Deep learning for accelerating computational homogenization schemes: application to flows in porous media

Shakoor, Modesar; Itier, Vincent; Mennesson, Jose

Document type :

Autre communication scientifique (congrès sans actes - poster - séminaire...): Communication dans un congrès avec actes

Title :

Deep learning for accelerating computational homogenization schemes: application to flows in porous media

Author(s) :

Shakoor, Modesar [Auteur]
Centre for Materials and Processes [CERI MP - IMT Nord Europe]
Itier, Vincent [Auteur]
Ecole nationale supérieure Mines-Télécom Lille Douai [IMT Nord Europe]
Centre de Recherche en Informatique, Signal et Automatique de Lille - UMR 9189 [CRIStAL]
Mennesson, Jose [Auteur]
Centre de Recherche en Informatique, Signal et Automatique de Lille - UMR 9189 [CRIStAL]
Centre for Digital Systems [CERI SN - IMT Nord Europe]
Ecole nationale supérieure Mines-Télécom Lille Douai [IMT Nord Europe]

Conference title :

9th European Congress on Computational Methods in Applied Sciences and Engineering (ECCOMAS)

City :

Lisbon

Country :

Portugal

Start date of the conference :

2024-06-03

English keyword(s) :

Computational homogenization
Deep learning
Porous media flows

HAL domain(s) :

Informatique [cs]/Modélisation et simulation

English abstract : [en]

While most approaches for modelling flows in porous media rely on Darcy’s law, a computational homogenization approach where both fine scale and coarse scale equations are of Navier-Stokes type has recently been developed ...
Show more >While most approaches for modelling flows in porous media rely on Darcy’s law, a computational homogenization approach where both fine scale and coarse scale equations are of Navier-Stokes type has recently been developed [1]. This alternative approach is promising as it eases the extension to inertial flows under a principle of multiscale virtual power. Its drawback, however, is that its implementation is quite involved. At each integration point of the Finite Element (FE) mesh of the coarse scale domain, indeed, are attached fine scale domains with their own meshes. Although solving this FExFE or FE² problem at each time increment of the simulation involves a computational complexity that is decreased by several orders as compared to what a single-scale approach would entail, it is still quite demanding.In this work, a deep learning model is developed in order to reduce the computational cost of the fine scale problems. This model is based on a recurrent neural network previously developed for fracture mechanics [2]. In the present work, however, because the objective is to integrate this model within the computational homogenization scheme, several new challenging issues are addressed. Among them, it is ensured that the model is intrinsically constrained for convexity and incompressibility. Solutions for coupling an FE code with a neural network are assessed as well.This presentation will briefly summarize the computational homogenization scheme for porous media flows, and then detail the deep learning model and its coupling with the FE code. The capabilities of the proposed approach as compared to single-scale and then FE² simulations will be demonstrated both in terms of accuracy and computation time.[1] M. Shakoor and C.H. Park, Computational homogenization of unsteady flows with obstacles, International Journal for Numerical Methods in Fluids, 95(4), 499-527, 2023.[2] K. Shinde, V. Itier, J. Mennesson, D. Vasiukov and M. Shakoor, Dimensionality reduction through convolutional autoencoders for fracture patterns prediction, Applied Mathematical Modelling, 114, 94-113, 2023.Show less >

Language :

Anglais

Peer reviewed article :

Oui

Audience :

Internationale

Popular science :

Non

ANR Project :

Modélisation et apprentissage des interfaces dans les écoulements biphasiques

Collections :

Centre de Recherche en Informatique, Signal et Automatique de Lille (CRIStAL) - UMR 9189

Source :

Harvested from HAL

Deep learning for accelerating computational ... BibTeX CSV Excel RIS

Deep learning for accelerating computational ...

BibTeX

CSV

Excel

RIS