Date and Place: Thursdays in Room 32349. For detailed dates see below!
Content
In the Scientific Computing Seminar we host talks of guests and members of the SciComp team as well as students of mathematics, computer science and engineering. Everbody interested in the topics is welcome.
List of Talks

Thu25Oct2018
11:45SC Seminar Room 32349
Viktor Leonhardt, TU Kaiserslautern
Title:
Simulation and Control of a KiteAbstract:
Producing clean renewable energy is a common energy source. But the current need for this type of energy is not yet sufficient to replace coal or gasbased energy sources. In the last decade, an alternative wind generation method, namely, Airborne Wind Energy System (AWES) is explored, which uses the strong winds at high altitudes. This thesis focuses on a framework to simulate and control of an Airborne Wind Energy System (AWES). In particular, a novel concept of a threetethered ground station called ”3r GroundGen AWES” is investigated. Thus, a behaviour based control is used to navigate the airborne system without the use of a predefined trajectory.

Thu15Nov2018
11:30SC Seminar Room 32349
Dr. HuaDong Yao, Division of Fluid Dynamics, Department of Mechanics and Maritime Sciences, Chalmers University of Technology
Title:
Hybrid CAA Methods for Predicting FlowInduced Exterior and Interior NoiseAbstract:
As one of the contaminating productions of flows, noise is emphasized to reduce for many civil and military applications. For commercial aircraft and ground vehicles (cars, trucks, highspeed trains, etc.), the noise pollutes urban and cabin environments. The noise signature of submarines is a crucial concern for cloaking them. Prediction of the noise using numerical methods is a challenging task. Since the noise has much lower energy than the noise sources in the flows, an approximation or modeling in the numerical methods developed for flow simulations can introduce significant spurious noise, which could even mask the real noise. Furthermore, mistreatments in the numerical schemes and modeling for acoustics simulations also account for the spurious noise generation. To tackle these numerical challenges, computational aeroacoustics (CAA) has been developed since the breakthrough by Lighthill in 1952. Lighthill proposed the acoustic analogy that isolates the flow and noise simulation based on the different physical characteristics between acoustic waves and hydrodynamic variables. The idea of the isolation is the principle foundation of hybrid CAA methods nowadays, which couple computational fluid dynamics (CFD) methods with linearized equations that describe acoustic wave generation and propagation with/without hydrodynamic convection. We will review the hybrid CAA methods that are commonly used for the noise prediction. The CFD methods include DNS, LES, DES, and RANS coupled with the stochastic modeling of hydrodynamic velocity fluctuations. The acoustic equations are formulated by means of the acoustic analogy methods (the Lighthill formulation and the FWH formulations, etc.), the boundary element method, the finite element method, and a simplified form of the acoustic perturbation equations. The hybrid methods will be introduced in conjunction with the practices for highlift facilities, landing gears, flapside edges, plenum fans, sail masts, vehicle sideview mirrors, and isotropic turbulence.

Thu22Nov2018
11:30SC Seminar Room 32349
Simon Extra, Brandenburg University of Technology CottbusSenftenberg
Title:
Preliminary Core Engine Design by Cascaded Optimization ProcessesAbstract:
Design of aero engines is a rather complex process using sophisticated software tools for obtaining optimal performance. Several projects involving the application of optimization tools to aero engine design have been undertaken to improve the design workflow. Most of these projects, however, focused on the optimal design of a specific component of the core engine only, where processes for compressor, combustor and turbine were based on proven preliminary design tools and yielded convincing results and improvements. This reflects the current engine design approach where optimization of components is performed separately by teams of specialists. Interfaces between the components are defined by an initial performance calculation and fixed throughout component optimization. This restricts the design space substantially and provides no information about the quality of the interface definitions used.
To overcome these limitations, the design space has to be extended and a coupled method for preliminary core engine design has to be used: Such an approach will be proposed and discussed in the presentation. It is based on collaborative optimization (CO) which allows a separate optimization of all component, whilst still observing overall optimization goals for the core engine. The strategy influences the interfaces between the components and coordinates the subtasks through a global optimizer to enforce a consistent, optimal solution. The subtasks use the component optimization processes already developed and tested, which had to be modified for implementing and coupling them in an overall automated optimization process. 
Thu06Dec2018
11:30SC Seminar Room 32349
Anne Schreuder, TU Kaiserslautern
Title:
Computing Greeks with Adjoint Monte CarloAbstract:
Pricing financial product is very important in the finance industry. Likewise is the computation of their price sensitivities. In this talk we are particular interested in the situation in which the price of a financial product depends on its future (expected) value at a given time T.
The question is when and how we can compute the dependency of the future expected value on the present value. A straightforward solution is using a stochastic version of the Euler method, which allows a very efficient reformulation as adjoint computation of Algorithmic Differentiation. 
Thu13Dec2018
9:00SC Seminar Room 32349
Avraam Chatzimichailidis, Fraunhofer ITWM
Title:
Second Order Methods applied to Deep Neural NetworksAbstract:
Optimizing deep neural networks involves finding a good enough minimum of a highly nonlinear and
nonconvex function. State of the art first order methods suffer from pathological curvature of the loss
landscape and successful convergence relies on the right metaparameter tweaking. Extending the optimizer
to second order eliminates these problems, at the cost of having to compute the inverse Hessian of the
deep neural network, which takes O(N^3).The Roperator allows efficient Hessianvectorproduct computation of DNNs in O(N), without having to
store the whole Hessian. Combining this operator together with the Lanczos algorithm, an iterative eigenvalue
solver, allows for an efficient computation of eigenvalues in DNNs.A framework is built that is able to visualize the loss landscape of DNNs together with the
trajectory taken by the optimizer. This is done by performing a PCA over the network parameters
at different points of the trajectory and choosing the two directions in parameter space with the most
variance. 
Thu13Dec2018
9:45SC Seminar Room 32349
Dr. Stefanie Günther, SciComp
Title:
Simultaneous ParallelinLayer Training for Deep Residual NetworksAbstract:
Deep residual networks (ResNets) have shown great promise to model complex data relations with applications in image classification, speech recognition, or text processing, among others. Despite the rapid methodological developments, compute times for ResNet training however can still be tremendous, measured in the order of hours or even days. While common approaches to decrease the training runtimes mostly involve dataparallelism, the sequential propagation through the network layers creates a scalability barrier where training runtimes increase linearly with the number of layers.
This talk presents an approach to enables concurrency accross the network layers and thus overcome this scalability barrier. The proposed method is inspired by the fact that the propagation through a ResNet can be interpreted as an optimal control problem. In this context, the discrete network layers are interpreted as the discretization of a timecontinuous dynamical system. Recent advances in parallelintime integration and optimization methods can thus be leveraged in order to speed up training runtimes. In particular, an iterative multigridreductionintime approach will be discussed, which recurively divides the time domain (i.e. the layers) into multiple time chunks that can be processed in parallel on multiple compute units. Additionally, the multigrid iterations enable a simultaneous optimization framework where weight updates are based on inexact gradient information.

Thu13Dec2018
11:00SC Seminar Room 32349
Title:
Parallel Multilevel ILU Preconditioner to Solve Large Linear System of EquationsAbstract:
The solution of large sparse linear systems is a ubiquitous problem in chemistry, physics, and engineering applications. Krylov subspace based iterative methods are preferred to solve the linear system instead of direct methods as they are faster and use less memory. These methods use preconditioners to accelerate the convergence of underlying iterative methods.
Parallel performance of iterative methods is largely determined by the scalability of the underlying iterative solver and its preconditioner. In particular, scalability of the preconditioner is the most challenging operation. For example, the incomplete LU decomposition (ILU) preconditioner algorithm is serial in nature. One should modify the serial ILU algorithm to extract the parallelism.
New approaches, where the kernels of the ILU preconditioners are split up into finegrained taskbased implementations to deal with intranode concurrency are promising. In this talk, we present parallel multilevel ILU preconditioner. Our future goals are shared memory implementation of the ILU preconditioner in linear solver library GaspiLS, followed by GASPI based distributed memory implementation.

Thu13Dec2018
11:45SC Seminar Room 32349
Javad Fadaie Ghotbi, Fraunhofer ITWM
Title:
Parallel Algebraic Multigrid Using GaspiLSAbstract:
Solving a linear system of N unknowns with Multigrid methods are optimal because they can solve with O(N) work. This optimality property is crucial for scaling up huge simulation on parallel computers. To accomplish this, the problem geometry guides us to design the multigrid component with the underlying system in mind. Algebraic multigrid is a technique for solving linear system based on multigrid framework, but without need any explicit geometric information. AMG contains the fundamental multigrid ingredients that based only on matrix entries. Various AMG algorithms with different efficiency properties have developed by researchers which target different problem classes.
In this presentation, we will introduce the AMG method, starting with a depiction of classical AMG and move on to Parallel AMG and recent developments. 
Thu17Jan2019
11:30SC Seminar Room 32349
Steffen Schotthöfer, TU Kaiserslautern
Title:
Sensitivity analysis on chaotic dynamical systemsAbstract:
More scientists and engineers use computer simulations, consequently the sensitivity analysis on the simulated systems is an essential part of that developement. Unfortunately, conventional approaches to compute the sensitivity of a system may fail if the system is chaotic. The results of these methods can be too large, often by orders of magnitude.
In this talk we discuss the Non Intrusive Least Squares Shadowing approach to compute the sensitivity of chaotic dynamical system. This method overcomes the failure of the conventional aproaches and is for many real world applications efficiently computable. 
Thu24Jan2019
11:30SC Seminar Room 32349
Dr. Rauno Cavallaro, Aerospace Engineering Group, Department of Bioengineering and Aerospace Engineering, Carlos III University Madrid
Title:
Towards tighter and higherfidelity multidisciplinary optimization for the design of unconventional aircraft: the aeroelastic PrandtlPlane caseAbstract:
Future aeronautics needs to face several challenges. As emphasized by the EU institutions, it is of foremost importance enhancing aircraft performances for a more sustainable and greener aviation. Very ambitious goals have been set for the near future, difficult to be achieved without radical technological novelties. For this reason, in terms of airframes, alternative aircraft configurations, e.g., Joined Wings and PrandtlPlanes, are currently studied by researchers all over the world. A further challenge is the need to reduce design time, risks and costs of aircraft development. This can be pursued introducing wider Multidisciplinary Analysis and Optimization (MDAO) with physicalbased analysis tools abinitio; higher fidelity tools guarantee a higher level of confidence in the design process, reducing risks of mispredictions with consequent need to step back to previous design phases and increase costs. Extended MDAO helps exploring design space and find more performant solutions not discoverable otherwise if considering few disciplines in isolation.
The two above aspects are synergic in the sense that for unconventional configurations, the physicalbased MDAO seems mandatory to assess the gains and convenience with higher level of confidence. Especially with Joined Wings, it has been observed a strong coupling between disciplines which calls for a tight MDAO since the early design stages. Regarding aeroelasticity of Joined Wings, two fundamental aspects need to be considered. The first one is the observed strong coupling between flight dynamics and aeroelasticity, with consequences not only on flying qualities, but also on flutter instabilities onset. The second aspect is the need for a higherfidelity assessment of unsteady aerodynamic forces to explore aeroelastic behavior of such innovative configuration in transonic regimes.
This presentation shows a unified flight dynamics/aeroelasticity framework analysis and results obtained on a PrandtlPlane. Moreover, an automated approach to correct unsteady aerodynamic generalized forces, typically obtained with tools solving potential flows, with CFDbased tools is presented.

Thu31Jan2019
11:30SC Seminar Room 32349
Kerstin Nicolay, TU Kaiserslautern
Title:
A Parallel Method for Community Detection in Large NetworksAbstract:
The analysis of networks on extremely large data sets has become essential in recent research. One main interest is the detection of communities within the network, but common examples like the social network Facebook or the citation network of Wikipedia easily exist of over a billion nodes. The problem of optimizing the so called modularity (a measure for determining the quality of a community) is NPhard, therefore the talk presents a heuristic but very fast approach of Blondel et al., the Louvain method, for finding communities in large networks. Furthermore, an analysis of a possible parallelization of the method is given.

Thu07Feb2019
11:30SC Seminar Room 32349
Théo Jeanneau, TU Kaiserslautern
Title:
The Smoothed Particles Hydrodynamics methodAbstract:
In high performance computing, most of the focus is on grid based methods, computed exclusively by CPUs. However, meshfree method can be a viable alternative depending on the situation, and it is even more the case when the computing can be helped by GPUs. The Smoothed Particles Hydrodynamics (SPH) is one of these methods, and also the most widely established meshfree method. It is particularly used in the fields of astrophysics, solid mechanics, and fluid dynamics.
In this presentation, we will introduce the SPH method, observe the results given for the example of a dam break, and the speedup obtained by computing on a GPU in addition to the CPU, using the DualSPHysics code.