Announcement Detail
Thursday, December 11, 2025
3pm EST
Join via Zoom: https://us06web.zoom.us/j/
USACM UQ Virtual Seminar
Structure-exploiting data-driven methods for large-scale uncertainty quantification
Speaker
Ionut-Gabriel Farcas, Virginia Tech
Abstract
Advances in numerical algorithms and computing enable the simulation of complex, real-world problems with ever-increasing realism and accuracy.
However, these simulations remain computationally expensive even on large supercomputers, which prevents their routine use in many-query tasks like uncertainty quantification (UQ) and sensitivity analysis (SA), which require ensembles of simulations.
In this presentation, we will show how structure-exploiting data-driven learning can enable efficient UQ and SA in large-scale, real-world problems.
In the first part, we will discuss a sensitivity-driven dimension-adaptive sparse grid interpolation approach. This method explores and exploits, via adaptive refinement, the fact that in many real-world problems (i) only a subset of the uncertain parameters are important and (ii) these parameters interact anisotropically. The goal is to preferentially refine the directions associated with the important inputs and their interactions. The power and usefulness of this approach is demonstrated in a realistic description of turbulent transport in the edge of fusion experiments. In a scenario with more than 264 million degrees of freedom and eight uncertain inputs, our approach requires a mere total of 57 high-fidelity simulations.
For certain simulation scenarios, relying on a single model can be insufficient. In the second part, we discuss a context-aware multi-fidelity Monte Carlo sampling algorithm in which a hierarchy - instead of a single model - of high- and low-fidelity models is used to perform UQ. In particular, we consider data-driven low-fidelity models and address the question of how much training data is required to train them for multi-fidelity Monte Carlo sampling. This is relevant for data-driven models which, in single-fidelity settings, can require large training sets. In contrast, in our context-aware algorithm, we trade off increasing the size of training sets with using the low-fidelity models for multi-fidelity sampling, meaning that small training sets can be sufficient. Crucially, in this framework, low-fidelity models can become too accurate for multi-fidelity methods, which is in stark contrast with standard surrogate modeling where more training data usually implies greater accuracy. Numerical experiments in a plasma micro-turbulence simulation scenario with 12 uncertain inputs show speedups of up to two orders of magnitude compared to standard methods, which corresponds to a runtime reduction from 72 days to about four hours on 32 cores on parallel machines.
Upcoming webinar speakers:
Thursday, January 8: Thomas Swinburn, University of Michigan
Monday, February 8: Wei Chen, Northwestern University
Thursday, March 12: Hannah Lu, The University of Texas at Austin
Thursday, April 16: Jinlong Wu, University of Wisconsin-Madison