46th SPEEDUP Workshop on High-Performance Computing
“Uncertainty Quantification and HPC”

University of Bern
August 31st - September 1st, 2017

Sponsored by: Speedup   UniBe   Foundations of Mathematics and Informatics in Computational Science

The 46th SpeedUp Workshop focusses on Uncertainty Quantification and HPC. The program consists of a hands-on tutorial on Thursday, August 30 and the actual workshop with invited speakers and posters on Friday, September 1, 2017.

The invited talks illustrate various aspects of Uncertainty Quantification (UQ) which combines probability theory and statistics with physical processes in the real world. UQ problems include prediction, model validation, parameter estimation, data assimilation and inverse problems. It is a field of growing interest and this workshop tries to illustrate this with invited talks from different fields of application of UQ. These talks will be complemented by presentations on mathematical and methodological aspects of UQ.


The 46th SpeedUp Workshop takes place at the UniS building of the University of Bern (Schanzeneckstrasse 1, 3012 Bern). UniS is conveniently located just next to the main train station in Bern.


The Tutorial (31. August 2017) is in room A201 (3rd floor)


the Workshop (1. September 2017) is in room A-126 (basement).

Lunch will be served in the UNIESS Bistro (mainfloor of the UniS building). Coffee breaks and poster display are in the lobby in front of room A-126.


PDF version

Tutorial on Thursday, 31 August 2017

09:00 - 12:00
Block A
Maximilian Koschade Technische Universität München
Hands-on Tutorial on Bayesian Multi-Level Monte Carlo
  • Introduction to Multi-Level Monte Carlo for Uncertainty Quantification
  • Hands-on session: predicting the statistics of an expensive stochastic PDE solver using lower-fidelity, cheaper solvers
  • General knowledge about the Python programming language
  • Laptop with a Unix/Linux operating system (for Windows user I recommend a Linux virtual machine)
  • Python3 with the following packages: to be announced (minimum requirement numpy and pymc3). I recommend using anaconda (as well as Jupyter notebook) and will provide an environment file from which automatically all dependencies can be installed
12:00 - 13:30
13:30 - 16:30
Block B
Marcel Schöngens CSCS, Zürich/Lugano
Hands-on Tutorial on ABCpy
  • ABCpy overview: philosophy, architecture, parallelism
  • Hands-on: Implementation of a simple example inference problem using ABCpy and Jupyter notebooks
  • Hands-on: Bring your own model calibration problem and solve it with ABCpy
  • General knowledge about the Python programming language
  • Laptop with a Unix/Linux operating system (for Windows user I recommend a Linux virtual machine)
  • Python3 with the following packages pre-installed (pip install): ipython, jupyter notebook, abcpy

Workshop on Friday, September 1st 2017

09:00 - 09:30
Registration and Coffee
09:30 - 09:45
09:45 - 10:30
Gernot Plank Technische Universität Graz
Personalized models of total heart function
Cardiovascular diseases are with a prevalence of 45% a major cause of death in the industrialized world. Despite steady therapeutic improvements, many cardiovascular diseases of high epidemiological relevance cannot be cured and are treated by mitigating symptoms. Image-based patient-specific models of cardiac function are a highly promising approach to comprehensively and quantitatively characterize cardiovascular function in a given patient. Such models are anticipated to play a pivotal role in future precision medicine as a method to stratify diseases, optimize therapeutic procedures, predict outcomes and thus better inform clinical decision making. Key challenges to be addressed are two-fold. Expensive computational models must be made efficient enough to be compatible with clinical time frames and generic models must be specialized based on clinical data which requires complex parameterization and data assimilation procedures.
10:30 - 11:00
Speed-poster presentation
11:00 - 11:30
Break and Poster Session
11:30 - 12:15
Phaedon-Stelios Koutsourelakis Technische Universität München
Physics-conversant machine learning: from molecular dynamics to stochastic PDEs
This talk is concerned with the development and adaptation of probabilistic machine learning strategies for the solution of various problems in physical modeling. It is consistent with the emergence of data-driven discovery, commonly referred to as the fourth paradigm in science, for extracting governing equations from data in cases where models (or closures) remain elusive.
12:15 - 13:45
13:45 - 14:15
General Assembly of the SpeedUp Society
14:15 - 15:00
Marcel Schöngens CSCS, Zürich/Lugano
ABCpy: Approximate Bayesian Computation at Scale
ABCpy is a highly modular, scientific library for Approximate Bayesian Computation (ABC) written in Python. In enables domain scientists to easily apply ABC to their research without being ABC experts; using ABCpy they can easily run large parallel simulations without much knowledge about parallelization, even without much additional effort to parallelize their code. Further, ABCpy enables ABC experts to easily develop new inference schemes and evaluate them in a standardized environment, and to extend the library with new algorithms. These benefits come mainly from the modularity of ABCpy. We give an overview of the design of ABCpy, and we provide a performance evaluation concentrating on parallelization.
15:00 - 15:30
Coffee Break
15:00 - 15:45
Andrea Barth Universität Stuttgart
Quantification of Uncertainty via Multilevel Monte Carlo Methods
Multilevel Monte Carlo methods were introduced to decrease the computational complexity of the calculation of, for instance, the expectation of a random quantity. More precisely, in comparison to standard Monte Carlo methods, the computational complexity is (asymptotically) equal to the calculation of one sample of the problem on the finest discretization grid used. The price to pay for this increase in efficiency is that the problem must be solved not only on one (fine) grid, but on a hierarchy of discretizations. This implies, first, that the solution has to be represented on all grids and, second, that the variance of the detail (the difference of approximate solutions on two consecutive grids) converges with the refinement of the grid. In this talk, I will give an introduction to multilevel Monte Carlo methods in the case when the variance of the detail does not converge uniformly. The idea is illustrated by the calculation of the expectation for an elliptic problem with a random (multiscale) coefficient and then extended to approximations of discontinuous nature, e.g. Poisson noise.
15:45 - 16:30
Oliver Fuhrer MeteoSwiss, Zürich
Mostly sunny or partly cloudy? HPC and UQ challenges in weather forecasting
Weather forecasts are inherently uncertain and techniques to quantify this uncertainty within the strict time-to-solution requirements is a considerable high-performance computing challenge. This talk will outline current practice in high-performance computing and uncertainty quantification for weather forecasting and highlight some of the current outstanding challenges.
16:30 - 17:00
Closing and Farewell Apéro

Call for Posters

The SpeedUp workshop is a great opportunity for young researchers to present their work to the HPC community. Posters on all aspects of High-Performance Computing are eligible. Please register your poster. All registered posters participate in the competition for the Best SpeedUp Poster Award 2017.


Online registration here.

Local Organizer

Dominik Obrist (University of Bern)

SPEEDUP Committee

Rolf Krause (USI Lugano), Peter Arbenz (ETH Zurich), Vittoria Rezzonico (EPF Lausanne), Helmar Burkhart (University of Basel), Dominik Obrist (University of Bern), Olaf Schenk (USI Lugano), Simone Deparis (EPF Lausanne), Andreas Adelmann (PSI Villigen), Ales Janka (HES-SO Fribourg), Bastien Chopard (University of Geneva), Joost VandeVondele (ETH Zurich), Jan Hesthaven (EPF Lausanne), Henrik Nordborg (HSR Rapperswil).