Rohit Pochampalli, Chair for Scientific Computing (SciComp), TU Kaiserslautern

Title: New Approaches in Data Driven Turbulence Modeling.

Abstract:

RANS equations combined with turbulence models perform adequately at lower computational cost than high fidelity methods. One equation turbulence models such as Spalart-Allmaras(SA), produce good results for specific applications, such as flows past wings and airfoils. However, the latter suffers from inaccuracies on quantities such as lift and drag in certain regimes, such as separated flows. A direct means to improve existing models is by introducing correction functionals onto terms of the model. For a specific flow configuration, these correction functionals can be estimated by inverse problems based on experimental measurements. Data driven techniques generalize this process by learning the correction functional in terms of flow variables, using data obtained from inverse problems.
We present two new approaches that primarily address statistical aspects of turbulence modeling with machine learning. In particular, the correction functional varies depending on several factors including the flow conditions and topology. The modeling approaches taken up here reflect the circumstance that the data at hand does not conform to the classical assumptions of statistical learning theory. Consequently, the focus is on developing frameworks that can treat non-identically distributed data. The first approach introduces a means to reduce the dependency of data on flow domains. The second approach transforms the data into a high dimensional representation that captures the topology of the flow domain. Finally, the performance of the machine learning enhanced turbulence model is compared to the SA model.

How to join:

The talk is held online via Jitsi. You can join with the link https://jitsi.uni-kl.de/SciCompSeminar_01. Please follow the rules below:

• Use a chrome based browser (One member with a different browser can crash the whole meeting).
• Mute your microphone and disable your camera.
• If you have a question, raise your hand.

More information is available at https://www.rhrk.uni-kl.de/dienstleistungen/netz-telefonie/konferenzdienste/jitsi/.

Guillermo Suárez, Chair for Scientific Computing (SciComp), TU Kaiserslautern

Title: Turbulence Modeling: An inverse problem

Abstract:

The numerical simulation of turbulence using the Reynolds-Averaged Navier-Stokes (RANS) equations by considering Boussinesq’ eddy viscosity assumption is a well established mathematical model in both, industrial applications and research. However, this assumption, derived at the end of 19th century, presents severe limitations even for simple shear flows, causing major errors on the results. To compensate for these deficiencies, new turbulence models have been proposed.

Traditional development of turbulence models relied on mathematical fundamentals as well as on empirical results. Nowadays, thanks to the outstanding progress in machine learning methods and the easy availability to computers, an alternative approach for turbulence modeling is to use high fidelity data.

In this presentation we will discuss a new methodology based on data-driven methods to improve the capabilities of RANS equations, more in precisely, to overcome for the constrains induced by the Boussinesq’ approximation.

How to join:

The talk is held online via Jitsi. You can join with the link https://jitsi.uni-kl.de/SciCompSeminar_02. Please follow the rules below:

• Use a chrome based browser (One member with a different browser can crash the whole meeting).
• Mute your microphone and disable your camera.
• If you have a question, raise your hand.

More information is available at https://www.rhrk.uni-kl.de/dienstleistungen/netz-telefonie/konferenzdienste/jitsi/.

Dr. Pierluigi Piersimoni, Department of Physics and Technology, University of Bergen, Norway

Title: An introduction to charged particle therapy and imaging

Abstract:

Particle therapy, also known as hadrontherapy, particularly with proton beam, has seen an increased interest in the last decades for cancer treatment; a fact testified by the constantly growing number of dedicated medical centers spreading all over the world. The advances in particle therapy are closely tied to advances in accelerator technology, which allows nowadays the existence of compact, relatively inexpensive accelerator facilities.
The distinct pattern of energy deposited in matter by heavy charged particles, characterized by the Bragg peak arising at the end of their path, make them attractive for patient treatments from a physical, biological and clinical point of view. In principle, knowing the exact range of the projectile particles in the target would allow to deposit most of the dose on the tumor and to avoid undesired dose to the nearby healthy tissues. However, due to the stochastic nature of energy deposition, the exact calculation of the particle range is not trivial, especially for complex biological structures such as a human body. Currently, the relative stopping power (RSP, that is the stopping power of a certain material relative to that of water) needed for treatment planning is calculated starting from x-ray computed tomography (CT) and then converted to RSP using calibration procedures. This is one of the major sources of range uncertainties (up to 3-4%) and can reduce the efficacy of hadrontherapy treatments. Proton CT (pCT) has been acknowledged as capable to reduce range uncertainties to 1% or below, by directly measuring the beam RSP for a volume of interest.
In the first part of this presentation an overview of the development of particle therapy and the evolution of medical dedicated accelerators will be given. The second part will be focused on pCT and the challenges involved in building a fast, efficient pCT system.

How to join:

The talk is held online via Jitsi. You can join with the link https://jitsi.uni-kl.de/SciCompSeminar_03. Please follow the rules below:

• Use a chrome based browser (One member with a different browser can crash the whole meeting).
• Mute your microphone and disable your camera.
• If you have a question, raise your hand.

More information is available at https://www.rhrk.uni-kl.de/dienstleistungen/netz-telefonie/konferenzdienste/jitsi/.

Prof. Dr. Ralf Keidel, HS Worms

Title: SIVERT Research Group Goals

Abstract:

SIVERT (a newly founded research group) aims to unlock the potential of particle therapy and pCT by improving the reliability and precision of diagnoses and therapeutic results through robust, real-time visualization and reconstruction methods. Researching and developing novel machine learning approaches is the central theme of all activities, because of their promising potential in addressing the multiple competing requirements. In this talk we will try to give support for the development of research questions. Details on Monte Carlo simulation strategies, data management, data conversion, historical evaluation methods and shortcomings of the pCT detector and software systems are also addressed.

How to join:

The talk is held online via Jitsi. You can join with the link https://jitsi.uni-kl.de/SciCompSeminar_04. Please follow the rules below:

• Use a chrome based browser (One member with a different browser can crash the whole meeting).
• Mute your microphone and disable your camera.
• If you have a question, raise your hand.

More information is available at https://www.rhrk.uni-kl.de/dienstleistungen/netz-telefonie/konferenzdienste/jitsi/.

Dr. Michael Herbst, École des Ponts ParisTech, France

Title: High-throughput density-functional theory calculations: An interdisciplinary challenge

Abstract:

Density-functional theory (DFT) is one of the most widespread methods for simulating the quantum-chemical behaviour of electrons in matter. Applications of DFT cover a wide range of fields including materials research, chemistry or pharmacy. In recent years so-called high-throughput screening calculations have emerged, where the aim is to automatically perform simulations on a large range of compounds and extract a few interesting ones for further investigation. Such use cases require a careful balance between accuracy and runtime inside the chosen algorithms and as such pose novel challenges with respect to physical models, reliability and performance of DFT codes. Tackling such problems is inherently a multidisciplinary task.

With this in mind we recently started the the density-functional toolkit (DFTK, https://dftk.org), a novel DFT package written in pure Julia. Our code provides a joint software platform for DFT, which is accessible to different scientific communities. In particular it is fast enough for practical calculations, but also flexible to support toy problems for mathematical development in the field. Despite being less than two years old our code already features a sizable feature set including mixed MPI-thread-based parallelism, support for arbitrary floating point types, as well as an integration with wide-spread codes in the field. Based on such aspects we have already managed to sucessfully push the state of the art, e.g. by designing novel black-box algorithms for high-throughput screening calculation or by investigating computable a posteriori error estimators for DFT.

How to join:

The talk is held online via Jitsi. You can join with the link https://jitsi.uni-kl.de/SciCompSeminar_06. Please follow the rules below:

• Use a chrome based browser (One member with a different browser can crash the whole meeting).
• Mute your microphone and disable your camera.
• If you have a question, raise your hand.

More information is available at https://www.rhrk.uni-kl.de/dienstleistungen/netz-telefonie/konferenzdienste/jitsi/.

PD Dr. Stefan Langer, DLR Braunschweig

Title: On boundary value problems for RANS equations and two-equation turbulence models – Talk 1

Abstract:

Currently, in engineering computations for high Reynolds number turbulent flows, turbulence modeling continues to be the most frequently used approach to represent the effects of turbulence. Such models generally rely on solving either one or two transport equations along with the Reynolds-Averaged Navier-Stokes (RANS) equations. The solution of the boundary-value problem of any system of partial differential equations (PDEs) requires the complete delineation of the equations and the boundary conditions, including any special restrictions and conditions. In the literature, such a description is often incomplete, neglecting important details related to the boundary conditions and possible restrictive conditions, such as how to ensure satisfying prescribed values of the dependent variables of the transport equations in the far field of a finite domain. In these two lectures, which build on each other, we consider the following topics:

Talk 1:

Often a severe loss of reliability and efficiency of current solution methods for boundary-value problems for the compressible RANS equations can be observed. We discuss the possible influence of boundary values, as well as near-field and far-field behavior, on the solution of the RANS equations coupled with transport equations for turbulence modeling. An analysis is performed to analyze the near-wall and far-field behavior of the turbulence model variables. This allows an assessment of the decay rate of these variables required to realize the boundary conditions in the far field. A multigrid algorithm using implicit multistage Runge-Kutta methods as a smoother is developed. It is shown that (almost) all well known solution methods proposed in the literature about CFD are specializations of this general ansatz. It is demonstrated that for complex applications smoothers with a large amount of implicitness are required to reliably approximate steady-state solutions.

How to join:

The talk is held online via Jitsi. You can join with the link https://jitsi.uni-kl.de/SciCompSeminar_10. Please follow the rules below:

• Use a chrome based browser (One member with a different browser can crash the whole meeting).
• Mute your microphone and disable your camera.
• If you have a question, raise your hand.

More information is available at https://www.rhrk.uni-kl.de/dienstleistungen/netz-telefonie/konferenzdienste/jitsi/.

Dr. Roy Charles Swanson, NASA Langley Research Center

Title: On boundary value problems for RANS equations and two-equation turbulence models – Talk 2

Abstract:

Currently, in engineering computations for high Reynolds number turbulent flows, turbulence modeling continues to be the most frequently used approach to represent the effects of turbulence. Such models generally rely on solving either one or two transport equations along with the Reynolds-Averaged Navier-Stokes (RANS) equations. The solution of the boundary-value problem of any system of partial differential equations (PDEs) requires the complete delineation of the equations and the boundary conditions, including any special restrictions and conditions. In the literature, such a description is often incomplete, neglecting important details related to the boundary conditions and possible restrictive conditions, such as how to ensure satisfying prescribed values of the dependent variables of the transport equations in the far field of a finite domain. In these two lectures, which build on each other, we consider the following topics:

Talk 2:

In this work, we address the issue of a properly defined boundary-value problem, what we call a well-defined problem. To obtain a unique numerical solution for the well-defined problem, well-posedness is required. The basic requirements of a well-posed problem are reviewed, and the current status of proving well-posedness, including the relevance of the boundary data and a bounded solution, is briefly discussed. This sets the stage for the focus on well-posedness as it pertains to considering the RANS equations and the transport equations for modeling the effects of turbulence as a weakly coupled system. Emphasis is placed on the equations of the turbulence model, and how turbulence modeling can be interpreted as a parameter identification problem, specificaly an inverse
problem. A compelling argument (although not a proof) for ill-posedness is made for both direct and inverse problems. This argument is based on using surface pressure and skin-friction data that produces very different results for the distribution of the turbulent viscosity. Numerical examples for the RANS equations and two-equation turbulence models are presented that confirm the theoretical findings.

How to join:

The talk is held online via Jitsi. You can join with the link https://jitsi.uni-kl.de/SciCompSeminar_11. Please follow the rules below:

• Use a chrome based browser (One member with a different browser can crash the whole meeting).
• Mute your microphone and disable your camera.
• If you have a question, raise your hand.

More information is available at https://www.rhrk.uni-kl.de/dienstleistungen/netz-telefonie/konferenzdienste/jitsi/.

Prof. Qiqi Wang, Massachusetts Institute of Technology (MIT), USA

Title: The climatic butterfly effect — do numerical simulations capture the statistics of chaotic systems?

Abstract:

The butterfly effect is a well-known phenomenon in fluid dynamics. A small perturbation can lead to large differences later in a chaotic dynamical system, such as turbulent flows. Lorenz famously posed the question, does the flap of a butterfly’s wings in Brazil set off a tornado in Texas? The answer is now widely accepted to be yes. This result has significant consequences to simulations that resolve chaotic dynamics, e.g., in fluid flows.

Whereas a tiny perturbation can change the state of a chaotic system, it is unclear whether it can change the long-time statistics. Statistics of many turbulent flows are known to be stable, insensitive to initial conditions. Ergodic theory provided a foundation for such stability. Indeed, for many unsteady flow simulations to be meaningful, we must believe that their statistics are not super sensitive to perturbations such as numerics. Many researchers rationalize his belief with the concept and theory of shadowing in dynamical systems.

Having dedicated a decade of research into the theory of shadowing, the speaker has found problems in this theory. Even systems that satisfy the most idealized assumptions in the shadowing theory can be arbitrarily sensitive to parameter perturbations. This question thus resurfaces: do numerical simulations capture the statistics of chaotic systems? In this talk, we will illustrate why the theory of shadowing cannot answer this question. We will then construct a simple mathematical model in which arbitrarily small perturbations can significantly influence the statistics of a stable, ergodic system. In Lorenz’s analogy, an intelligent butterfly could control the climate. Engineers must construct numerical simulations more meticulously to predict the statistics of chaotic fluid flows.

To download the presentation slides, please click here!

How to join:

The talk is held online via Jitsi. You can join with the link https://jitsi.uni-kl.de/SciCompSeminar_12. Please follow the rules below:

• Use a chrome based browser (One member with a different browser can crash the whole meeting).
• Mute your microphone and disable your camera.
• If you have a question, raise your hand.

More information is available at https://www.rhrk.uni-kl.de/dienstleistungen/netz-telefonie/konferenzdienste/jitsi/.

Michaela Schmitt, TU Kaiserslautern

Title: Recent approaches in Opacity Optimization

Abstract:

In the field of scientific visualization flow visualization is a long existing technique. Scientists and engineers aim to derive information from visualizations of flow data, but, less important structures frequently obscure features of interest.
Thus, visual clutter and occlusion are common problems when it comes to inspecting flow data and different approaches have been developed to remedy for this problem.
Two major approaches exist to solve this problem, one of them being seeding algorithms. By smartly placed seeding points the resulting visualization can be optimized. However, this approach is not performant for real time interaction since it is costly to recompute seed lines.
The second class of approaches are selection algorithms. Instead of optimizing the choice of seed points, the set of lines is precomputed and clutter then reduced by smart selection of the lines computed earlier. Therefore, the generation of lines (e.g. numerical integration) is only performed during the preprocessing and will not be recomputed which makes it easier to interactively apply changes. Belonging to the selection-based algorithms is opacity optimization. It aims to maximize the opacity of important structures while removing clutter by minimizing the opacity of less important structures.

In this talk we will take a look on recent advances in the field of opacity optimization, focusing on the work from Zeidan et al., which proposed a novel technique utilizing moment-based techniques for signal reconstruction. Optimized opacities are calculated per fragment on an underlying, possibly arbitrary geometry.
We will compare it to other recent work and conclude with a discussion of the advantages and disadvantages found in the different approaches.

How to join:

The talk is held online via Jitsi. You can join with the link https://jitsi.uni-kl.de/SciCompSeminar_15. Please follow the rules below:

• Use a chrome based browser (One member with a different browser can crash the whole meeting).
• Mute your microphone and disable your camera.
• If you have a question, raise your hand.

More information is available at https://www.rhrk.uni-kl.de/dienstleistungen/netz-telefonie/konferenzdienste/jitsi/.

Mitesh Mittal, TU Kaiserslautern

Title: Deep learning approaches for medical image data segmentation and
classification

Abstract:

Deep neural networks are now the state-of-the-art machine learning
models across a variety of areas, from computer vision, recommendation
systems, natural language processing, and are also widely deployed in
academia and industry. Their potential for use in medical imaging
technology, medical data analysis, medical diagnostics and healthcare
in general, are slowly being realized. We will look at recent advances
and some associated challenges in deep learning applied to medical
image segmentation and image classification.

The image classification problems arise in diabetic retinopathy
diagnosis. Methods based on UCI experiments often fail on real diabetic
retinopathy data. Bayesian deep learning enables the estimation of
model uncertainty and sets benchmarks based on the uncertainty in
classification. The uncertainty level determines whether an expert
referral is necessary for better diagnosis. The BDL method outperforms
current UCI experiments which overfit their uncertainty to the dataset
and provide better diagnosis diabetic retinopathy diagnosis.

In medical image segmentation, we look at deep Convolutional Neural
Networks (CNN) that use data augmentation on available annotated
samples to give better performance. This model can be trained end-to-
end from very few images and the architecture takes into account the
context and localization of the image patches.
This approach outperforms the prior best method (a sliding-window
convolutional network) on the ISBI challenge for segmentation of
neuronal structures in electron microscopic stacks by large margins.

The striking performance of these methods motivates their application
to other complex problems in medical image data processing.

How to join:

The talk is held online via Jitsi. You can join with the link https://jitsi.uni-kl.de/SciCompSeminar_13. Please follow the rules below:

• Use a chrome based browser (One member with a different browser can crash the whole meeting).
• Mute your microphone and disable your camera.
• If you have a question, raise your hand.

More information is available at https://www.rhrk.uni-kl.de/dienstleistungen/netz-telefonie/konferenzdienste/jitsi/.