Pascal Bienz – Extraction of Handwritten Medical Data from Paper Forms in Malawi

Supervisor: Dr. Loïc Baboulaz

Audiovisual Communications – LCAV

Baobab Health, a technological NGO in Malawi, is replacing Malawi’s traditional paper-based system of medical records with an electronic system in hospitals and clinics, with the goal of improving healthcare delivery and reducing its costs through better planning.
While the electronic system includes entries of patients since its beginning (1.5 million patients registered), it is also necessary to populate it with historical handwritten data in paper format to allow for deeper analysis of the health situation in Malawi. For various reasons, this is a time-consuming and costly process for Baobab Health.
The proposed project is to develop a fully working product that efficiently extracts information from a large set of digital images representing medical records. It should require minimum human intervention and be 100% accurate due to the nature of the data itself.

Daniel Wolfensberger – Measuring strategies for mobiles nodes in a heterogeneous sensor network

Supervisor: William Christopher Evans

Distributed Intelligent Systems and Algorithms Laboratory – DISAL

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.
A small number of mobile nodes (e.g.,
miniature helicopters) may bring an important advantage to an otherwise static sensor network.
The final goal of this project is to develop a strategy for choosing points in space and time that a mobile node can measure such that this picture is greatly improved.
The use of geostatistics will make it possible to quantify the quality of the interpolation, per example by defining a fitness function based on the kriging variance.
Meta-heuristic optimization algorithms, like particle swarm optimization can then be used to find out a near optimal sampling strategy for the mobile nodes.

Fabrizio Rompineve Sorbello – Aspects of hard scatterings in current LHC analysis

Supervisor: Prof. Stefano Frixione

Institute of Theoretical Physics – ITP & CERN

The primary aim of this project is that of exploiting a set of computer codes, collectively known as aMC@NLO (see amcatnlo.cern.ch), that are able to compute, numerically and in a fully automated manner, the cross sections for any user-defined scattering processes at the first non-trivial order in the perturbation theory of the coupling constant of QCD (the theory of strong interactions, which is dominant at the LHC). Such cross sections may or may not be combined with an Event Generator simulation, which allows one to obtain final states that are faithful representations of those actually occurring in high-energy hadron collisions.
The idea is that of investigating aspects of hard scatterings that pose challenging problems in current LHC analyses. These include the production of a Standard-Model Higgs boson in association with up to two light jets, and with a top-antitop pair, and that of a W boson is association with light jets. The achievement of this project requires a command of the aMC@NLO codes, the ability to understand the physics of an Event Generator, and the capability to write a physics analysis to be employed in the latter. Although most of the aMC@NLO codes are set up and tested, parts of this project could require the writing of code add-ons.

Cyril Misev – Single Particle Simulation in 3D Tokamak magnetic fields

Responsibles: Dr. Jonathan Graves, David Pfefferlé

Centre de Recherches en Physique des Plasmas – CRPP

The objective of this project is to contribute to the improvement of an existing code used for simulating the behaviour of a population of particles in a magnetic field. Improvements will take into account calculation speed and accuracy (physical and numerical).
The guiding centre particle trajectory code to be improved presently employs tri-linear interpolation of the 3D magnetic field in order to propagate the trajectory equations for the single particle position as it moves through the magnetic field. This can cause error, especially where the grid is coarse relative to the variation in the magnetic field.
The external magnetic field contributions are provided from an equilibrium code through Fourier modes in the poloidal and toroidal directions, but discretized in the third (radial) direction. However in the guiding centre code, the toroidal and poloidal coordinates are presently discretised and interpolated. The goal is in part to avoid this unnecessary discretisation step in the poloidal and toroidal coordinates, and thus to calculate the set of magnetic field values at the exact position of the particle at every timestep of the simulation.
To further reduce computation time, the code should find an optimum minimum number of toroidal and poloidal modes yielding a prescribed acceptable error in the equilibrium from the equilibrium code. A special requirement to the code is that physical properties, most importantly div(B)=0 must be satisfied. Regarding the discretisation in the radial direction, the goal will be to implement a variable (non-equidistant) grid.
Convergence studies of trajectories as well as simulation time comparisons with the initial code will be done using different interpolation techniques and radial grid point distributions will be used to test the new code in terms of its efficiency.

Duccio Malinverni – Dynamical Monte-Carlo simulation of polymers in confined space: Implementation of a new algorithm

Responsible: Prof. Pablo De Los Rios

Laboratory of Statistical Biophysics – LBS

Polymers confined in space are present in many real-life applications: Compacted DNA, thin layers, physics of dielectrics among others. The conformation of such polymers leads to a competition between imposed external geometry and internal arrangement of the monomers composing the polymer. The complexity of these arrangement naturally lead to the use of computer simulations of these systems in the field of statistical physics of polymers. Among the family of Monte-Carlo simulations, two types of algorithms exist:

The statical approach consists of sequentially adding a monomer to the polymer chain in a random direction, rejecting the chains which violate compatibility with the general chain (no monomer overlap, chain inside the geometry domain) until a given number of polymers of N monomers are generated.
The dynamical method starts with a chain of N monomers, and stochastically moves a randomly chosen monomer in a random direction, then again checking if the new conformation is compatible with the internal and external constraints until a given number of polymer conformations are generated.
As the number of monomers is increased, and the confined space decreased, the dynamical method tends to reject more and more conformations mainly due to a conflict with the confined geometry.
In this project, a new dynamical algorithm is implemented in which the number of these rejections should be decreased while generating conformations. The goal of this project is to implement this algorithm, validate it against theoretical (end-to-end distance and gyration radius, auto-correlation function,…) and other numerical methods, and benchmark its performance.

Dana Christen – Multi-level preconditioner for solving the Navier-Stokes equations in hemodynamics applications

Responsibles: Dr. Simone Deparis, Mr. Gwenol Grandperrin

Chair of Modelling and Scientific Computing – CMCS

We present a multi-level algorithm to approximate the inverse of the fluid block in a Navier-Stokes saddle-point matrix where the coarse level is defined as a restriction of the degrees of freedom to the degrees of a lower order finite elements approximation.
A one-level scheme involving P1 and P2 finite elements is studied in details and several transfer operators are compared by means of two reference problems.
Numerical results show that restriction and prolongation operators based on projection techniques lead to faster GMRES convergence of the fluid part, when compared to operators based on interpolation techniques.

Federico Hernan Martinez Lopez – Solving didactical problems with CUDA

Supervisors: Prof. Roger Hersch and Remi Bloch

Peripheral Systems Laboratory – LSP

The objective of the project is to implement some algorithms which by nature expose high parallelization and some others that may not be so obvious in the GPGPU. The main tasks will be:
– Identification of massively parallel regions of the algorithms
– Implementation of the algorithms in GPGPU
– Identification and implementation of performance improvement in running time
– Analysis of the theoretical vs practical speedup
– Comparison of results with different parallel architectures (Multicore, clusters, etc)

Ivan Slijepcevic – A Parallel Particle Swarm Optimization engine for a universal optimization environment

Supervisor: Prof. Matteo Dal Peraro

Laboratory for Biomolecular Modeling – LBM

In some optimization problems the evaluation of the fitness function may require complex operations, that cannot be represented by a simple algebraic expression. These operations may typically involve manipulations of files or complex data structures, as well as calls to an external programs.
Parallel Optimization Workbench (POW) is a python-based optimization framework helping the developer to tackle these problems with minimal production of code. In order to evaluate the fitness function, POW allows the manipulation of any data structure as well as the call of external programs. The exploration of the search space is performed by an enhanced version of Particle Swarm Optimization (PSO Kick and Reseed, PSO-KaR), working in a parallel fashion using MPI libraries.
The aim of this project is to implement PSO-KaR in C++ and integrate it in the existing POW framework. The code will be benchmarked against the original Python implementation on multicore workstations first, and finally on a cluster.

Lidia Stepanova – Reduced order models for the simulation of pathological heart valves

Supervisor: Dr. Simone Deparis and Dr. Toni Lassila

Chair of Modelling and Scientific Computing – CMCS

The human heart contains four biological valves (mitral, aortic, pulmonary, and tricuspid) that regulate the flow in the atria and the ventricles. Malfunction of one of these valves, either by stenosis (stiffening resulting in a inability to open properly) or by regurgitation (leakage or flow reversal due to an inability to close properly), is a relatively common condition affecting the function of the heart and potentially leading to the onset of heart failure. There is much interest in the modelling and simulation of heart valves. especially pathological ones, in order to predict and prescribe possible surgical therapies in patient-specific cases. In the case that insufficient data and/or modelling capability is available to fully capture the complex 3D interaction between the valves and the blood flow through the heart, we want to capture the general behavior of the valve in the sense of mean flow, intraventricular pressure, and other clinically relevant variables. To this end many works have been devoted to deriving lumped parameter models for heart valves. They do not model the entire 3D geometry and fluid-structure interaction of the valve, but rather work based on simplified fluid dynamics principles and integrated quantities of velocity and pressure across the valve surface. The objectives of this project are to first perform an overview of existing reduced order models for the heart valves, and then to choose a suitable subset of models to prototype in MatLab and further to implement in LifeV library. Comparisons between predictions given by different reduced models for valves should be made first between artificial test cases, and finally on real patient data obtained as part of a project related to the simulation of pathological left ventricles with regurgitant mitral valves.

Laurent Fasnacht – Verifying equivalence of Python programs to their C/C++ counterparts

Supervisor: Prof. George Candea

Dependable Systems Laboratory – DSLAB

When developing HPC software, it is common to first write a prototype in a high level language (MATLAB, Python, R, …), to ensure the correctness of the algorithm and have a first usable implementation. However, these programming languages, while allowing quick code writing, don’t allow the same performance as low-level languages, like C or C++. Therefore, developers usually reimplement the algorithms in a more efficient language, and then make some checks to give themselves some confidence that both implementations give the same results. As writing good tests is difficult, it would be very useful to have tools that are able to prove equivalence between both implementations (check that for all possible inputs, outputs are equivalent).

This project focuses on proving equivalence between programs in Python and C. The main idea is to write a tool chain based on selected existing tools to convert both languages to a common representation, on which it becomes possible to apply automated reasoning techniques to formally prove equivalence. As the result of the proof depends on the quality of the tools doing the conversion, they have to be checked in depth. The tool chain will then be integrated in a framework that enables developers to easily verify their code.

Andrea Di Blasio – Numerical simulation of blood solutes by Isogeometric Analysis

Supervisor: Dr. Luca Dede’

Chair de Modelling and Scientific Computing – CMCS

Solutes and drugs are transported in the circulatory system by the blood for which absorption processes occur at the arterial walls. The comprehension, modeling, and simulation of these phenomena, both in physiological and pathological conditions, represents a relevant topic of interest in biomedical applications. Different mathematical models can be considered by coupling the Navier-Stokes equations representing the blood flow with the advection-diffusion equations describing the transport of the solutes and eventually the diffusion processes in the arterial wall, thus defining heterogeneous coupled models.
The project focus on the numerical approximation by means of Isogeometric Analysis of the Navier-Stokes equations coupled with the advection-diffusion models for the dynamics of the solutes in the blood and in the arterial walls. Firstly, the case of steady problems should be considered. Then, unsteady problems could be solved by using suitable numerical scheme for the approximation in time of the coupled problem. In both the cases, two-dimensional problems can be studied.

Elena Queirolo – Numerical methods for trajectory optimisation

Supervisor: Prof. Assyr Abdulle and Martin Huber

Chaire d’analyse numérique et mathématiques computationnelles – ANMC

The optimal design of the trajectory of a spacecraft is an important problem in aeronautics. For example, the task could be to find the best trajectory when the destination and the maximal fuel consumption are given. The aim of the project is to study a numerical method for such trajectory optimization modeled as an optimal control problem. First, we derive the first order optimality conditions of the optimal control problem and reformulate them as a constrained Hamiltonian system with two-point boundary conditions. Then, we use symplectic partitioned Runge-Kutta methods for the discretization in time and analyze their properties. Finally, we obtain the numerical method by combining these Runge-Kutta methods with a multiple shooting algorithm. We illustrate the capabilities of the numerical method by solving a model problem.

Philippe De Gol – Stochastic dynamics of ice calving

Supervisor: Prof. Paolo Perona and Dr. Benoît Crouzy

Group AHEAD (Applied HydroEconomics and Alpine environmental Dynamics)

Ice calving is the mechanism with which the advancing front of outlet glaciers is lost as a consequence of deep fracture (see Figure). The result of this process is that the glacier front advances and sometimes shortens (calves). Eventually, At short time scales the current position of the front seems to be statistically fluctuating around a mean position dictated on longer time scales by the either glacial global advancement or retreat phases (see Figure, from Post 1985). Many deterministic physically based models exist to predict when calving occurs and the related ice mass amount.
In this work, we look at the dynamics of the front with the idea of ascribing it to a stochastic process made of drifts (front advancement) and jumps (calving). We intend to study the stochastic dynamic of this process analytically, and the related statistics concerning the front position, the size of the calves, related intertimes, etc. Eventually, data from satellite images will be used to test the model. An ambitious student with strong mathematical skills and passion for glacial zone processes is therefore expected to choose this topic.

Loïc Perruchoud – Tracking leg movements in high-speed videos of insect locomotion

Biomedical Imaging Group – BIG

Supervisors: Prof. Michael Unser, Dr. Cédric Vonesch, Pavan Ramdya

In order to study locomotion in Drosophila, one must be able to quantify with high precision their walking behaviours. Therefore, the goal of this project is to build a computer vision software that extracts information on the positions and orientations of various leg segments from video input of Drosophila in an automated way. This work was separated into two major parts.
The first part reported the tracking of the body of a fly. This task was completed using an active snake. We first introduce the basics of active snakes and presented how the evolution of the curve define by the snake can be formulated as an optimization problem. Then, we introduced various fly models that we used. We finally showed how it is possible to track the body of a fly using active snake with a shape regularization energy term. The tracking algorithm showed good performances and robustness.
In the second part, we have been able to extend the optimization procedure used to track the body to the legs. We defined a parametric model of legs that can be attached to the body as a function of the active snake tracking the body. The tracking problem was formulated as an optimization problem thanks to an energy term based on the response of the fly to a steerable ridge filter. We showed that by adding two geometric constraint energy terms, we were able to obtain promising results for the tracking of the legs of the fly.

Jérémie Despraz – Combining UV Heating and Cooling in the Code GEAR

Laboratory of Astrophysics – LASTRO

Supervisors: Prof. Georges Meyan, Dr. Yves Révaz

The focus of this project is the study and implementation of a new cooling and heating function in the cosmological code GEAR, taking into account cosmic background radiation and gas metallicity.