Surrogate modelling using sparse polynomial chaos expansions: a machine learning flavour
Open access
Author
Date
2024-06-19Type
- Other Conference Item
ETH Bibliography
yes
Altmetrics
Abstract
Computational models have become an integral component across various domains of applied sciences and engineering. These models, often referred to as simulators, play a crucial role in predicting the behavior of complex natural or man-made systems. They empower engineers and scientists to assess a system's performance within a virtual environment, subsequently aiding in the optimization of its design and operation. Simulators, including high-fidelity finite element models, frequently encompass a multitude of parameters, and require large computational times, even when harnessing the full extent of available computing power. Moreover, the intricate nature of these systems leads to consider uncertainty in their governing parameters, environmental conditions, and operational factors. In this context, uncertainty quantification (UQ) methodologies have garnered widespread attention within both academic and industrial spheres. These methodologies offer solutions for addressing issues related to structural reliability, sensitivity analysis and optimal design. One well-known brute-force approach, Monte Carlo simulation, leverages random number generation to tackle these challenges. However, it often demands thousands to millions of simulations to attain accurate predictions, rendering it impractical for high-fidelity simulators. In contrast, surrogate models provide an effective means of resolving UQ problems by establishing a precise approximation of the simulator's response, utilizing a limited number of runs at strategically chosen values (referred to as the experimental design) and employing a learning algorithm. In this respect, surrogate modelling is nothing but ad-hoc supervised (machine) learning dedicated to the range of ``small data''. In this presentation, we will start by introducing the fundamental features of surrogate models and their connection with machine learning. Subsequently, we will delve into a comprehensive exploration of polynomial chaos expansions (PCE), including their sparse variant tailored for high-dimensional problems . Furthermore, we will address recent advancements of sparse PCEs in the context of structural dynamics . The talk will be illustrated with applications from geophysics, earthquake engineering and wind turbine design. Show more
Permanent link
https://doi.org/10.3929/ethz-b-000680984Publication status
publishedPublisher
ETH Zurich, Risk, Safety and Uncertainty QuantificationEvent
Subject
Uncertainty Quantification; Surrogate models; Sparse polynomial chaos expansionsOrganisational unit
03962 - Sudret, Bruno / Sudret, Bruno
Notes
Distinguished Lecture held on June 19, 2024.More
Show all metadata
ETH Bibliography
yes
Altmetrics