Séminaire Calcul Scientifique et Modélisation
Responsables : Christele Etchegaray et Martin Parisot
We discuss a new swarm-based gradient descent (SBGD) method for non-convex optimization. The swarm consists of agents, each is identified with position $x$ and mass $m$. There are three key aspects to the SBGD dynamics: (i) persistent transition of mass from agents at high to lower ground; (ii) a random marching direction, aligned with the steepest gradient descent; and (iii) a time stepping protocol which decreases with $m$.
The interplay between positions and masses leads to dynamic distinction between `heavier leaders’ near local minima, and `lighter explorers’ which explore for improved position with large(r) time steps. Convergence analysis and numerical simulations demonstrate the effectiveness of SBGD method as a global optimizer.
.
Simuler numériquement de manière précise l'évolution des interfaces séparant différents milieux est un enjeu crucial dans de nombreuses applications (multi-fluides, fluide-structure, etc). La méthode MOF (moment-of-fluid), extension de la méthode VOF (volume-of-fluid), utilise une reconstruction affine des interfaces par cellule basée sur les fractions volumiques et les centroïdes de chaque phase. Cette reconstruction d'interface est solution d'un problème de minimisation sous contrainte de volume. Ce problème est résolu dans la littérature par des calculs géométriques sur des polyèdres qui ont un coût important en 3D. On propose dans cet exposé une nouvelle approche du calcul de la fonction objectif et de ses dérivées de manière complètement analytique dans le cas de cellules hexaédriques rectangulaires et tétraédriques en 3D. Les résultats numériques montrent un gain important en temps de calcul.
Conventional methods, like the continuous Finite Element SUPG, face limitations in maintaining these structures due to the stabilization techniques employed, which do not effectively vanish when the discrete divergence is zero.
What we propose is to use the Global Flux procedure, which has proven to be successful in preserving 1-dimensional equilibria [1,2], to define some auxiliary variables guiding a suitable discretization of both the divergence and stabilization operators [3]. This approach enables the natural preservation of divergence-free solutions and more intricate equilibria involving various sources. Moreover, this strategy facilitates the identification of discrete equilibria of the scheme that verify boundary or initial conditions. We use the Deferred Correction time discretization, obtaining explicit arbitrarily high order methods.
Numerous numerical tests validate the accuracy of our proposed scheme compared to classical approaches. Our method not only excels in preserving (discretely) divergence-free solutions and their perturbations but also maintains the original order of accuracy on smooth solutions.
[1] Y. Cheng, A. Chertock, M. Herty, A. Kurganov and T. Wu. A new approach for designing moving-water equilibria preserving schemes for the shallow water equations. J. Sci. Comput. 80(1): 538–554, 2019.
[2] M. Ciallella, D. Torlo and M. Ricchiuto. Arbitrary high order WENO finite volume scheme with flux globalization for moving equilibria preservation. Journal of Scientific Computing, 96(2):53, 2023.
[3] W. Barsukow, M. Ricchiuto and D. Torlo. Structure preserving methods via Global Flux quadrature: divergence-free preservation with continuous Finite Element. In preparation, 2024.
Institut de Mathématiques de Bordeaux UMR 5251
Université de Bordeaux
351, cours de la Libération - F 33 405 TALENCE
Tél: +33 (0)5 40 00 60 70
institut
math.u-bordeaux.fr
