Master Projects 2013

Vincent Zimmern – Bioinformatic Analysis of RNA-Seq Experiments

EPFL supervisor: Prof. Patrick Fraering

Merck Serono Chair in Neuroscience – CMSN

   Host Institution: University of Texas Southwestern, Dallas, USA

               Supervisor: Prof. Joachim Herz

During my master’s project at the University of Texas Southwestern Medical Center, I worked as a bioinformatician for the Herz and Yu labs, both of which collaborate on research related to neurodegenerative disease. The main project consisted in analyzing large RNA-sequencing datasets by aligning millions of RNA reads to mammalian genomes, deducing gene expression levels from these alignments, using statistical tests to assess the amount of alternative splicing and differential expression in important genes, and using clustering and gene ontology techniques in R and other statistical computing environments to obtain insight into disease mechanisms. This approach was applied to murine models of amyotrophic lateral sclerosis (ALS), Alzheimer’s disease, and frontal temporal lobar dementia (FTLD). Results from these analyses may lead to new insights in how to treat these neurodegenerative diseases.

Loïc Perruchoud – Measuring strategies for mobile nodes in a heterogeneous sensor network

EPFL supervisors: Prof. Alcherio Martinoli, Mr. William Christopher Evans

Distributed Intelligent Systems and Algorithms Laboratory – DISAL

Measuring and monitoring environmental fields, such as temperature, humidity, pollutant concentration or even radioactivity is important to understand and to manage the environment. To achieve such a goal, sensor networks are a very powerful tool. However, sensor readings are only valid locally and data in between the sensor nodes are obtained by interpolation. Particularly, in the case of environmental sensor networks, nodes may be deployed in areas with limited access, and further the ideal node distribution is often unclear at deployment time. As a result, interpolation between nodes may under some conditions give a poor picture of the underlying environmental processes. We are exploring the advantages of a small number of mobile nodes (e.g., miniature helicopters) may bring to an otherwise static sensor network.

What path should the mobiles nodes follow to maximize the utility of their travel time? To answer this question we consider a single mobile device and we assume that the path of this mobile device is computed by a base station that has access to the data measured by both static and mobile sensors. We also assume that the field does not change during a single flight.

Kriging interpolation is widely used in geostatistics. It estimated the interpolate values by taking a weighted average of the surrounding sampled values, weighted in function of the spatial correlation. Kriging has the advantage of giving an estimation of the interpolation error at each interpolated point. In this report, we study how we can dynamically adapt the path of a mobile sensor by minimizing the kriging error estimation in real-time. We propose an adaptive path planning algorithm based on the ant colony optimization metaheuristic that reduces the estimation of the kriging interpolation error in real-time.


Jérémie Despraz – Super Ball Bot: Structures for Planetary Landing and Exploration

EPFL supervisors: Prof. Auke Ijspeert, Mr. Mostafa Ajallooeian

Biorobotics Laboratory – BIOROB

Host institution: National Aeronautics and Space Administration (NASA), San Francisco, USA  Supervisor: Mr. Vytas SunSpiral

The goal of this Master project was to develop central pattern generator (CPG) based controls for a Planetary Exploration Tensegrity Robot. Development occured primarily in a physics-based tensegrity simulator, and involved an exploration of various candidate tensegrity structures and associated controls to roll the robot over various terrains. Tensegrity structures are uniquely appropriate for distributed CPG based control due to their compliant tension network which mirrors the properties of the CPG network. It is believed that tensegrity robot technology can play a critical role in future planetary exploration.

Artist view of tensegrity robots on the surface of Titan

Download “Super Ball Bot: Structures for Planetary Landing and Exploration”

Pascal Bienz – Coil sensitivity estimation for parallel magnetic resonance imaging

EPFL supervisors: Prof. Michael Unser, Mr. Emrah Bostan and Dr. Stamatis Lefkimmiatis

Biomedical Imaging Group – LIB

Parallel MRI (pMRI) increases the speed of the MRI acquisition by simultaneously receiving the data through a number of receiver coils with different spatial sensitivities. These sensitivity maps are used to reconstruct the MR image from missing k-space samples. Unfortunately, they are patient dependent and have to be calibrated for each person. The goal of this project is to perform joint reconstruction of the MR image from reduced k-space sampling and the coil sensitivity maps in order to reduce the acquisition time.
This project outlines the physical principles that underlie the acquisition of MRI data and explains the mathematical model of a reconstruction. By first assuming that the sensitivity maps are known during the iterative reconstruction scheme, we compare two different regularizers; namely, the total variation, and a regularizer based on the Schatten norm of the hessian. We finally show a way to perform the joint reconstruction of the MR image and the sensitivity maps.

Varun Sharma – Large Eddy Simulations of Wind Turbine Wakes

EPFL supervisor: Prof. M. Parlange

Laboratory of Environmental Fluid Mechanics and Hydrology – EFLUM

Dana Christen – 3D simulation of heterogeneous interface rupture

EPFL supervisors: Prof. Jean-François Molinari, David Kammer

Computational Solid Mechanics Laboratory, LSMS

This work presents the development of a framework for the three-dimensional simulation of slip fronts at the interface between a deformable solid body and a rigid surface.
Experimental observations of this phenomenon have been described by Ben-Dav et al. (2010) and their reproduction in two-dimensional numerical simulations has been reported by Rdiguet et al. (2013) and Kammer et al. (2012).
Preliminary simulations matching the original experimental setup reveal the geometry to be challenging to accurately simulate in the dimensions. Results from subsequent simulations using a modified geometry with properly scaled parameters are presented. While unachieved mesh convergence prevents the observation of slip fronts at the interface, results show the newly developed workflow to be working.

Nicolò Pagan – Coupling 1D system code OLGA with TransAT CFD

EPFL supervisor: Prof. François Gallaire

About ASCOMP, Zurich

Responsible: Dr. Djamel Lakehal

This work is about coupling a 3D CFD code (TransAT) and a 1D code (OLGA). The interest in this
work lies in the possibility of simulating real-world problems by saving computational time required
by a fully 3D simulation through the use of a 1D model in the sub-domains where a more detailed
description of the flow is not necessary.
A coupling scheme is proposed that enforces conservation of mass, momentum and energy across
a coupling boundary. Time discretization is based on a semi-implicit scheme. The implementation
of the conservation laws is validated on simple test cases. The fact that different physical models
are used in each code, however, may lead to discrepancies near the coupling boundary interface.
More complex problems are considered and the stability of the algorithm is tested for multi-phase flows, with phases crossing the coupling boundary in opposite directions. These flows are not usually studied in the framework of a coupling scheme due to their complexity. The results are satisfactory even though they are affected by the lack of 3D information in the domain simulated by the 1D code OLGA.
The coupling scheme has been applied to a real problem in the context of oil and gas applications.
We considered a multi-phase flow in a domain characterized by a complex 2D sub-domain, i.e. a
dome structure laying on the seabed and containing a jet spill of oil and gas slugs, connected to a 1D
sub-domain simulating a long riser. The robustness of the implemented coupling algorithm is shown
for this multi-phase problem with different phase flow directions and a phase-change interface model
applied at the coupling boundary.
This work shows mathematically the advantages of semi-implicit coupling schemes. We also highlight
the difficulties coming from having different physical models used by two different solvers and propose some specific solutions to mitigate those issues.
The stability of the method and the documentation provided concerning the validation cases are
sufficient for a release of the current version of its implementation in the TransAT code.

Mahmoud Jafargholi – Real time solution of network optimization problems

ABB Corporation Research Center, Baden-Dättwil, Switzerland

EPFL Supervisor: Prof. Colin Jones

Solving nonlinear optimization problems for real-time applications is a challenging task and has been restricted to applications with rather slow cycle times. The main reason for this restriction can be traced to the computational demand that optimization-based algorithms pose to the computational hardware. Although there exists a variety of both commercial and free solvers for nonlinear optimization problems, their rather generic nature prevents them from exploiting the problem structure inherent in some specific problem classes to solve a problem instance as efficient as possible. Furthermore, the complexity and/or licenses of standard solvers restrict their deployment on non-standard platforms such as embedded control systems. The effectiveness of structure-exploiting algorithms was recently shown for model predictive control problems with the solver FORCES ( A tailored high-speed nonlinear optimization solver for network problems will increase the range of possible applications which can be tackled with today’s hardware platforms.

Emmanuel Froustey – A practical inverse problem approach for phase imaging

EPFL supervisor: Prof. Michaël Unser

Biomedical Imaging Group

The problem of imaging transparent (phase-only) specimen is commonly encountered in biology. As an alternative to hardware-based solutions, one recovers the phase information by solving the Transport of Intensity Equation (TIE) that relates the phase and the intensity images. The fundamental advantage of the TIE is that it allows one to obtain the phase image by using a standard bright-field microscope.

The goal of the project is to solve the TIE by means of an inverse problem approach. An ImageJ plugin will be implemented to compute TIE solutions and validate the applicability on real datasets.

Mustafa Abd-Ur-Rehman – Parallelization of the Webots simulation loop & the ODE physics engine

EPFL supervisor: Prof. Auke Ijspeert

The Biorobotics Laboratory

Work at Cyberbotics on the multi-threading parallelisation of the physics engine used in Webots (ODE), follwing up a previous work with them on the parallelisation of the ODE collider. This master project will focus on the parallelisation of the LCP solver and more generally the parallelisation of the Webots simulation loop, in order to improve the overall performance of the Webots on multi-core computer systems. The investigation in particular the openmp parallel programming framework.

Diego Marcos Gonzales – Learning dictionaries for change detection in multi-temporal and hyperspectral Remote Sensing images

EPFL Supervisors: Prof. Jean-Philippe Thiran, Frank de Morsier and Devis Tuia

Signal Processing Laboratory 5 – LTS5

The recent advances in Sparse coding and Dictionary learning have shown to be very useful for classification problem. A dictionary is learned per class and a test sample is attribute to the dictionary resulting in the best reconstruction after sparse coding. This has an important potential for classification and change detection problems in remote sensing imagery.  This project aim at focusing more particularly to hyperspectral images which are challenging since high-dimensional. Adding the spatial and temporal information in the classification/detection process will be investigated and compared with state-of-the-art classifier.