TL;DR
This paper introduces an optimization-based domain decomposition (DD) approach to create reduced-order models (ROMs) for complex, time-dependent fluid dynamics problems that depend on various parameters. By combining DD with ROM techniques like POD-Galerkin and neural networks, they achieve significant speedups for simulations while maintaining good accuracy, making real-time analysis of these problems feasible.
What Problem Does This Paper Address?
Simulating fluid dynamics, especially when dealing with time-dependent and parameter-dependent scenarios (like varying viscosity or flow rates), is computationally very expensive. This limits their use in real-time applications or for extensive parameter studies. Traditional methods often require massive computational resources. Domain decomposition (DD) breaks down large problems into smaller, manageable subdomains, and reduced-order modeling (ROM) aims to create simplified models that capture the essential dynamics with much less computational cost. However, effectively combining these two powerful techniques for complex, unsteady, and parametric fluid dynamics problems, particularly ensuring accurate coupling between subdomains, remains a challenge. This paper addresses this by developing an optimization-based DD framework that allows for efficient ROM construction.
Methods
The core of this work is an optimization-based domain decomposition (DD) approach. The problem is formulated as an optimal control problem where the objective is to minimize the "coupling error" at the interface between subdomains. This allows for a separation of solvers for each subdomain.
For creating reduced-order models (ROMs), they explore two main techniques:
- POD-Galerkin: This is a classical intrusive ROM method. It uses Proper Orthogonal Decomposition (POD) to generate reduced basis spaces for each subdomain and then applies a Galerkin projection to the governing equations within these reduced spaces.
- POD-NN (Neural Network): This is a non-intrusive ROM method. After generating POD reduced spaces, a neural network is trained to directly map physical parameters and time to the reduced coefficients of the solution. This bypasses the need for an optimization loop in the online phase.
They also perform a convergence analysis for the non-stationary incompressible Navier-Stokes equations within this framework.
Datasets
The paper uses two benchmark fluid dynamics problems for evaluation:
- Backward-facing step flow: This involves a channel with a sudden expansion, featuring a physical domain with specific dimensions and boundary conditions.
- Lid-driven cavity flow: This is a standard benchmark problem using a unit square domain with a moving lid.
For both cases, the "high-fidelity" solutions (monolithic, full-order models) are used to generate snapshots for training the POD and neural network models. The parameters explored include viscosity ($\nu$) and inlet/lid velocity magnitude ($\bar{U}$).
Key Insights and Ideas
- Optimization as a Coupling Mechanism: Framing the DD problem as an optimal control problem, where the goal is to minimize interface error, provides a robust way to couple subdomains. This is a key novelty.
- POD-Galerkin Robustness: The POD-Galerkin approach is found to be more robust and accurate, especially in transient regions and near discontinuities, compared to the POD-NN method.
- POD-NN Speed vs. Accuracy Trade-off: The POD-NN method offers significant speedups in the online phase but can be less precise, particularly when dealing with discontinuities in time or small solution norms.
- Sensitivity to Discontinuities: Both POD-NN and the overall DD approach can be sensitive to initial/boundary condition discontinuities. Restricting analysis to time intervals where these discontinuities vanish improves POD-NN performance.
- Computational Efficiency: The DD-ROM approach drastically reduces the number of iterations required for optimization compared to the full-order model (FOM), leading to substantial computational savings.
Key Results
- Significant Speedup: The DD-ROM approach achieved a substantial reduction in computational effort. For the backward-facing step, the average number of optimization iterations for the ROM solver was 24, compared to 170 for the FOM. Similar improvements were seen for the lid-driven cavity flow.
- Accuracy: The POD-Galerkin ROM generally provided good accuracy, with relative errors often below 1% compared to the monolithic solution.
- POD-NN Performance: While very fast, the POD-NN model's accuracy varied. It performed well for velocity fields but struggled with pressure fields, especially in the presence of time discontinuities. However, with careful training set selection (e.g., excluding initial transient phases), its performance improved.
- Parameter Dependence: The methods were tested across ranges of physical parameters (viscosity, velocity magnitude), demonstrating their ability to handle parameter-dependent problems.
Limitations and Future Work
- Sensitivity to Discontinuities: The POD-NN method's sensitivity to initial and boundary condition discontinuities is a noted limitation. The authors suggest that for problematic regions, a hybrid approach combining POD-NN with FOM or POD-Galerkin might be beneficial.
- Domain Splitting: The choice of domain splitting can impact convergence speed, a common issue in DD methods.
- Future Work: The authors suggest extending the algorithm to more subdomains for complex engineering problems, exploring multi-physics or fluid-structure interactions by coupling existing solvers, and performing sensitivity analysis with respect to model parameters and problem geometry. They also mention investigating the use of POD-NN in combination with other methods for improved robustness.
Summary updated