Florian DE VUYST, Ecole Normale Supérieure de Cachan: Toward live CFD-computing interaction and visualization using GPU acceleration
Multicore coprocessors like GPU allow for hardware parallel acceleration on workstations or laptops. With today’s more than 1000 cores per chip, it is possible to get theoretical Teraflo performance and about 20 Teraflops are expected by 2020. However, computational methods now have to pay attention to the Byte-per-flop ratio that measures the quantity of data to read/write in memory compared to the floating operations. We will especially discuss Lattice Boltzmann methods for the Navier-Stokes equations that are suitable methods for manycore computing. We will end the talk by a demo of interactive CFD computation and some comments on prospective new ways of computer-assisted design.
Ronan GOLHEN, MaxSea International: Ship route optimization as fast marching method
We present our time optimization algorithm for sailing vessel : MaxSea Routing. Sailing Ship performance is modeled using series of polar curves, indicating vessel speed for given wind speed and angle. Isochrones construction is a kind of adaptative space meshing, to resolve the time optimization problem, given a weather forecast. We want now to extend this algorithm to other kind of optimization, for example fuel consumption for cargo ship. We have modeled this kind of optimization as a 3D wave front propagation. 2D for ship route, 1D for motor thrust / consumption. We are currently exploring the Ordered Upwind Method, a kind of fast marching method. We will present current results, open problems.
Tuesday, October 1st, 10:00 a.m. to 11:30 a.m., Turing lecture hall, building 1, Paris-Rocquencourt center. Coffee from 9:45 a.m.
Alain DERVIEUX, INRIA Sophia Antipolis: A priori-based mesh adaptation for CFD
A. Carabias 1, E. Mbinky 2, H. Alcin 3, A. Belme 4, A. Loseille 2, F. Alauzet 2, S. Wornom 4, A. Dervieux 1
1 INRIA, BP 93, 06902 Sophia-Antipolis
2 INRIA, B.P. 105, 78153 Le Chesnay
3 INRIA, 200 Rue Vieille Tour, 33405 Talence
4 UPMC, 4 place Jussieu, 75252 Paris
5 LEMMA, 06410 Biot
Our approach to mesh adaptation relies on non-discretised formulations for finding a mesh which is the best in some well-identified sense [1, 2]. A mesh is represented by a continuous metric field M. We start from the compressible Navier-Stokes system, denoted in a variational way:
N S(u, φ) = 0 ∀ φ.
We focus on k-exact approximations, for which the approximation error vanishes if the exact solution is a polynomial of degree k, which we write in terms of a local interpolator:
u_exact =Π_M u_exact.
The approximation error norm is converted into a continuous functional by choosing a priori estimates. A priori estimates rely on what we call the implicit error, a discrete error expressed by
δu_M = Π_M u_exact − u_M,
and solution of a discrete error equation:
∂(NS)/∂u(u, δu_M , φ_M) = RHS.
In the goal-oriented analysis, we introduce a functional
δj = (g, u_exact − u_M)
and the corresponding adjoint state u∗. Choosing the discrete test function as
φ_M = Π_M u∗,
we get an estimate of the error functional expressed in terms of interpolation errors for fluxes hq:
δj ≈ Σ(fq (u, u∗ ), Π_M hq (u_exact ) − hq (u_exact ).
The right-hand side is minimised with respect to the metric M. This produces the optimal metric, which, in turn, is converted into an optimal mesh. Applications to second-order accurate computation of turbulent flows and to third-order accurate computations will be discussed.
REFERENCES
[1] A. Loseille, A. Dervieux, F. Alauzet, “Fully anisotropic goal-oriented mesh adaptation for 3D steady Euler equations”, J. Comp. Phys. (2010) 229: 2866-2897.
[2] A. Belme, A. Dervieux, F. Alauzet, “Time accurate anisotropic goal-oriented mesh adaptation for unsteady flows”, J. Comp. Phys. (2012) 231: 6323-6348.
Loïc MARECHAL, INRIA Paris-Rocquencourt: Génération automatique de maillages hexaédriques par la méthode octree
Les méthodes de générations de maillages basées sur une structure octree ont la réputation, justifiée, d’être robustes, mais de piètre qualité. Mes travaux consistent à en tirer le meilleur parti sans perdre sa généricité et son efficacité. Je présenterai un logiciel permettant de produire des maillages purement hexaédriques et conformes, avec des possibilités d’adaptation, de représentation des arêtes vives ou non-manifolds et d’insertion de couches limites.
Tuesday, November 5th, 10:00 a.m. to 11:30 a.m., Turing lecture hall, building 1, Paris-Rocquencourt center. Coffee from 9:45 a.m.
Tomás Chacón REBOLLO, University of Sevilla: Méthodes stabilisées et conditions inf-sup : application aux équations primitives de l’océan
Les méthodes stabilisés proportionnent des techniques de discrétisation des fluides incompressibles qui permettent d’utiliser la même interpolation pour vélocité et pression. Il existe une large variété de méthodes stabilisées, depuis les méthodes purement pénalisées jusqu’aux modèles de turbulence de type VMS (Variational Multi-Scale). La stabilité de la discrétisation de la pression pour ces méthodes est basée sur des conditions inf-sup spécifiques, adaptées à leur structure. Dans cette conférence d’un coté on va présenter quelques méthodes d’éléments finis stabilisées et les conditions inf-sup qui leur sont associées, et d’un autre on va justifier leur stabilité via une caractérisation comme méthodes mixtes avec des paires stables vitesse-pression. Finalement on va présenter leur application à l’approximation numérique des équations primitive de l’océan et quelques tests numériques.
Eric SAVIN, ONERA and Ecole Centrale de Paris: Kinetic models of elastic waves in bounded media with applications in structural acoustics
The dynamic response of built-up structures to low-frequency excitations can be predicted efficiently by reduced models derived from modal analyses. This representation is however much less relevant in their higher frequency (HF) ranges of vibration, when such systems exhibit a typical diffusive behavior. In this talk we will present some recent developments concerning the mathematical modeling and numerical simulations of the transient dynamics of engineering structures, based on the semiclassical analysis of HF solutions of wave systems. The transport models we consider are constructed adopting a kinetic point of view of wave propagation phenomena in heterogeneous and/or random media. The latter describes the asymptotic evolution properties of the underlying kinetic and strain energy densities. Nodal/spectral discontinuous “Galerkin” finite element methods are used for numerical simulations in order to account for their possible discontinuities at the boundaries and interfaces. The concurrence of the proposed framework with the engineering approach will be illustrated by several examples.
Tuesday, December 3rd, 10:00 a.m. to 11:30 a.m., Turing lecture hall, building 1, Paris-Rocquencourt center. Coffee from 9:45 a.m.
Marina VIDRASCU, INRIA Paris-Rocquencourt: Asymptotic expansions and domain decomposition
in collaboration with Pippo Geymonat, Soane Hendili, and Françoise Krasucki
Tho topic of this talk is to show how domain decomposition type methods combined to asymptotic expansions are a powerful tool to solve various heterogenous multi-scale problems. We will focus on the description of an efficient and reliable multi-scale method, using the matched asymptotic expansions and domain decomposition, as a suitable solution algorithm for elasticity problems with a large number of heterogeneities. In essence, we assume that the displacement and stress fields admit two asymptotic expansions, one far from the hetero-geneities (the outer one) another one close (the inner one). The asymptotic approach allows the replacement of the initial problem by a set of new ones where the layer of heterogeneities is replaced by a surface (in 3d) or a line in (2d) on which particular jumping conditions for displacement and stresses are defined. The order 0 outer problem is independent of the heterogeneities. For the first order outer problem the transmission coecients are given by several elementary inner problems posed on a representative cell. The number of such problems depends on the nature of the heterogeneities, they all have the same structure, and can be solved by domain decomposition. The relevant algorithms will be presented and numerical results will illustrate the method capabilities.
Quang Huy TRAN, IFP Energies Nouvelles: Bases mixtes ondelettes-gaussiennes pour le calcul des structures électroniques : exploration préliminaire en 1-D
in collaboration with Dinh Huong Pham, Valérie Perrier, and Luigi Genovese
Dans certaines simulations moléculaires quantiques, on souhaite obtenir une grande précision sur la fonction d’onde électronique au voisinage de chaque noyau (ensemble protons/neutrons d’un atome). La difficulté vient de ce que : (1) en théorie, celle-ci y exhibe un point de rebroussement ; (2) en pratique, pour des raisons liées au coût des calculs, il n’est pas avantageux de raffiner, même localement, le maillage. Pour ajouter le moins de degrés de liberté possible, l’idée est de considérer une fonction supplémentaire par noyau ayant la forme d’une gaussienne contractée, suffisamment proche de celle de la “vraie” solution. Les premiers essais en 1-D démontrent la pertinence d’une telle stratégie.
Tuesday 7th of January, 10:00 a.m. to 11:30 a.m., Turing lecture hall, building 1, Paris-Rocquencourt center. Coffee from 9:45 a.m.
Robert EYMARD, Université Paris-Est Marne-la-Vallée: Some convergence results for numerical schemes approximating two phase flow in porous media
This talk is focused on the mathematical study of numerical schemes, devoted to the approximation of two-phase flow in porous media.
The main difficulty in this study is the difference between the continuous setting, where a large class of test functions is available, and the discrete one, where the discrete space operators are limiting the possibilities for choosing test functions.
This difficulty is explored on different examples of two-phase flow models. In some of them, convergence proofs are only completed in the case of two-point flux approximation for the diffusion terms.
Gilbert ROGE, Dassault Aviation: Modélisation et méthodes numériques pour l’aérodynamique
L’apport du Calcul Scientifique dans la conception de forme aérodynamique est incontestable depuis une quinzaine d’années. La recherche de méthodes précises et performantes n’a pas cessé depuis les années 1980. Des problèmes faisant intervenir plusieurs disciplines (interactions fluide-structure – aéroélasticité, et donc CFD-CSM, pour traiter le phénomène de flottement – flutter par exemple ; aéroacoustique pour atteindre les objectifs ACARE – Horizon2020, …) sont maintenant résolus de manière industrielle. Ceci nécessite une fiabilité – robustesse et une efficacité compatible des exigences métier (certification, temps de cycle de design, …). La bonne utilisation (choix …) du contexte HPC (évolution rapide des architectures, loi de Moor, …) est un élément clef de la réussite. Malgré les avancées réalisées, des grands challenges sont encore à relever.
Tuesday, February 11, 10:00 a.m. to 11:30 a.m., Turing lecture hall, building 1, Paris-Rocquencourt center. Coffee from 9:45 a.m.
Patrick JOLY, ENSTA et INRIA Saclay – Île-de-France: From 3D Maxwell’s equations to transmission lines theory: an asymptotic approach
This work has been motivated by applications to non destructive testing of electric network using electromagnetic waves. Our goal is to derive simplified 1D models for the propagation of electromagnetic waves in thin co-axial cables, which are allowed to de lossy, heterogeneous and non cylindrical which prevents the use of standard techniques such as separation of variables and modal decomposition. For this we use an asymptotic approach considering the transverse dimension of the cable as a small parameter (as for beam theory in elasticity for instance). Doing so, we obtain and justify a generalized telegraphist’s model whose main properties will be presented.
Frédéric TAILLADE, EDF – R&D: Mise en œuvre de la réflectométrie temporelle et fréquentielle pour le diagnostic de structures de génie civil
La surveillance des structures de génie civils (ponts, enceintes de confinement, tunnels, barrages…), de structures aéronautiques en composites ou métalliques (portes, voilures…), des réseaux de transports (chaussées, lignes ferroviaires), des terrains (digues, sols…), du bâti, etc. requiert le développement de dispositifs de mesures durables, peu intrusifs et si possible déportés afin d’assurer la continuité de service et un haut niveau de sécurité.
De manière générale, l’auscultation visant à la maintenance prédictive des structures suppose la mise en œuvre d’un système de diagnostic en temps réel qui repose sur des capteurs délivrant une information fiable multi-physique et massivement distribuée en vue d’une analyse fondée sur des modèles de comportement et des outils avancés de traitement du signal intégrant la connaissance métier.
L’utilisation de capteurs intrinsèquement répartis, utilisant des méthodes d’interrogations TDR (Time Domain Reflectometry) et FDR (Frequency Domain Reflectometry) qu’elles soient dans le domaine optique ou électrique, délivrant une information en chaque point du capteur, constituent un ingrédient majeur de cette stratégie de contrôle de santé à grande échelle.
Nous focaliserons notre présentation sur quelques exemples d’applications de mise en œuvre de contrôle santé de structures de génie civil reposant sur la propagation d’ondes dans les lignes électriques. Nous montrerons ainsi la manière de transformer des éléments de structures en lignes électriques moyennant leur adaptation (diagnostic du remplissage des conduits de précontrainte extérieure, diagnostic des armatures de terre armées, mesure de la teneur en eau du béton, etc.). A travers ces exemples, nous tenterons de montrer les avantages, les limites et les besoins pour répondre aux enjeux d’auscultation en génie civil.
Tuesday, March 4, 10:00 a.m. to 11:30 a.m., Turing lecture hall, building 1, Paris-Rocquencourt center. Coffee from 9:45 a.m.
Luis ALMEIDA, CNRS & Laboratoire Jacques-Louis Lions, Université Paris 6 & INRIA Paris-Rocquencourt: Mathematical modeling of epithelial hole closure and wound healing
We will describe some work done in collaboration with A. Jacinto’s lab on epithelial wound healing in drosophila pupae and some more recent work on hole closure in in vitro systems (monolayers of MDCK cells) done in collaboration with B. Ladoux’s team. These works concern mathematical modeling of the contraction of actomyosin cables and/or lamellipodal crawling.
Pierre BADEL, Ecole Nationale Supérieure des Mines de St-Étienne: Identification of vascular soft tissue mechanical properties
The mechanics of living tissues is involved in many natural, vital, functions and processes. Unfortunately, it is also involved in many diseases with, often, fatal consequences if left untreated. Research on hard tissues like bone has reached advanced levels and is now mostly focused on mechano-chemo-biological interactions. The domain of soft tissues remains much less explored. Their mechanical properties are indeed very complex, and characterization is challenging.
The identification of the mechanical properties of vascular tissues is one the main topics of our group. Both in vivo and in vitro methods are developed in order to target both fundamental knowledge on the mechanics of such tissues and clinical objectives. In any case, inverse methods are used due to the complexity of the experimental data which can be obtained.
This seminar will address a short overview of the methods that we use for specific applications. Then, a more detailed focus will be proposed on in vitro identification of hyperelastic models for human ascending aortic aneurysm and rupture characterization. The context, the experimental challenges, and associated identification methods will be presented. Examples of the most striking results for understanding aneurysm mechanics and rupture will be shown and discussed.
Tuesday, April 1st, 10:00 a.m. to 11:30 a.m., Turing lecture hall, building 1, Paris-Rocquencourt center. Coffee from 9:45 a.m.
Pierre-Yves LAGREE, Institut Jean Le Rond d’Alembert, Université Paris 6: Granular flows from a continuum point of view
Granular matter (sand, rocks, gravels…) flows like a viscous fluid or resists to shear stress like a solid. It evolves from one state to the other over a distance typically of a few grain diameters. Here we adopt the fluid point of view and discribe granular matter as a continuum media with the so called µ(I) rheology: it is a plastic rheology were the normal and shear stress are proportional (Coulomb friction law). This non newtonian rheology is implemented in a 2D Navier-Stokes solver (Gerris). Moreover, comparison with Contact Dynamics simulations (the displacement of all the grains is computed) and experiments show the reliability of the μ(I)-rheology when modeling complex flow of granular matter in cases of avalanche flows (Bagnold flow), collapse of columns and flow in hourglass.
Peter BASTIAN, Universität Heidelberg: High-performance computing for flows in porous media
Simulation of flow and transport processes in porous media provides a formidable challenge and application field for high-performance computing. Relevant continuum-scale models include partial differential equations of elliptic, parabolic and hyperbolic type which are coupled through highly nonlinear coefficient functions. The multi-scale character and uncertainties in the parameters constitute an additional level of complexity but provide also opportunities for high-performance computing.
The first part of the talk is concerned with high-order DG methods for density driven flow. Using techniqes from spectral finite element methods an implementation is presented where the time per degree of freedom scales independent of the polynomial degree. It is shown that such schemes may be beneficial on todays HPC architectures relying on multiple cores with wide SIMD capabilities.
The second part of the talk will focus on the efficient solution of incompressible two-phase flow with a fully-coupled discontinuous Galerkin (DG) method. This approach will be compared with a standard cell-centered finite volume method and an efficient preconditioner for the arising linear systems will be discussed.
Tuesday, June 3, 10:00 a.m. to 11:30 a.m., Turing lecture hall, building 1, Paris-Rocquencourt center. Coffee from 9:45 a.m.
Edward LUKE, Mississippi State University: An autoparallelizing framework for irregular computations
Loci is computational framework that provides a new programming paradigm for developing numerical models that can be automatically mapped onto large scale parallel computing architectures. The programming paradigm makes use of a programming model that combines relational algebra to describe irregular access patterns with a simple logic programming model that facilitates composition of applications from simple kernels. The resulting programming model contains no explicit parallel directives, yet has demonstrated application scalability on distributed memory superclusters to thousands of processors. This talk will discuss the design philosophy of the Loci framework, programming model, and basic parallelization strategy. Performance and productivity comparisons with other programming models including MPI, OpenMP, Chapel, and Liszt on the LULESH hydrocode benchmark will be discussed. In addition, a brief overview of experience with production CFD codes such as the Loci/CHEM multiphysics solver will be dicussed. The Loci framework is free software provided under the GNU LGPL license.
Tuesday, July 1, 10:00 a.m. to 11:00 a.m., Turing lecture hall, building 1, Paris-Rocquencourt center. Coffee from 9:45 a.m.
