Project 02: Advanced Parallelization Techniques for datadriven Uncertainty Quantification

Autor

Sebastian Krumscheid, Linus Seelinger

Herausgeber

Uwe Ehret, Martin Frank, KIT-Zentrum MathSEE

Beteiligtes Institut

KIT-Zentrum Mathematik in den Natur-, Ingenieur- und Wirtschaftswissenschaften (KIT-Zentrum MathSEE)
Steinbuch Centre for Computing (SCC)

Genre

Veranstaltung

Beschreibung

02 Advanced Parallelization Techniques for data-driven Uncertainty
Quantification
MATH PI: TT-Prof. Dr. Sebastian Krumscheid, Steinbuch Centre for Computing (SCC), Junior Research
Group Uncertainty Quantification (SCC-UQ) & Institute for Applied and Numerical Mathematics
(IANM)
SEE PI: Dr. Linus Seelinger, YIG Prep Pro Fellow
Department(s): Mathematics
Type of position: 75% FTE, E13 TV-L
In the field of natural sciences, the integration of Bayesian inference and Uncertainty Quantification
(UQ) shows great potential in improving our understanding of complex systems and enhancing the
reliability of scientific predictions. However, UQ often faces challenges as it requires a large number of
simulation runs, which necessitates advanced UQ algorithms and high-performance computing (HPC).
Unfortunately, intricate data dependencies and technical obstacles hinder parallel computing for
inverse (i.e., data-driven) UQ. To address this problem, UM-Bridge introduces a novel software
architecture to UQ, which solves these technical challenges and lays the foundation for developing
sophisticated parallelization strategies at a relatively low compute cost. Additionally, it allows for the
seamless application of these strategies to various scientific models.
Consequently, UM-Bridge enables comprehensive study of advanced parallelization techniques for UQ
with application to realistic large-scale models. The objectives of this project are twofold. Firstly, we
aim to develop new parallelization methods for inverse UQ that will enable efficient Bayesian inversion
on modern HPC systems. We will consider approaches such as hierarchical Markov chain Monte Carlo
(MCMC) methods combined with sample pooling techniques obtained from fast, yet approximate
surrogate models and parallel, multiple-try MCMC proposal methods. Secondly, we will adapt these
new methods for use in Earth System Sciences, specifically for data-driven modeling in computeintensive Earth system models. This will be done collaboratively in the context of Simulation and Data
Life Cycle Labs at the Steinbuch Centre for Computing.
Requirements for this position:
- Solid knowledge of probability theory and statistics, as well as numerical analysis and
- mathematical modelling.
- Experience with uncertainty quantification techniques is advantageous.
- Good programming skills, including experience with parallel programming.

Laufzeit (hh:mm:ss)

00:05:53

Serie

KCDS Virtual Open House 2023 - Fall

Publiziert am

23.10.2023

Fachgebiet

Informatik

Lizenz

KITopen-Lizenz