Adaptive Learning and Prediction of Multiscale Dynamical Systems

Computer simulations are an indispensable tool for analyzing complex, multiscale systems in a wide range of scientific and engineering disciplines. In many such systems, the governing dynamics occur at the micro scale, whereas quantities of interest manifest themselves on orders of magnitude larger macro scale. To reduce the computational cost, the simulations are ideally performed directly on the macro and not the micro scale. However, finding an effective macro-scale model that reflects well its micro-scale counterpart remains a key challenge for Computational Science. In this thesis, we apply machine learning algorithms to automatically identify a system's macro-scale representation, effective macro-scale dynamics and the transfer of information between micro and macro scales. We present Adaptive Learning of Effective Dynamics (AdaLED), a framework for building online macro-scale surrogate models of complex multiscale systems based on autoencoders and ensembles of recurrent neural networks. The framework trains the surrogate on the fly and continuously monitors its prediction accuracy. Depending on the accuracy, it automatically switches the control of the simulation between the micro and macro scale dynamics to achieve a balance between speed-up and accuracy. Compared to other approaches, AdaLED enables for the first time the creation of adaptive surrogate models with online accuracy estimation that operate on the level of a time step. AdaLED models allow us to accelerate parts of dynamics that are simpler to learn while leaving the more complex ones to the micro-scale dynamics. The thesis further touches on the problem of coupling known micro and macro models and handling open boundary conditions for particle dynamics, used in particle-continuum coupling in multiscale fluid dynamics. We propose ideas for generalizing existing algorithms, limited to static cubical and spherical domains, to complex time-varying geometries. In particular, we show the generalization of the method for cross-boundary mass flow and discuss the challenges of generalizing the remaining boundary force method required to remove near-boundary density artifacts. The thesis also addresses the essential and often overlooked aspects of software design of coupled simulations based on multiple existing single-model solvers. We propose a novel software design that helps achieve portable, solver-agnostic implementations of the coupling model while minimizing the overhead of portability. The design presented in the thesis allows us to reuse the same coupling codes with multiple combinations of single-model solvers, thus increasing its usefulness and impact.

Grantor dc:publisher

ETH Zurich

Year dc:date

2022

Author dc:creator

• Kičić, Ivica

Contributors dc:contributor

• Koumoutsakos, Petros; id_orcid0000-0001-8337-2122
• Pezzè, Mauro
• Zavadlav, Julija

• machine learning; surrogate modeling; Software design; molecular dynamics; Coupled models
• info:eu-repo/classification/ddc/004
• info:eu-repo/classification/ddc/600
• Data processing, computer science
• Technology (applied sciences)

Statement dc:rights

• info:eu-repo/semantics/openAccess
• Creative Commons Attribution 4.0 International

Licence

• http://creativecommons.org/licenses/by/4.0/

Language dc:language

en

Identifier

https://doi.org/10.3929/ethz-b-000600553

OAI identifier oai:identifier

oai:www.research-collection.ethz.ch:20.500.11850/600553