Choose timezone
Your profile timezone:
Séminaire de Statistique et Optimisation
Séminaire des doctorant.es SO
by , , ,
→
Europe/Paris
Salle K. Johnson (1R3, 1er étage)
Salle K. Johnson
1R3, 1er étage
Description
Robert Koprinkov
An Analysis of Gaussian Process Regression for Multiphysics Problems
Abstract: Multiphysics problems are ubiquitous in engineering. They are usually resolved through fixed point iteration procedures involving several numerical solvers, each modelling a different domain of physics. Multi-query problems consider parametrized engineering problem, and involve the solution of the problem for many different parameter values. Previous calculations can be used to build a surrogate model of the numerical solvers, for example by interpolating previous solutions. These surrogate models can then be used to reduce the computational cost of solving the multi-query problem. This poster analyzes one such surrogate modelling procedure for multi-query multiphysics problems, which is based on the interpolation of the Proper Orthogonal Decomposition (POD) basis vector components using Gaussian Process Interpolation.
Jules Ripoll
Adapting flow models for real image editing
Abstract: Text to image generative models have made impressive strides in recent years. Their text conditioning allows for flexible and expressive image generation, leaving end users with countless possibilities. Many contributions have been made in order to adapt these models to image editing, with convincing results. While progress has been made, such methods can still collapse when editing real images. In this work, we propose single image fine-tuning to adapt the model to new images, thus making our image a more likely sample from our model under a given prompt. We highlight a relationship between computation and robustness, where one can allocate more time for better results. We demonstrate our method’s capabilities with extensive qualitative results as well as state-of-the-art identity preservation score on human faces.
Nathan Gorse
A Statistical Approach to Modeling Extreme Coastal Flooding Under Non-Standard Conditions
Abstract: We are interested in the influence of time-varying meteoceanic conditions on coastal flooding. We specifically work on the surge and focus on extreme observations in the functional domain. Yet, our observations do not meet standard hypotheses as they are dependent and have short-tailed distributions. We propose a three-stage methodology. Firstly, we introduce an autoregressive model in order to remove the temporal dependence between cycles and show that its residuals satisfy the required assumptions.
Then, we show how to combine existing techniques related to Pareto processes to simulate extreme residuals. Finally, applying the reverse transformations, we simulate new extreme conditions over tidal cycles. This step depends on an initial time series that can be chosen to tune the desired level of extremes. The developments are applied to the site of Gâvres in the south of the French Brittany.
Adrián Padilla-Segarra
Reconstruction of Fluid Flows using Physics-Informed, Boundary-Constrained Gaussian Processes
Abstract: Recently, there is a growing interest in adapting data–driven and data assimilation methods to fluid mechanics frameworks. In this contribution, we address the question of how to precondition a Gaussian process (GP) prior with physics-information pertaining to incompressible fluid flows past an aerodynamic profile. In particular, we define a prior distribution such that its samples of the velocity field satisfy a divergence-free condition and a slip condition on the profile boundary in a continuous manner. We perform GP regression with observations issued from high fidelity simulations of incompressible flows past a cylinder profile and past a NACA airfoil.
