Geometrically parametrised reduced order models for studying the hysteresis of the Coanda effect in finite element-based incompressible fluid dynamics

Jose Raul Bravo, Giovanni Stabile, Markus Hess, J. A. Hernandez, Riccardo Rossi, Gianluigi RozzaJournal of Computational Physics 2024 · DOI

TL;DR

This paper introduces a novel framework for creating computationally efficient reduced-order models (ROMs) of fluid dynamics problems with changing geometries. By combining Proper Orthogonal Decomposition (POD) with Empirical Cubature Method (ECM) for hyper-reduction, they can accurately simulate complex phenomena like the Coanda effect and its hysteresis, even with dynamic geometric changes, offering significant speedups over traditional methods.

What Problem Does This Paper Address?

Simulating fluid dynamics, especially with moving or deforming boundaries, is computationally expensive. This is particularly true for phenomena like the Coanda effect, where a fluid jet attaches to a surface, which can lead to complex, bifurcating solutions and hysteresis (where the system's behavior depends on its history). Traditional full-order models (FOMs) struggle to capture these dynamics efficiently, especially when the geometry changes over time. The challenge is to develop reduced-order models that can accurately represent these complex behaviors, including hysteresis, while offering substantial computational speedups, even when dealing with time-dependent geometric variations.

Methods

The authors propose a framework that integrates two key techniques:

  • Proper Orthogonal Decomposition (POD): This is a standard technique for dimensionality reduction. It identifies the most dominant modes (patterns) in a set of simulation snapshots (data) to create a lower-dimensional representation of the system.
  • Empirical Cubature Method (ECM) for Hyper-reduction: This is the novel aspect. ECM is used to create "hyper-reduced" models (HROMs). Instead of using all the degrees of freedom from the ROM, ECM intelligently selects a reduced set of elements and weights to approximate the system's residual. This significantly reduces the computational cost during the "online" stage (when running simulations for new parameters) by avoiding the need to evaluate the full ROM. The paper details how ECM is applied to select these critical elements and weights.

The framework is applied to a contraction-expansion channel geometry, which is known to exhibit the Coanda effect. Two types of geometric mappings are used: an affine mapping (simpler, straight walls) and a nonaffine mapping (more complex, curved walls) to demonstrate the framework's versatility.

Datasets

The study uses data generated from full-order model (FOM) simulations of the contraction-expansion channel. These FOM simulations serve as the "ground truth" for training the ROMs and HROMs. The paper defines specific "training trajectories" (sequences of geometric parameter changes) and "testing trajectories" to evaluate the models' ability to capture both known and unseen behaviors, including hysteresis. No external datasets are used; the data is generated internally through simulations.

Key Insights and Ideas

  • Dynamic Geometry is Key for Hysteresis: The paper strongly suggests that time-dependent geometric deformations are crucial for observing and capturing hysteresis in the Coanda effect. Static geometric changes are insufficient to reveal this history-dependent behavior.
  • ECM is Essential for Speedup: While POD reduces the model's dimension, ECM is what enables significant computational speedups by drastically reducing the number of elements and computations needed in the online stage.
  • Training Data Richness Matters: Including snapshots from both stable branches of the bifurcation (e.g., jet attaching to the upper and lower walls) in the training set is vital for developing robust ROMs/HROMs that can accurately predict behavior on unseen trajectories, especially when those trajectories might explore the opposite branch.
  • Truncation Tolerances are a Trade-off: The choice of truncation tolerances for both the solution (ESOL) and residuals (ERES) directly impacts accuracy and speedup. Lower tolerances generally lead to higher accuracy but less speedup. Critically, a less stringent residual tolerance (ERES = 1e-3) often leads to unstable or inaccurate HROMs, especially for challenging test cases.
  • HROMs Can Be "Good Enough": Even with some errors, HROMs can be valuable for applications where detecting the Coanda effect is paramount, even if the exact bifurcation branch isn't perfectly captured. For higher precision, ROMs with moderate speedups are better.

Key Results

  • Accurate Simulation of Coanda Effect and Hysteresis: The proposed ROM/HROM framework successfully captures the Coanda effect and its associated hysteresis loops in the contraction-expansion channel, even with dynamic geometric changes.
  • Significant Speedups: HROMs achieved speedups of up to several hundred times compared to the FOM, while maintaining physically acceptable solutions. ROMs without hyper-reduction offered much smaller speedups (less than an order of magnitude).
  • Impact of Geometric Mapping: The nonaffine mapping (with curved walls) generally required more POD modes than the affine mapping (with straight walls) to achieve similar accuracy.
  • Performance on Training vs. Testing Trajectories: Models trained on data encompassing both stable bifurcation branches (Example 2) performed significantly better on unseen testing trajectories compared to models trained on data from only one branch (Example 1).
  • Error Analysis: Percentage errors for the QoI (vertical velocity at a probe point) were generally higher than for the full solution field. HROMs with stringent residual tolerances (ERES < 1e-4) showed much lower errors.

Limitations and Future Work

  • Computational Cost of Training: The paper notes that the training stage for HROMs can be computationally expensive, particularly in terms of memory, especially when dealing with a large number of elements and POD modes.
  • Multidimensional Parameter Spaces: The current study focuses on a single geometric parameter. Future work will explore multidimensional parameter spaces, which will require more advanced data management and parallelization techniques.
  • Parallelization: The authors are actively developing parallelization strategies for the entire workflow (simulation, data management, SVD) to handle larger problems and improve scalability.
  • Alternative ROM Techniques: For applications requiring very high accuracy and speedup, the authors suggest exploring nonlinear manifold ROMs and autoencoder neural networks.
  • Mesh Refinement: The study used relatively small meshes. Future work will investigate the impact of higher mesh resolution and increased accuracy requirements.

Summary updated

BibTeX
@article{bravo2024geometrically,
  author = {Jose Raul Bravo and Giovanni Stabile and Markus Hess and J. A. Hernandez and Riccardo Rossi and Gianluigi Rozza},
  title = {{Geometrically parametrised reduced order models for studying the hysteresis of the Coanda effect in finite element-based incompressible fluid dynamics}},
  journal = {Journal of Computational Physics},
  year = {2024},
  doi = {10.1016/j.jcp.2024.113058},
}
Figure