A Principled Approach to Design Using High Fidelity Fluid-Structure Interaction Simulations

Wensi Wu, Christophe Bonneville, Christopher J. EarlsFinite 2020 · DOI

Introduction

In the design of aerospace and marine structures, understanding the interplay between fluid flow and structural deformation—known as Fluid-Structure Interaction (FSI)—is critical. High-fidelity FSI simulations provide the necessary accuracy for these complex environments but come with a significant "computational tax." A single simulation can require hundreds of CPU cores and days of runtime. This computational cost makes traditional optimization methods, like grid search or gradient-based approaches, impractical. This paper presents a workflow utilizing constrained Bayesian Optimization (BO) to navigate these complex design spaces efficiently, achieving optimal results with a minimal number of high-fidelity simulations.

The FSI Framework

The researchers employ a partitioned approach to FSI, allowing specialized solvers to handle different physical domains. As shown in the system architecture, the fluid dynamics are managed by OpenFOAM, while structural dynamics are resolved using CU-BENs.

Schematic of the partitioned FSI coupling architecture between OpenFOAM and CU-BENs.

These solvers communicate through a custom Coupler that utilizes an implicit quasi-Newton method (IQN-ILS). This ensures that the fluid and structure reach a stable equilibrium at every time step, maintaining the high fidelity required for accurate engineering design.

Optimization via Bayesian Logic

To overcome the "black box" nature of FSI simulations, the framework uses Bayesian Optimization. Instead of a random search, BO constructs a Gaussian Process (GP)—a statistical surrogate model that serves as a "best guess" of the design landscape.

An acquisition function, specifically "Expected Improvement," guides the search by identifying areas where the model predicts a superior design or where uncertainty is high. This principled approach allows the optimizer to find global optima in non-linear spaces in as few as 8 to 12 simulations, a drastic reduction compared to standard methods.

Verification and Application

The framework was first verified using the Turek & Hron FSI3 benchmark, a standard problem involving an elastic flag attached to a cylinder. The simulation accurately captures the transition from a reference state to a deformed configuration under fluid loading.

Reference and deformed configurations illustrating the FSI simulation process.

Following verification, the method was applied to the design of an Underwater Unmanned Autonomous Vehicle (UUAV) sail plane. By varying material stiffness ($E$) and the angle of attack ($ heta$), the optimizer successfully identified a configuration ($ heta = 3.74^\circ$, $E = 35.39$ MPa) that minimized drag while adhering to strict physical constraints regarding lift and structural deformation.

Key Insights and Results

  • Efficiency: The BO approach reached optimal design regions significantly faster than traditional methods, proving that high-fidelity design is possible even with a limited computational budget.
  • Counter-Intuitive Physics: One notable finding was that a less stiff beam could actually result in lower tip deflection. This occurs when the softer structure triggers higher modes of oscillation that interact more favorably with the fluid's vortex shedding, a phenomenon that human intuition might overlook without high-fidelity modeling.
  • Accuracy: The FSI framework demonstrated high precision, matching established benchmarks within 2% for drag/lift and approximately 10% for structural displacement.

Conclusion and Future Work

This research demonstrates that constrained Bayesian Optimization is a powerful tool for high-fidelity engineering design. While currently focused on 2D domains and a limited number of design parameters, future work will explore scaling this approach to 3D simulations and higher-dimensional design spaces. The ability to find optimal, physically valid designs with a handful of simulations represents a significant step forward in making high-fidelity FSI a standard part of the engineering design cycle.

Summary updated

BibTeX
@misc{wu2020principled,
  author = {Wensi Wu and Christophe Bonneville and Christopher J. Earls},
  title = {{A Principled Approach to Design Using High Fidelity Fluid-Structure Interaction Simulations}},
  year = {2020},
  eprint = {2008.09687},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2008.09687},
}
Figure