SFB 1313 Milestone Presentation by Tim Brünnette

January 7, 2025 / pa

SFB 1313 doctoral researcher at the Institute for Modelling Hydraulic and Environmental Systems (research project B04) | 7 January 2025 | 5 pm CET

Tim Brünnette, SFB 1313 doctoral researcher at the Institute for Modelling Hydraulic and Environmental Systems (research project B04), will give his milestone presentation on "Efficient Uncertainty Quantification for Non-Deterministic Engineering Processes" on 7 January 2025.

Date: Tuesday, 7 January 2025
Time: 5 pm CET
Title: "Efficient Uncertainty Quantification for Non-Deterministic Engineering Processes"
Location: Multi Media Lab (MML), U1.003, Pfaffenwaldring 61, 70569 Stuttgart, Campus Vaihingen

Abstract

Uncertainty quantification is an invaluable tool for relevant predictions. A lot of this uncertainty is epistemic, i.e. stemming from a lack of knowledge about model parameters, boundary conditions or similar. On top of this, many processes retain some inherent irreducible stochasticity, the so called aleatoric uncertainty. Multiple runs of a predictive model are at the core of almost all methods providing a quantification of uncertainty. These runs are designed to explore a range of uncertain inputs which are then translated into a range of possible outputs. If the aleatoric uncertainty has a relevant and complex effect on the process outcome, it means that even multiple runs of the same input are necessary to fully capture the uncertainty. For many models, this has a prohibitively high computational cost. We discuss ways to quantify the uncertainty in a more efficient way, reducing the number of expensive model runs.

To this end, we employ several techniques each tailored to their respective application. Our foremost example are fractures, whose random-appearing geometries are highly relevant for flow-driven applications in the underground, such as CO2 sequestration, contaminant transport or geothermal energy systems. We employ two main strategies: We can either circumvent the expensive model runs through the use of reduced models or surrogates predicting the quantities of interest. Alternatively, we can lift the hood and change how the stochasticity is implemented in the fracturing model, making the inherent uncertainty easier to track.

As further examples of applications, we will look at simulations of airplane crash debris distributions and of noise in quantum computers. For the former, the trick  for increased efficiency lies in sensibly reusing previously computed data, while for the latter a clever model choice enables partial analytical results.  

To the top of the page