E-Thesis 132 views
Data-driven reduced order modelling based on deep learning / RUI FU
Swansea University Author: RUI FU
DOI (Published version): 10.23889/SUThesis.67587
Abstract
Reduced Order Modelling (ROM) has become a powerful tool in computational simulations, especially for complex fluid dynamics. ROM-based models (ROMs) offer significant reductions in computational cost while maintaining acceptable accuracy. These models rely on dimensional reduction to simplify complex...
| Published: |
Swansea University, Wales, UK
2024
|
|---|---|
| Institution: | Swansea University |
| Degree level: | Doctoral |
| Degree name: | Ph.D |
| Supervisor: | Feng, Y. |
| URI: | https://cronfa.swan.ac.uk/Record/cronfa67587 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Abstract: |
Reduced Order Modelling (ROM) has become a powerful tool in computational simulations, especially for complex fluid dynamics. ROM-based models (ROMs) offer significant reductions in computational cost while maintaining acceptable accuracy. These models rely on dimensional reduction to simplify complex dynamic systems into lower-dimensional representations. It is essential for large simulations and real-time analysis when computational efficiency is crucial. The Parametric Reduced Order Models (PROMs) extend the capabilities of traditional ROMs by considering changes in physical parameters or system configurations that affect the variability of dynamic system responses. This parametric feature enables PROMs to adapt to various dynamic environments, enhancing their flexibility and applicability.In recent years, deep learning has performed excellently in handling large datasets and complex systems. Integrating deep learning methods into ROMs and PROMs has opened up new areas for improving the performance and applicability of ROMs and PROMs. With deep learning methods, ROMs are significantly enhanced in robust capabilities for feature extraction, pattern recognition, and the ability to learn complex relationships in data. For PROMs, deep learning helps achieve higher accuracy, better generalisation and faster adaptation to new conditions. In fields such as fluid dynamics and climate modelling, dynamic systems always have significant challenges with their field complexity and high dimensionality. The ROMs/PROMs with deep learning applied in these fields not only enable more efficient computations for simulations but also the extraction of deeper insights from simulation data, enabling improvements in predictive modelling and decision-making. This thesis presents several advancements in the field of Reduced Order Modeling (ROM)for fluid flow problems. The primary contributions are threefold:1. Non-Linear Non-Intrusive ROM(NL-NIROM): A novel nonlinear non-intrusive ROM(NL-NIROM) is introduced for fluid dynamic field time-series prediction, which is based on an Autoencoder network and the self-attention mechanism. Additionally, a linear non-intrusive reduced order model(L-NIROM) is also developed on proper orthogonal decomposition (POD) and self-attention mechanism. The advantage of this NL-NIROM is that it can capture more non-linearity information than POD-based L-NIROM.2. Parametric Non-Linear Non-Intrusive ROM (P-NLNIROM): A new parametric non-linear non-intrusive reduced-order model (P-NLNIROM) for the fluid flow system simulations on different parameter sets is presented. This model is constructed by the Stacked Auto-Encoder(DAE) and Deep Residual Learning Neural network(ResNet). In addition, transfer learning(TL) is employed to extend the prediction range within the parametric space, thereby greatly improving the predictive capabilities of the P-NLNIROM.3. Error Estimate Parametric ROM: Building upon the P-NLNIROM, an error estimate parametric ROM is implemented to correct the predicted simulations caused by P-NLNIROM for higher accuracy and reliability in simulation outputs. To evaluate the effectiveness of this ROM, a numerical experiment on the lock exchange scenarios is presented. It serves as an example case to demonstrate the performance and practical applicability of the error estimate parametric ROM with fluid dynamics simulations. |
|---|---|
| Item Description: |
A selection of content is redacted or is partially redacted from this thesis to protect sensitive and personal information. |
| Keywords: |
non-instrusive reduce order modelling, deep learning methods, computational fluid dynamics, N-S equations |
| College: |
Faculty of Science and Engineering |
| Funders: |
EPSRC grant: PURIFY (EP/V000756/1) |