FEM@LLNL Seminar Series
We are happy to announce a new FEM@LLNL seminar series, starting in 2022, which will focus on finite element research and applications talks of interest to the MFEM community. We have lined up some excellent speakers for our first year and plan to keep adding more. Videos will be added to a YouTube playlist as well as this site's videos page.
SignUp
Fill in this form to signup for future FEM@LLNL seminar announcements.
Next Talk
David Moxey (King's College, London)
Time TBD, September 12, 2023
Previous Talks
Natasha Sharma (University of Texas at El Paso)
A Continuous Interior Penalty Method Framework for Sixth Order CahnHilliardtype Equations with applications to microstructure evolution and microemulsions
July 18, 2023
Abstract: The focus of this talk is on presenting unconditionally stable, uniquely solvable, and convergent numerical methods to solve two classes of the sixthorder CahnHilliardtype equations. The first class arises as the socalled phase field crystal atomistic model of crystal growth, which has been employed to simulate a number of physical phenomena such as crystal growth in a supercooled liquid, crack propagation in ductile material, dendritic and eutectic solidification. The second class, henceforth referred to as Microemulsion systems (ME systems) appears as a model capturing the dynamics of phase transitions in ternary oilwatersurfactant systems in which three phases namely a microemulsion, almost pure oil, and almost pure water can coexist in equilibrium. ME systems have several applications ranging from enhanced oil recovery to the development of environmentally friendly solvents and drug delivery systems. Despite the widespread applications of these models, the major challenge impeding their use has been and continues to be a lack of understanding of the complex systems which they model. Thus, building computational models for these systems is crucial to the understanding of these systems. The presence of the higher order derivative in combination with a timedependent process poses many challenges to the creation of stable, convergent, and efficient numerical methods approximating solutions to these equations. In this talk, we present a continuous interior penalty Galerkin framework for solving these equations and theoretically establish the desirable properties of stability, unique solvability, and firstorder convergence. We close the talk by presenting the numerical results of some benchmark problems to verify the practical performance of the proposed approach and discuss some exciting current and future applications.
Freddie Witherden (Texas A&M University)
FSSpMDM — Accelerating Small Sparse Matrix Multiplications by RunTime Code Generation
June 20, 2023
Abstract: Small matrix multiplications are a key building block of modern highorder finite element method solvers. Such multiplications describe the act of applying a specific finite element operator onto a set of state vectors. The small and irregular size of these multiplications makes them poor candidates for generic matrix multiplication routines. Moreover, for elements with a tensor product construction, the operators themselves can exhibit a significant degree of sparsity. In this talk, I will describe the code generation strategies employed by our Fixed Size Sparse MatrixDense Matrix (FSSpMDM) routine in libxsmm and show how these result in performant operator kernels for prismatic and hexahedral elements. Strategies will be described for both x8664 (AVX2/AVX512) and AARCH64 (NEON/SVE) instruction sets. Results will be presented on recent Intel and Apple CPUs and compared against the wellknown GiMMiK C code generation library.
Frank Giraldo (Naval Postgraduate School)
Using HighOrder ElementBased Galerkin Methods to Capture Hurricane Intensification
May 16, 2023
Abstract: Properly capture hurricane rapid intensification (where the winds increase by 30 knots in the first 24 hours) remains challenging for atmospheric models. The reason is that we need LEStype scales 𝒪(100m) which is still elusive due to computational cost. In this talk, I describe the work that we are doing in this area and how elementbased Galerkin Methods are being used to approximate spatial derivatives. I will also discuss the timeintegration strategy that we are exploring for this class of problems. In particular, we are exploring process Multirate methods whereby each process in a system of nonlinear partial different equations (PDEs) uses a timeintegrator and timestep commensurate with the wavespeed of that process. We have constructed Multirate methods of any order using extrapolation methods. Along this same idea, we have also developed a multimodeling framework (MMF) designed to replace the physical parameterizations used in weather/climate models. Our approach is to view the coarsescale and finescale models through the lens of Variational MultiScale (VMS) methods in order to give MMF a more rigorous mathematical foundation. Our end goal is to use MMF in order to better resolve the inner core of hurricanes. In addition, I will show some results using flux differencing discontinuous Galerkin Methods for constructing both Kinetic Energy Preserving and Entropy Stable methods and discuss why we need scalable models in order to achieve our goals. Our model, NUMA, is a 3D nonhydrostatic atmospheric model that runs on large CPU clusters and on GPUs.
Leszek F. Demkowicz (University of Texas at Austin)
Full Envelope DPG Approximation for Electromagnetic Waveguides. Stability and Convergence Analysis
April 25, 2023
Abstract: The presented work started with a convergence and stability analysis for the socalled full envelope approximation used in analyzing optical amplifiers (lasers). The specific problem of interest was the simulation of Transverse Mode Instabilities (TMI). The problem translates into the solution of a system of two nonlinear timeharmonic Maxwell equations coupled with a transient heat equation. Simulation of a 1 m long fiber involves the resolution of 10 M wavelengths. A superefficient MPI/openMP hp FE code run on a supercomputer gets you to the range of ten thousand wavelengths. The resolution of the additional thousand wavelengths is done using an exponential ansatz e^{ikz} in the zcoordinate. This results in a nonstandard Maxwell problem. The stability and convergence analysis for the problem has been restricted to the linear case only. It turns out that the modified Maxwell problem is stable if and only if the original waveguide problem is stable and the boundedness below stability constants are identical. We have converged to the task of determining the boundedness below constant. The stability analysis started with an easier, acoustic waveguide problem. Separation of variables leads to an eigenproblem for a selfadjoint operator in the transverse plane (in x,y). Expansion of the solution in terms of the corresponding eigenvectors leads then to a decoupled system of ODEs, and a stability analysis for a twopoint BVP for an ODE parametrized with the corresponding eigenvalues. The L^2orthogonality of the eigenmodes and the stability result for a single mode, lead then to the final result: the inverse boundedness below constant depends inversely linearly upon the length L of the waveguide. The corresponding stability for the Maxwell waveguide turned out to be unexpectedly difficult. The equation is vectorvalued so a direct separation of variables is out to begin with. An exponential ansatz in z leads to a nonstandard eigenproblem involving an operator that is nonself adjoint even for the easiest, homogeneous case. The answer to the problem came from a tricky analysis of the eigenproblem combined with the perturbation technique for perturbed selfadjoint operators. The use of perturbation theory requires an assumption on the smallness of perturbation of the dielectric constant (around a constant value) but with no additional assumptions on its differentiability (discontinuities are allowed). In the end, the final result is similar to that for the acoustic waveguide  the boundedness below constant depends inversely linearly on L.
Joachim Schöberl (Vienna University of Technology)
The Netgen/NGSolve Finite Element Software
March 28, 2023
Abstract: In this talk we give an overview of the open source finite element software Netgen/NGSolve, available from www.ngsolve.org. We show how to setup various physical models using FEniCSlike Python scripting. We discuss how we use NGSolve for teaching finite element methods, and how recent research projects have contributed to the further development of the NGSolve software. Some recent highlights are matrixvalued finite elements with applications in elasticity, fluid dynamics, and numerical relativity. We show how the recently extended framework of linear operators allows the utilization of GPUs for linear solvers, as well as timedependent problems.
Vikram Gavini (University of Michigan)
Fast, Accurate and Largescale Abinitio Calculations for Materials Modeling
March 7, 2023
Abstract: Electronic structure calculations, especially those using density functional theory (DFT), have been very useful in understanding and predicting a wide range of materials properties. The importance of DFT calculations to engineering and physical sciences is evident from the fact that ~20% of computational resources on some of the world's largest public supercomputers are devoted to DFT calculations. Despite the wide adoption of DFT, the stateoftheart implementations of DFT suffer from cellsize and geometry limitations, with the widely used codes in solid state physics being limited to periodic geometries and typical simulation domains containing a few hundred atoms. This talk will present our recent advances towards the development of computational methods and numerical algorithms for conducting fast and accurate largescale DFT calculations using adaptive finiteelement discretization, which form the basis for the recently released DFTFE opensource code. Details of the implementation, including mixed precision algorithms and asynchronous computing, will be presented. The computational efficiency, scalability and performance of DFTFE will be presented, which demonstrates a significant outperformance of widely used planewave DFT codes.
Som Dutta (Utah State University)
Quantifying the Potential of Covid19 Transmission Across Scales: Using SEM based NavierStokes solver and the CEAT
February 7, 2023
Abstract: The ongoing Covid19 pandemic has redefined our understanding of respiratory infectious disease transmission. The primary modes of transmission of the SARSCoV2 virus has been identified to be airborne, with human generated respiratory aerosols being the main carrier of the virus. Understanding the dispersion of these aerosols/droplets generated during speaking and coughing, has helped quantify potential for transmission and design effective mitigation strategies. Through my talk I will present how models at two ends of the spatiotemporal resolution spectrum helped quantify the physics and aid NASA Ames administrators design mitigation strategies. For the higher spatiotemporal resolution I will illustrate how the highorder SEM based NavierStokes solver Nek5000/NekRS was utilized to develop the models, including algorithms developed through CEED. I will present the two main modes of respiratory aerosol/droplet dispersal indoors, first at a shorter timescale through expiratory events like coughing, and second at a longer timescale through the room ventilation system induced flow and turbulence. At the other end of the spatiotemporal resolution, I will talk briefly about Covid19 Exposure Assessment Tool (CEAT), a novel tool developed to account for multiple factors that affect transmission. I will end my talk by briefly discussing how we can bridge the scales and heterogeneities in the problem with the aid of cutting edge computing and datadriven methods, so that we are fully prepared for the next pandemic. The work presented here has been facilitated by funding through DOE's National Virtual Biotechnology Laboratory (NVBL).
Stefan Henneking (University of Texas at Austin)
Bayesian Inversion of an AcousticGravity Model for Predictive Tsunami Simulation
January 10, 2023
Abstract: To improve tsunami preparedness, earlyalert systems and realtime monitoring are essential. We use a novel approach for predictive tsunami modeling within the Bayesian inversion framework. This effort focuses on informing the immediate response to an occurring tsunami event using nearfield data observation. Our forward model is based on a coupled acousticgravity model (e.g., Lotto and Dunham, Comput Geosci (2015) 19:327340). Similar to other tsunami models, our forward model relies on transient boundary data describing the location and magnitude of the seafloor deformation. In a realtime scenario, these parameter fields must be inferred from a variety of measurements, including observations from pressure gauges mounted on the seafloor. One particular difficulty of this inference problem lies in the accurate inversion from sparse pressure data recorded in the nearfield where strong hydroacoustic waves propagate in the compressible ocean; these acoustic waves complicate the task of estimating the hydrostatic pressure changes related to the forming surface gravity wave. Our spacetime model is discretized with finite elements in space and finite differences in time. The forward model incurs a high computational complexity, since the pressure waves must be resolved in the 3D compressible ocean over a sufficiently long time span. Due to the infeasibility of rapidly solving the corresponding inverse problem for the fully discretized spacetime operator, we discuss approaches for using compact representations of the parametertoobservable map.
Lin Mu (University of Georgia)
An Efficient and Effective FEM Solver for Diffusion Equation with Strong Anisotropy
December 13, 2022
Abstract: The Diffusion equation with strong anisotropy has broad applications. In this project, we discuss numerical solution of diffusion equations with strong anisotropy on meshes not aligned with the anisotropic vector field, focusing on application to magnetic confinement fusion. In order to resolve the numerical pollution for simulations on a nonanisotropyaligned mesh and reduce the associated high computational cost, we developed a highorder discontinuous Galerkin scheme with an efficient preconditioner. The auxiliary space preconditioning framework is designed by employing a continuous finite element space as the auxiliary space for the discontinuous finite element space. An effective line smoother that can mitigate the highfrequency error perpendicular to the magnetic field has been designed by a graphbased approach to pick the line smoother that is approximately perpendicular to the vector fields when the mesh does not align with anisotropy. Numerical experiments for several benchmark problems are presented to validate the effectiveness and robustness.
Garth Wells (University of Cambridge)
FEniCSx: design of the next generation FEniCS libraries for finite element methods
November 8, 2022
Abstract: The FEniCS Project provides libraries for solving partial differential equations using the finite element method. An aim of the FEniCS Project has been to provide highperformance solver environments that closely mirror mathematical syntax, with the hypothesis that highlevel representations means that solvers are faster to write, easier to debug, and can deliver faster runtime performance than is reasonably possible by hand. Using domainspecific languages and code generation techniques, arguably the FEniCS libraries delivered on these goals for a set of problems. However, over time limitations, including performance and extensibility, become clear and maintainability/sustainability became an issue.Building on experiences from the FEniCS libraries, I will present and discuss the design on the next generation of tools, FEniCSx. The new design retains strengths of the past approach, and addresses limitations using new designs and new tools. Solvers can be written in C++ or Python, and a number of design changes allow the creation of flexible, fast solvers in Python. In the second part of my presentation, I will discuss highperformance finite element kernels (limited to CPUs on this occasion), motivated by the Center for Efficient Exascale Discretizations 'bakeoff' problems, and which would not have been possible in the original FEniCS libraries. Double, single and halfprecision kernels are considered, and results include (i) the observation that kernels with vector intrinsics can be slower than autovectorised kernels for common cases, and (ii) a cacheaware performance model which is remarkably accurate in predicting performance across architectures.
Dennis Ogiermann (University of Bochum)
Computing Meets Cardiology: Making Heart Simulations Fast and Accurate
September 13, 2022
Abstract: Heart diseases are an ubiquitous societal burden responsible for a majority of deaths world wide. A central problem in developing effective treatments for heart diseases is the inherent complexity of the heart as an organ. From a modeling perspective, the heart can be interpreted as a biological pump involving multiple physical fields, namely fluid and solid mechanics, as well as chemistry and electricity, all interacting on different time scales. This multiphysics and multiscale aspect makes simulations inherently expensive, especially when approached with naive numerical techniques. However, computational models can be extraordinarily useful in helping us understanding how the healthy heart functions and especially how malfunctions influence different diseases. In this context, also information about possible weaknesses of therapies can be obtained to ultimately improve clinical treatment and decision support. In this talk, we will focus primarily on two important model classes in computational cardiology and their respective efficient numerical treatment without compromising significant accuracy. The first class is the problem of computing electrocardiograms (ECG) from electrical heart simulations. Since ECG measurements can give a wide range of insights about a wide range of heart diseases they offer suitable data to validate our electrophysiological models and verify our numerical schemes on organscale. Known numerical issues, arising in the context of electrophysiological models, will be reviewed. The second class addresses bidirectionally coupled electromechanical models and their efficient numerical treatment. Focus will be on a unified spacetime adaptive operator splitting framework developed on top of MFEM which proves highly efficient so far for the investigated model classes while still preserving high accuracy.
Ricardo Vinuesa (KTH)
Modeling and Controlling Turbulent Flows through Deep Learning
August 23, 2022
Abstract: The advent of new powerful deep neural networks (DNNs) has fostered their application in a wide range of research areas, including more recently in fluid mechanics. In this presentation, we will cover some of the fundamentals of deep learning applied to computational fluid dynamics (CFD). Furthermore, we explore the capabilities of DNNs to perform various predictions in turbulent flows: we will use convolutional neural networks (CNNs) for nonintrusive sensing, i.e. to predict the flow in a turbulent open channel based on quantities measured at the wall. We show that it is possible to obtain very good flow predictions, outperforming traditional linear models, and we showcase the potential of transfer learning between friction Reynolds numbers of 180 and 550. We also discuss other modelling methods based on autoencoders (AEs) and generative adversarial networks (GANs), and we present results of deepreinforcementlearningbased flow control.
Jeffrey Banks (RPI)
Efficient Techniques for Fluid Structure Interaction: Compatibility Coupling and Galerkin Differences
July 26, 2022
Abstract: Predictive simulation increasingly involves the dynamics of complex systems with multiple interacting physical processes. In designing simulation tools for these problems, both the formulation of individual constituent solvers, as well as coupling of such solvers into a cohesive simulation tool must be addressed. In this talk, I discuss both of these aspects in the context of fluidstructure interaction, where we have recently developed a new class of stable and accurate partitioned solvers that overcome addedmass instability through the use of socalled compatibility boundary conditions. Here I will present partitioned coupling strategies for incompressible FSI. One interesting aspect of CBCbased coupling is the occurrence of nonstandard and/or highderivative operators, which can make adoption of the techniques challenging, e.g. in the context of FEM methods. To address this, I will also discuss our newly developed Galerkin Difference approximations, which may provide a natural pathway for CBCs in an FEM context. Although GD is fundamentally a finite element approximation based on a Galerkin projection, the underlying GD space is nonstandard and is derived using profitable ideas from the finite difference literature. The resulting schemes possess remarkable properties including nodal superconvergence and the ability to use large CFLone time steps. I will also present preliminary results for GD discretizations on unstructured grids using MFEM.
Paul Fischer (UIUC/ANL)
Outlook for Exascale Fluid Dynamics Simulations
June 21, 2022
Abstract: We consider design, development, and use of simulation software for exascale computing, with a particular emphasis on fluid dynamics simulation. Our perspective is through the lens of the highorder code Nek5000/RS, which has been developed under DOE's Center for Efficient Exascale Discretizations (CEED). Nek5000/RS is an open source thermal fluids simulation code with a long development history on leadership computing platformsit was the first commercial software on distributed memory platforms and a Gordon Bell Prize winner on Intel's ASCII Red. There are a myriad of objectives that drive software design choices in HPC, such as scalability, lowmemory, portability, and maintainability. Throughout, our design objective has been to address the needs of the user, including facilitating data analysis and ensuring flexibility with respect to platform and number of processors that can be used.
When running on largescale HPC platforms, three of the most common user questions are

How long will my job take?

How many nodes will be required?

Is there anything I can do to make my job run faster?
Additionally, one might have concerns about storage, postprocessing (Will I be able to analyze the results? Where?), and queue times. This talk will seek to answer several of these questions over a broad range of fluidthermal problems from the perspective of a Nek5000/RS user. We specifically address performance with data for NekRS on several of the DOE's preexascale architectures, which feature AMD MI250X or NVIDIA V100 or A100 GPUs.
Mike Puso (LLNL)
Topics in Immersed Boundary and Contact Methods: Current LLNL Projects and Research
May 24, 2022
Abstract: Many of the most interesting phenomena in solid mechanics occurs at material interfaces. This can be in the form of fluid structure interaction, cracks, material discontinuities, impact etc. Solutions to these problems often require some form of immersed/embedded boundary method or contact or combination of both. This talk will provide a brief overview of different lab efforts in these areas and presents some of the current research aspects and results using from LLNL production codes. Technically speaking, the methods discussed here all require Lagrange multipliers to satisfy the constraints on the interface of overlapping or dissimilar meshes which complicates the solution. Stability and consistency of Lagrange multiplier approaches can be hard to achieve both in space and time. For example, the wrong choice of multiplier space will either be overconstrained and/or cause oscillations at the material interfaces for simple statics problems. For dynamics, many of the basic time integration schemes such as Newmark's method are known to be unstable due to gaps opening and closing. Here we introduce some (nonNitsche) stabilized multiplier spaces for immersed boundary and contact problems and a structure preserving time integration scheme for long time dynamic contact problems. Finally, I will describe some ongoing efforts extending this work.
Robert Chiodi (UIUC)
CHyPS: An MFEMBased Material Response Solver for Hypersonic Thermal Protection Systems
April 26, 2022
Abstract: The University of Illinois at UrbanaChampaign's Center for Hypersonics and Entry Systems Studies has developed a material response solver, named CHyPS, to predict the behavior of thermal protection systems for hypersonic flight. CHyPS uses MFEM to provide the underlying discontinuous Galerkin spatial discretization and linear solvers used to solve the equations. In this talk, we will briefly present the physics and corresponding equations governing material response in hypersonic environments. We will also include a discussion on the implementation of a direct Arbitrary LagrangianEulerian approach to handle mesh movement resulting from the ablation of the material surface. Results for standard community test cases developed at a series of Ablation Workshop meetings over the past decade will be presented and compared to other material response solvers. We will also show the potential of highorder solutions for simulating thermal protection system material response.
Tamas Horvath (Oakland University)
SpaceTime Hybridizable Discontinuous Galerkin with MFEM
March 29, 2022
Abstract: Unsteady partial differential equations on deforming domains appear in many reallife scenarios, such as wind turbines, helicopter rotors, car wheels, freesurface flows, etc. We will focus on the spacetime finite element method, which is an excellent approach to discretize problems on evolving domains. This method uses discontinuous Galerkin to discretize both in the spatial and temporal directions, allowing for an arbitrarily highorder approximation in space and time. Furthermore, this method automatically satisfies the geometric conservation law, which is essential for accurate solutions on timedependent domains. The biggest criticism is that the application of spacetime discretization increases the computational complexity significantly. To overcome this, we can use the highorder accurate Hybridizable or Embedded discontinuous Galerkin method. Numerical results will be presented to illustrate the applicability of the method for fluid flow around rigid bodies.
Tobin Isaac (Georgia Tech)
Unifying the Analysis of Geometric Decomposition in FEEC
March 22, 2022
Abstract: Two operations take function spaces and make them suitable for finite element computations. The first is the construction of tracefree subspaces (which creates "bubble" functions) and the second is the extension of functions from cell boundaries into cell interiors (which create edge functions with the correct continuity): together these operations define the geometric decomposition of a function space. In finite element exterior calculus (FEEC), these two operations have been treated separately for the two main families of finite elements: full polynomial elements and trimmed polynomial elements. In this talk we will see how one constructor of tracefree functions and one extension operator can be used for both families, and indeed for all differential forms. We will also examine the practicality of these two operators as tools for implementing geometric decompositions in actual finite element codes.
Raphaël Zanella (UT Austin)
Axisymmetric MFEMBased Solvers for the Compressible NavierStokes Equations and Other Problems
March 1, 2022
Abstract: An axisymmetric model leads, when suitable, to a substantial cut in the computational cost with respect to a 3D model. Although not as accurate, the axisymmetric model allows to quickly obtain a result which can be satisfying. Simple modifications to a 2D finite element solver allow to obtain an axisymmetric solver. We present MFEMbased parallel axisymmetric solvers for different problems. We first present simple axisymmetric solvers for the Laplacian problem and the heat equation. We then present an axisymmetric solver for the compressible NavierStokes equations. All solvers are based on H^1conforming finite element spaces. The correctness of the implementation is verified with convergence tests on manufactured solutions. The NavierStokes solver is used to simulate axisymmetric flows with an analytical solution (Poiseuille and TaylorCouette) and an air flow in a plasma torch geometry.
Robert Carson (LLNL)
An Overview of ExaConstit and Its Use in the ExaAM Project
February 1, 2022
Abstract: As additively manufactured (AM) parts become increasingly more popular in industry, a growing need exists to help expediate the certifying process of parts. The ExaAM project seeks to help this process by producing a workflow to model the AM process from the melt pool process all the way up to the part scale response by leveraging multiple physics codes run on upcoming exascale computing platforms. As part of this workflow, ExaConstit is a nextgeneration quasistatic, solid mechanics FEM code built upon MFEM used to connect local microstructures and local properties within the part scale response. Within this talk, we will first provide an overview of ExaConstit, how we have ported it over to the GPU, and some performance numbers on a number of different platforms. Next, we will discuss how we have leveraged MFEM and the FLUX workflow to run hundreds of highfidelity simulations on Summit inorder to generate the local properties needed to drive the part scale simulation in the ExaAM workflow. Finally, we will show case a few other areas ExaConstit has been used in.
Guglielmo Scovazzi (Duke University)
The Shifted Boundary Method: An Immersed Approach for Computational Mechanics
January 20, 2022
Abstract: Immersed/embedded/unfitted boundary methods obviate the need for continual remeshing in many applications involving rapid prototyping and design. Unfortunately, many finite element embedded boundary methods are also difficult to implement due to the need to perform complex cell cutting operations at boundaries, and the consequences that these operations may have on the overall conditioning of the ensuing algebraic problems. We present a new, stable, and simple embedded boundary method, named "shifted boundary method" (SBM), which eliminates the need to perform cell cutting. Boundary conditions are imposed on a surrogate discrete boundary, lying on the interior of the true boundary interface. We then construct appropriate field extension operators, by way of Taylor expansions, with the purpose of preserving accuracy when imposing the boundary conditions. We demonstrate the SBM on largescale solid and fracture mechanics problems; thermomechanics problems; porous media flow problems; incompressible flow problems governed by the NavierStokes equations (also including free surfaces); and problems governed by hyperbolic conservation laws.
Future Talks
Ben Southworth (Los Alamos National Laboratory)
October 17, 2023
Youngsoo Choi (Lawrence Livermore National Laboratory)
November 7, 2023