilité des GNN.Dans la seconde moitié de la thèse, nous nous concentrons sur les applications de l’apprentissage des représentations de graphes en biologie computationnelle. La prédiction des gènes liés au cancer est une tâche cruciale, car l’identification de ces gènes peut éclairer les mécanismes sous-jacents à la maladie et aider à identifier de nouvelles cibles thérapeutiques. À cet égard, nous proposons un nouveau cadre basé sur les GNN pour identifier les gènes du cancer, en exploitant des données multi-omiques et plusieurs réseaux d’interactions entre gènes et protéines. Notre approche atteint des performances de pointe tout en demeurant interprétable.Par ailleurs, nous abordons la prédiction de la fonction des protéines via une approche novatrice combinant GNN et grands modèles de langage, en intégrant des informations sur la séquence et la structure des protéines pour prédire leur fonction sous forme de texte, contrairement aux modèles de classification antérieurs
Noise2Noise Image Reconstruction of Lifetime Maps in Halide Perovskite Thin Films
No full textInternational audienceWe present an unsupervised deep-learning approach for lifetime map reconstruction from noisy time-resolved fluorescence imaging (TR-FLIM) datasets. In the context of semiconductor and photovoltaic device characterisation, this method is critical for accurately predicting solar cell performance and detecting early signs of degradation. More precisely, we consider an unsupervised Noise2Noise (N2N) training framework combined with physics-driven modelling for the quantitative reconstruction of lifetime maps. The proposed approach incorporates a log-linear fit in the N2N loss function and parameterises the unknown maps as outputs of a shallow neural network with a multibranch architecture. By learning from multiple noisy acquisitions of the same scene, our method effectively allows an accurate estimation with shorter acquisition protocols, which translates into a lower risk of damage for the sample under consideration. Tests on simulated data and comparisons with available model-based approaches show that the proposed approach improves robustness w.r.t. noise levels with limited tuning of the regularisation/algorithmic parameters
Massive parallelization of projection-based depths
No full textThis article introduces a novel methodology for the massive parallelization of projection-based depths, addressing the computational challenges of data depth in high-dimensional spaces. We propose an algorithmic framework based on Refined Random Search (RRS) and demonstrate significant speedup (up to 7,000 times faster) on GPUs. Empirical results on synthetic data show improved precision and reduced runtime, making the method suitable for large-scale applications. The RRS algorithm (and other depth functions) are available in the Python-library data-depth (https://data-depth.github.io/) with ready-to-use tools to implement and to build upon this work
CYCLOPs: a Unified Framework for Surface Flux-Driven Cyclones Outside the Tropics
No full textInternational audienceCyclonic storms resembling tropical cyclones are sometimes observed well outside the tropics. These include medicanes, polar lows, subtropical cyclones, Kona storms, and possibly some cases of Australian East Coast cyclones. Their structural similarity to tropical cyclones lies in their tight, nearly axisymmetric inner cores, eyes, and spiral bands. Previous studies of these phenomena suggest that they are partly and sometimes wholly driven by surface enthalpy fluxes, as with tropical cyclones. Here we show, through a series of case studies, that many of these non-tropical cyclones have morphologies and structures that resemble each other and also closely match those of tropical transitioning cyclones, with the important distinction that the potential intensity that supports them is not present in the pre-storm environment but rather is locally generated in the course of their development. We therefore propose to call these storms CYClones from Locally Originating Potential intensity (CYCLOPs). Like their tropical cousins, the rapid development and strong winds of cyclops pose a significant threat and forecast challenge for islands and coastal regions
A Universal Uniform Approximation Theorem for Neural Networks
No full textInternational audienceWe show the existence of a fixed recurrent network capable of approximating any computable function with arbitrary precision, provided that an encoding of the function is given in the initial input. While uniform approximation over a compact domain is a well-known property of neural networks, we go further by proving that our network ensures effective uniform approximation - simultaneously ensuring: - Uniform approximation in the sup-norm sense, guaranteeing precision across the compact domain {[0,1]^d}; - Uniformity in the sense of computability theory (also referred to as effectivity or universality), meaning the same network works for all computable functions. Our result is obtained constructively, using original arguments. Moreover, our construction bridges computation theory with neural network approximation, providing new insights into the fundamental connections between circuit complexity and function representation. Furthermore, this connection extends beyond computability to complexity theory. The obtained network is efficient: if a function is computable or approximable in polynomial time in the Turing machine model, then the network requires only a polynomial number of recurrences or iterations to achieve the same level of approximation, and conversely. Moreover, the recurrent network can be assumed to be very narrow, strengthening the link our results and existing models of very deep learning, where uniform approximation properties have already been established