SC²S Colloquium - August 5, 2015

From Sccswiki
Jump to navigation Jump to search
Date: August 5, 2015
Room: 02.07.023
Time: 3:00 pm, s.t.

Agnes Koehler: Controlling a Toggle Switch Stochastic Modeling of Cell Populations

The two-gene Toggle Switch is a genetic circuit made of two mutually inhibiting proteins. This system is well known to present two stable equilibrium points in which one of the two proteins is present and fully represses the other, as well as one unstable equilibrium point in which both proteins repress each other with identical strength. The switch can be toggled by the addition of diffusible chemicals that weakens the repression capabilities of the proteins. The interest of my thesis was to control a single bacteria or even whole cell populations in silico at the unstable steady state by using two different control algorithms: PIC (proportional integral control) and OLC (open loop control). Applying this to single cells shows promising results that can be applied to in vivo experiments with the computed control parameters adapted to the biological system without tedious lab work. In particular, the application of the tuned control algorithms to whole cell populations reveals that although it seems counterintuitive, a population can be controlled by applying the same control to every single cell at the same time.

David Strätling: Concept Drift with Adaptive Sparse Grids

The ever increasing amount of data collected and stored, calls for algorithms to trans- form these huge amounts of data into useful information. In this context concept drifts are of increasing importance in the data mining and machine learning context, as today more and more data is read in streams rather than from static datasets. During the work on this thesis a classifier based on density estimation with adaptive sparse grids has been developed as part of the SGPP-Framework [21]. This classifier is able to adapt to the data using refinement. The data in batches and the batches are weight to cope with concept drifts. Multiple weighting methods have been implemented and various parameters allow adjusting the classifier to the dataset used. The different weighting methods, as well as the adaptivity have been tested separately and in combination on multiple datasets. On most datasets, the results showed an improvement in comparison to an ordinary classifier based on sparse grids. For other datasets the correct set of parameters has yet to be determined.

Stefan Gavranovic: Topology Optimization using GPGPU

In recent years a lot of research was invested in exploring and establishing the theory of topology optimization. The application field of topology optimiza- tion has expanded beyond structural analysis to include fluid flow, acoustics, heat transfer, nanophotonic devices and material designs. Therefore, engineers have an opportunity to mathematically obtain optimal designs for these cases where intuition and experience are not of great help. However, most of the con- ducted research was focused on 2D models. Due to high computational costs, performing topology optimization on 3D models may require hours of comput- ing time, or in some cases even days, which hinders rapid prototyping design process. Ideally, the engineers would like to have almost instantaneous feed- back in early stages of prototyping. Not as much research was conducted on improving computational efficiency as it was done for establishing the theory of topology optimization, hence it still stays an open topic for the research. Therefore, in this work we investigate the use of the multi-core architec- ture such as Graphics Processing Unit (GPU) by utilizing parallel programming framework Compute Unified Device Architecture (CUDA). Since the optimiza- tion process comes with a high computational price of performing the finite- element method (FEM) analysis at each optimization step, the main focus of this work is to implement an efficient solver for performing FEM analysis. In order to accomplish this, several steps are taken. Geometry under considera- tion is discretized with the help of hexahedral elements. This enables the use of highly efficient matrix-free geometric multigrid methods for solving linear sys- tem of equations. Geometric multigrid algorithm is adapted in such a way that it maps to GPU hardware, therefore resulting in execution times far superior to those when solving the problem on CPU. Furthermore, we address the problem of enforcing Dirichlet boundary con- ditions when using non conforming meshes. As the most suitable approach Nitsche method was chosen.