Overview

Project 1 Evaluating Model Poisoning Attacks in Federated Machine Learning
Project 2 Simulating the genomic consequences of long and short term balancing selection by heterozygote advantage
Project 3 Implementing a simulation of the phenotypic infinitesimal model in reproducible open source software
Project 4 Computational modeling of avascular tumors: exploring the hallmarks of cancer in silico
Project 5 Incorporating learned preconditioners in an open-source FEM solver
Project 6 Learning preconditioners for interior point methods
Project 7 Neural representation of massive 3D data
Project 8 Can you hear the components of a network?
Project 9 Architectured materials with tailored energy dissipation: Design with machine learning
Project 10 High-dimensional option pricing using dimension reduction
Project 11 Quasi-Optimal Entanglement Characterisation in Quantum States
Project 12 A Hybrid Finite Difference – Finite Element Method for Wave Equations
Project 13 Filtering and identification in Epidemics
Project 14 Analysis of weakly labeled digital cytology data
Project 15 Evaluating resilience to heat stress among dairy cows
Project 16 Detecting the sequential use of resting beds by cows after milking
Project 17 Detecting the location of cows before entering the milking parlor3
Project 18 Detecting spikes in positioning data applied to cow movement
Project 19 Machine Learning Accelerated Solution of Inverse Problems in Photonics
Project 20 Modelling the physical and chemical behaviour of magma and associated geothermal systems in COMSOL
Project 21 Modelling the propagation of fractures, such as magmatic dykes and hydrofractures - X-FEM
Project 22 Efficient solver for the Hermite element mass matrix
Project 23 Network usage at home: the electricity footprint
Project 24 Deep learning based image segmentation for enhanced seed phenotyping

Project 1 Evaluating Model Poisoning Attacks in Federated Machine Learning

Usama Zafar usama.zafar@it.uu.se

Salman Toor salman.toor@it.uu.se

In order to design robust defensive mechanisms, it is imperative to first fully understand the impact of different types of attacks under federated machine learning architecture. Hence, under this project the primary aims is to implement several state-of-the-art model poisoning attacks in FedML and evaluate their performance.

Read the full description: here.

Project 2 Simulating the genomic consequences of long and short term balancing selection by heterozygote advantage

Martin Johnsson martin.johnsson@slu.se

Balancing selection by heterozygote advantage occurs when a genetic variant has both positive and negative effects on selected traits, so that on average, the heterozygote genotype has the highest fitness. Such variants are a major problem in animal breeding programs, where monogenic genetic defects can rise to high frequency due to heterozygote advantage (see reviews by Hedrick 2015, Derks & Steensma 2021). This raises questions about how to efficiently detect and manage such damaging genetic variants.

In this project, you will use individual-based forward-in-time-simulation to investigate the genomic signature of balancing selection by heterozygote advantage on a long and short time scale, at different levels of strength of selection and recombination rate. You can work either with genetic program SLIM3 or with custom Python or R code. Results may also be compared with classical equation-based models.

Hedrick PW. (2015) Heterozygote Advantage: The Effect of Artificial Selection in Livestock and Pets, Journal of Heredity, https://doi.org/10.1093/jhered/esu070

Derks MFL and Steensma M (2021) Review: Balancing Selection for Deleterious Alleles in Livestock. Frontiers in Genetics, https://doi.org/10.3389/fgene.2021.761728

SLIM3 website: https://messerlab.org/slim/

Project 3 Implementing a simulation of the phenotypic infinitesimal model in reproducible open source software

Martin Johnsson martin.johnsson@slu.se

The infinitesimal model is the standard model of quantitative genetics, that underpins classical theory as well as the practice of genetic evaluation in animal breeding. Simulations of the infinitesimal model, in slightly different forms, were common in the literature but have largey been replaced by genomically explicit finite-locus models. This is to the detriment of both pedagogy and theory-building in genetics. Barton et al. (2017) relatively recently formalized the phenotypic infinitesimal model, and discussed how it can be simulated.

This project aims to implement reproducible simulation of the phenotypic infinitesimal model, with an eye to eventually including the model in the simsimqt R package.

Barton NH, Etheridge AM, Véber A. (2017) The infinitesimal model: Definition, derivation, and implications. Theoretical Population Biology. https://doi.org/10.1016/j.tpb.2017.06.001

Project 4 Computational modeling of avascular tumors: exploring the hallmarks of cancer in silico

Stefan Engblom stefane@it.uu.se

Erik Blom erik.blom@it.uu.se

In this project we wish to extend a pre-existing tumor model to investigate two mechanisms closely related to the hallmarks of cancer and study their impact on tumor malignancy:

The Warburg effect: tumor cells begin to produce energy from the less efficient process of glycolysis even when oxygen is abundant. Does the addition of this effect significantly alter the aggressiveness and the stability of the tumor growth?
Evolution of cellular adhesion: healthy cells that adhere to other cells in coordination help sustain the structure of tissues and organs in the body. A reduction in cell-cell adhesion in tumor cells can therefore lead to more aggressive growth and metastasis. If we allow the adhesion property to mutate across cell generations, will an initially benign tumor ultimately become a malignant one?

Read the full description: here.

Project 5 Incorporating learned preconditioners in an open-source FEM solver

Paul Häusner paul.hausner@it.uu.se

Jens Sjölund jens.sjolund@it.uu.se

Recently, we proposed data-driven preconditioners, based on graph neural networks, that have shown promising results without requiring time-intensive hand-engineering. In this project, the goal is to extend the FEniCSx library and the underlying DolfinX solver to support learned preconditioners which enables practitioners to efficiently use the data-driven optimization method out of the box.

Read the full description: here.

Project 6 Learning preconditioners for interior point methods

Paul Häusner paul.hausner@it.uu.se

Jens Sjölund jens.sjolund@it.uu.se

Recent advances in data-driven optimization aim to replace hand-engineered preconditioners with parameterized functions that can be trained against data. The aim of this project is to extend this setting of learned preconditioners to interior-point methods

Read the full description: here.

Project 7 Neural representation of massive 3D data

Jens Sjölund jens.sjolund@it.uu.se

The idea of this project is to explore whether so-called neural representations can compress massive 3D data from X-ray microtomography to a more manageable size, while retaining the features most pertinent for downstream tasks.

Read the full description: here.

Project 8 Can you hear the components of a network?

Jörn Zimmerling jorn.zimmerling@it.uu.se

In this project, we will build an experimental setup and implement an algorithm to solve an inverse eigenvalue problem. In an inverse eigenvalue problem, you are given the eigenvalues (and some information about the eigenvectors) and you reconstruct the matrix (with some prior knowledge on its structure).

Read the full description: here.

Project 9 Architectured materials with tailored energy dissipation: Design with machine learning

Mahmoud Mousavi mahmoud.mousavi@angstrom.uu.se

Aliae Welander aliae.welander@hitachienergy.com

Material design based on micro-architectures delivers a promising tool to replace the traditionally homogeneous materials with tailored heterogeneity, and it finds many applications in industry for weight-saving, impact energy absorption, load carrying, vibroacoustic control, and heat exchanging. Thanks to the progress of additive manufacturing along with the advances in computational engineering, the mechanical properties of the so-called architectured materials is possible to be captured. The inverse problem of identifying the tailored topology which can deliver a desired property, however, is still a challenge. Recent advances in machine learning and deep learning seems to provide a tool for overcoming this challenge. In this project, computational homogenization and machine learning is employed to address the inverse problem of identifying the lattice materials with a given energy dissipation profile

Read the full description: here.

Project 10 High-dimensional option pricing using dimension reduction

Elisabeth Larsson elisabeth.larsson@it.uu.se

An option pricing problem can be described either in terms of one or more stochastic diffusion processes for the underlying assets, and potentially additional processes for the volatility of the assets and or the interest rate, or as a PDE or PIDE (partial integro-differential equation) for the option price. The corresponding numerical methods are in the first case Monte Carlo methods and in the second any numerical PDE solver. What determines the optimal choice is mainly the number of underlying assets or processes, which in the PDE case corresponds to the number of dimensions. Monte Carlo methods are the most efficient for high-dimensional problems, while standard PDE methods are competitive for lower dimensions up to around five dimensions. Last year, a group of four students made first study on how to use a combination of a decomposition into lower-dimensional parts and meshfree kernel-based approximation to solve high-dimensional option pricing problems. This project is a continuation of this work and will in particular focus on the effects of subtracting a particular solution that captures the non-smooth initial condition of the PDE problem.

Read the full description: here.

Project 11 Quasi-Optimal Entanglement Characterisation in Quantum States

Prashant Singh prashant.singh@scilifelab.uu.se

Vahid Azimi Mouslou vahid.azimi-mousolou@physics.uu.se

Per Mattsson per.mattsson@it.uu.se

Quantum entanglement serves as a fundamental resource for various quantum information and communication protocols, including quantum teleportation, superdense coding, and quantum error correction. Entangled qubits exhibit the remarkable ability to surpass classical computers with exponential speedup, underscoring their crucial role in universal quantum algorithms. While quantum entanglement is an experimentally well-established phenomenon, its precise and reliable characterization remains an ongoing and challenging problem in the field of quantum science and technology. Within the scope of the project, the task will be to formulate the entanglement characterisation problem as a global optimisation problem. We will implement the objective function in Python, and the optimisation routines using PyTorch or Tensorflow.

Read the full description: here.

Project 12 A Hybrid Finite Difference – Finite Element Method for Wave Equations

Martin Almquist martin.almquist@it.uu.se

Vidar Stiernström vidar.stiernstrom@it.uu.se

High-order finite difference (FD) methods are computationally efficient for wave equations posed on moderately complex domains that can be meshed by structured grids. Finite element (FE) methods, on the other hand, offer unstructured meshing capabilities and can handle very complex domains, but are typically less efficient than FD methods because they come with a non-diagonal mass matrix that causes a linear solve in every time step. A hybrid FD-FE method has the potential to combine the computational efficiency of finite differences with the geometric flexibility of finite elements.

Read the full description: here.

Project 13 Filtering and identification in Epidemics

Stefan Engblom stefane@it.uu.se

In this project we are interested in designing data-driven epidemiological models which are effective in producing not only forecasts, but also in accurate "nowcasts" in order to provide for situation awareness during decision making.

Implement a convincing metric by which different filtering- and prediction algorithms can be suitably evaluated, specifically so under an epidemiological perspective involving hazardous situations.
Experiment with different identification algorithms and evaluate the qualities and trade-offs they have. -Can we offer any recommendations to the authorities?
Implement a few selected improved algorithms and/or produce optimal parameters and set-ups for existing ones.
Challenge. Given data for COVID-19, and a suitable computational set-up, can we perform dynamic parameter identification as well as for- and nowcasting for all of Sweden's 21 regions using 2 years of data in under 1 minute on a laptop? What is the best achievable quality of this output?

Read the full description: here.

Project 14 Analysis of weakly labeled digital cytology data

Nadezhda Koriakina nadezhda.koriakina@it.uu.se

Joakim Lindblad joakim.lindblad@it.uu.se

Natasa Sladoje natasa.sladoje@it.uu.se

In the presence of a tumor or during routine screening for abnormalities, samples from patients are classified as benign or malignant. This is performed either by cytological or histological examination. In cytology, a cytotechnologist traverses the sample to find suspicious cells. A digital image of a sample, often referred to as a whole slide image (WSI), can contain more than a hundred thousand cells. The task of finding a rather small portion of abnormal cells within many normal cells is extremely laborious and time-consuming. The goal of the project is to investigate whether weakly supervised CNN-based method(s) are able to detect lesions (and possibly additional patterns) in weakly labeled DCCL dataset.

Read the full description: here.

Project 15 Evaluating resilience to heat stress among dairy cows

Tomas Klingström tomas.klingstrom@slu.se

The purpose of the project is to compare weather data measurements from the Swedish Meterological and Hydrological Institute (SMHI) with milking data collected from the SLU Infrastructure for dairy data collection, Gigacow. This project is a step towards enabling researchers to integrate farm-level data from herd management systems and genome analysis collected by SLU Gigacow with weather data to conduct long term studies on how weather conditions affect animal health and milk production.

Objectives for the project are to:

Compare milk yield of dairy cows with external weather factors.
Use signal processing to eliminate confounding factors when measuring individual resilience to heat stress.

The expected outcome of the project is a to measure the impact of weather on the milk yield of dairy cattle.

Read the full description: here.

Project 16 Detecting the sequential use of resting beds by cows after milking

Lars Rönnegård lrn@du.se

The specific goal of this project is to identify where the cows lie down after being milked and how the resting beds (cubicles) are used by different cows. We have access to 3 years of positioning data on 200 dairy cows inside a free-stall farm whose positions have been collected every second.

Read the full description: here.

Project 17 Detecting the location of cows before entering the milking parlor

Lars Rönnegård lrn@du.se

The specific goal of this project is to identify where the cows are located prior to entering the milking parlor and in which order they leave and re-enter the group before and after milking. We have access to 3 years of positioning data on 200 dairy cows inside a free-stall farm whose positions have been collected every second.

Read the full description: here.

Project 18 Detecting spikes in positioning data applied to cow movement

Lars Rönnegård lrn@du.se

The specific goal of this project is to identify erroneous spikes in positioning data. We have access to 3 years of positioning data on 200 dairy cows inside a freestall farm whose positions have been collected every second. On some occasions, the signals are poor and the positions seem to fly away in spikes. The aim is to develop software that can automatically find these spikes in the data.

Read the full description: here.

Project 19 Machine Learning Accelerated Solution of Inverse Problems in Photonics

Prashant Singh prashant.singh@scilifelab.uu.se

Francesco Ferranti francesco.ferranti@vub.be

Inverse problems are very often encountered in design optimization workflows in photonics. Given design parameters and responses, the design optimization process involves inferring specific values of the design parameters satisfying desired/optimal values of the response. We will design an inverse (machine learning) generative model that can learn the mapping between various response behaviors, and the corresponding (possibly disjoint) parameter value locations in the design space.

Read the full description: here.

Project 20 Modelling the physical and chemical behaviour of magma and associated geothermal systems in COMSOL

Steffi Burchardt steffi.burchardt@geo.uu.se

Christoph Hieronymus christoph.hieronymus@geo.uu.se

Reactive flow in a porous medium, plus phase changes of the fluid. Both of these issues
could be tackled separately or in a combined way. For the reactive flow, components could be precipitating
out of the fluid, or the fluid could dissolve parts of the matrix. For phase changes, the correct material properties
(density, heat capacity, volumetric expansion, etc) for water are already implemented, but what is missing is
a) to account for any latent heats associated with the change (if relevant…), and b) that there
is a volume change which is commensurate with the density change. Results need to be tested /
benchmarked.

Read the full description: here.

Project 21 Modelling the propagation of fractures, such as magmatic dykes and hydrofractures - X-FEM

Steffi Burchardt steffi.burchardt@geo.uu.se

Christoph Hieronymus christoph.hieronymus@geo.uu.se

This is a method which allows the splitting of elements by introducing new numerical
nodes. With the appropriate physics, one can model fracture propagation with new fractures
forming in arbitrary directions and locations, rather than being limited to existing element boundaries.
X-FEM has already been developed for other FEM codes, and it has also been incorporated into
Comsol by others. However, it is not very user-friendly. It would be extremely useful in Earth sciences
to be able to use this tool (or perhaps a more user-friendly version of it).

Read the full description: here.

Project 22 Efficient solver for the Hermite element mass matrix

Gunilla Kreiss gunilla.kreiss@it.uu.se

When designing algorithms for solving PDE’s numerically an important component is often finding an appropriate efficient numerical solution method for the arising linear systems. Nowadays these systems are usually of very large dimension and pose the requirement to use iterative solution methods which are less demanding with respect to computer resources, compared to the direct solution methods, based on the Gauss elimination framework.

In the prevailing cases, the matrices in the arising linear systems are sparse and have a structure which depends on the discretization of the PDE and therefore it is usually beneficial to tailor the linear solver to the specific structure. In this project we consider the mass-matrix corresponding to a novel finite element method, the Hermite element method, and compare the efficiency of some iterative schemes with preconditioning with some direct methods.

For the standard FEM it is rather well-explored, for instance, in the so-called domain decomposition aproach. For the Hermite element method the structure of the matrix is different, and it will be insightful to investigate how this will influence the comparison between the different methods.

In the Hermite element method the approximation is based on Hermite interpolation. In a quadrilateral grid with rectangular elements, the degrees of freedom are related to function values and to values of derivatives at the corners of the elements. These elements have been included in the DEAL2 computational framework, and it is therefor not difficult to generate matrices corresponding to problems in one, two or three space dimensions, and of different sizes, to be studied for instance using Matlab.

This project aims to study the matrix structures, arising in Hermit polynomial basis functions Finite element setting and how to use these structures when solving the corresponding linear systems using some iteration methods.

The suggested framework is Matlab. The matrices will be provided and we will use the iterative solution methods from the Matlab library.

Project 23 Network usage at home: the electricity footprint

Mike Hazas mike.hazas@it.uu.se

Martin Stojanov martin.stojanov@it.uu.se

Usage of Internet-connected services and devices causes about 9% of total global electricity demand. We have regional estimates, for example in Europe video streaming is about 70% of all Internet traffic. However, we know much less about what services are used and when in homes. We are developing a methodology (link below) to quickly gather 24-hour samples of Internet usage from homes using NetFlow. It logs the IP endpoints, time and duration of the connection, and total data sent/received.

Usage of a given service (e.g. Netflix or Facebook) involves many different IP endpoints (advertising, tracking, analytics, and metadata/control channels). This project will work with a NetFlow dataset, to automatically group and classify the connections logged. Subclassification would also be of interest, for example "1 GB was attributable to a particular social networking service in the 24 hour sample, of which 400 MB was video and 200 MB was advertising."

https://www.it.uu.se/research/human-machine-interaction/digital-ecologies-lab/internet-microscope

For a description of the issues and challenges of attributing (based on reverse DNS lookup) and grouping NetFlow logs, have a look at sections 3.2.1.2 and 3.2.2 of this PhD thesis by Kelly Widdicks:
https://eprints.lancs.ac.uk/id/eprint/143606/1/2019widdicksphd.pdf

For an example of how well-analysed logs can lead to research outcomes, have a look at the paper by Widdicks et al. presented at the CHI conference in 2019:
https://doi.org/10.1145/3290605.3300696

Project 24 Deep learning based image segmentation for enhanced seed phenotyping

Filip Thor filip.thor@it.uu.se

Carl Nettelblad carl.nettelblad@it.uu.se

An important part of plant breeding processes is the evaluation of the crop’s attributes, or determining its phenotypes. In this project, we are interested in phenotyping oat seeds. Some key phenotypes that are important for oats are, for example, protein content, yield, thousand kernel weight, and seed morphology (length, width, circularity, etc.). The latter of these properties can be evaluated by simple camera photography and using conventional computer vision tools. The overall goal of this project is the development of a software solution where we can extract the necessary seed properties (number of seeds, shape, size, color, etc), without the hassle of manually ordering each batch.

Read the full description: here.