Project description
Surrogates or reduced models for real-world systems are core to plenty of tasks in science and engineering. They enable rapid simulation for uncertainty quantification, sub-components in multi-scale simulations and approximations in pervasive settings. Our project will tackle some of the pressing open questions that did arise in recent projects within SimTech and advance the construction of fast, reliable surrogates. We aim to build machine learning (ML) models based on experimental and simulation data. They have to fulfil the special requirements of surrogate modelling which differ from standard learning tasks. As a second focus, and together with our partner project PN 5-7, we will bridge the gap between data-scarce and data-rich regimes for the construction of surrogate models of complex mechanical systems. While machine learning typically relies on abundant data to train on, both experimental and simulation data are usually expensive to obtain and thus scarce. This project will study and enable ML-based surrogates by three ML techniques (sparse grids, greedy kernel methods, neural networks) in simulation contexts. On the one hand, we will reduce the demand for data by physics-enforcing penalization, which additionally is core to ensure the suitability of the learned models as surrogates. On the other hand, we will develop techniques to augment scarce experimental data exploiting the set of physics-based reduced models developed in PN 5-7, adapting ideas from ensemble learning and multi-fidelity modeling.
Project information
Project title | Surrogate modelling by simulation-enhanced machine-learning |
Project leaders | Dirk Pflüger (Bernard Haasdonk) |
Project duration | January 2019 - June 2022 |
Project number | PN 6-2 |
- Follow-up project 6-2 (II)
Towards parameter-dependent data-enriched physics-informed machine learning
Publications PN 6-2 and PN 6-2 (II)
2022
- R. Leiteritz, P. Buchfink, B. Haasdonk, and D. Pflüger, “Surrogate-data-enriched Physics-Aware Neural Networks,” in Proceedings of the Northern Lights Deep Learning Workshop 2022, in Proceedings of the Northern Lights Deep Learning Workshop 2022, vol. 3. Mar. 2022. doi: 10.7557/18.6268.