TL;DR
This paper explores how Physics-Informed Neural Networks (PINNs) can be optimized for turbulence modeling by investigating different levels of "solver constraints." They found that PINNs, especially with reduced equation sets, offer significant improvements in both accuracy and speed compared to traditional data-driven approaches, even with limited training data. This suggests a promising path for more efficient and accurate fluid dynamics simulations.
What Problem Does This Paper Address?
Simulating turbulent flows is crucial for many engineering applications, but traditional methods like Computational Fluid Dynamics (CFD) often rely on simplifying assumptions or "closed" numerical models. These simplifications can lead to inaccurate predictions or require excessive computational time. While data-driven neural networks offer faster predictions after training, they struggle with limited datasets and generalizing to unseen scenarios. Physics-Informed Neural Networks (PINNs) offer a middle ground by incorporating physical laws into the neural network's learning process, but their optimal configuration for complex problems like turbulence remains an open question. This paper specifically investigates how reducing the complexity of the governing equations within a PINN framework affects accuracy and computational cost, aiming to find a balance between physical fidelity and efficiency.
Methods
The core of this research involves implementing and comparing various PINN architectures and configurations for simulating turbulent flow over a periodic hill. The key methodological contributions are:
- PINN Architectures: They tested several PINN models, ranging from a full Reynolds-Averaged Navier-Stokes (RANS) equation model (Direct Reynolds Stress Model - DRSM) to progressively simplified models. These simplifications include:
- Reduced Boundary Enforcement: Models where not all variables (e.g., first-order stresses, pressure) are enforced at the boundaries.
- Reduced Equation Sets: Models that solve only a subset of the governing equations (e.g., Continuity Only Model - COM).
- Turbulence Closure Models: Incorporating traditional turbulence models like Mixing Length (MLM) and Turbulent Viscosity (TVM) within the PINN framework.
- Data-Driven Baseline: A simple Multi-Layer Perceptron (MLP) was used as a baseline to highlight the limitations of purely data-driven approaches with sparse datasets.
- Loss Function: PINNs utilize a combined loss function that includes both boundary condition errors (MSEbc) and physics-based errors from the governing equations (MSEpde).
- Training: Normalization of input data and early stopping were employed to improve convergence and prevent overfitting.
Datasets
The primary dataset used in this study is from Direct Numerical Simulation (DNS) data of turbulent flow over periodic hills. This dataset, provided by Xiao et al. [18], contains 29 different periodic hill geometries for a Reynolds number of 5600. This dataset is noted for its detail and suitability for data-driven models.
Key Insights and Ideas
- The Power of Physics (Even Simplified): PINNs consistently outperform the purely data-driven model, even with limited training data. This highlights the advantage of incorporating physical laws, even if the equations are simplified.
- Reduced Complexity, Increased Efficiency: Surprisingly, models with fewer governing equations (like the Continuity Only Model) achieved very high accuracy for velocity predictions and significantly reduced solve times. This suggests that for certain applications, a full set of RANS equations might be computationally overkill.
- Assumptions Matter (A Lot): Introducing poor assumptions within traditional turbulence models (like the Mixing Length Model) within a PINN framework leads to significant convergence issues and poor accuracy, especially at boundaries. This underscores the importance of carefully selecting or deriving closure models.
- Unclosed Systems are Viable: The study proposes and demonstrates that numerically "unclosed" equation sets (where not all variables are explicitly solved for) can be effectively modeled by PINNs, leading to reduced computational demands.
- Data-Driven Limitations: The simple data-driven MLP struggled to capture complex flow features and showed high errors, particularly for transverse velocity and pressure, reinforcing the challenges of using sparse datasets for complex fluid dynamics.
Key Results
- Data-Driven Model: Achieved errors of 2.9% and 67.5% for streamwise and transverse velocities, respectively, highlighting its limitations with sparse data.
- Direct Reynolds Stress Model (DRSM) PINN: Significantly reduced errors to 0.613% for streamwise velocity.
- Continuity Only Model (COM) PINN: Achieved exceptionally low errors of 0.147% for streamwise and 2.562% for transverse velocities, with a solve time of only 0.74 hours. This was the most accurate and efficient model for velocity prediction.
- Mixing Length Model (MLM) PINN: Showed poor convergence at boundaries and higher errors (6.92% streamwise, 15.7% transverse), despite a low solve time.
- Turbulent Viscosity (TVM) and Turbulent Viscosity and Kinetic Energy (TVKEM) PINNs: Showed progressively improved accuracy over MLM, with TVKEM achieving errors of 0.47% (streamwise) and 5.62% (transverse).
Limitations and Future Work
- Dataset Size: The authors acknowledge that the limited size of the dataset is a primary reason for the data-driven model's shortcomings, particularly in capturing complex nonlinear dynamics and pressure fields.
- Generalizability: While the study focused on a periodic hill flow, further validation on cases with more complex turbulence phenomena (e.g., vortex shedding, flow separation) is needed.
- Boundary Condition Enforcement: The paper suggests that the optimal strategy for boundary condition enforcement might involve an initial full enforcement followed by relaxation, but this requires further investigation.
- Unclosed Systems: While promising, the effectiveness of unclosed systems needs to be proven across a wider range of flow cases and turbulence regimes.
- Computational Cost: Although PINNs offer speedups, further optimization of training strategies and architectures is always beneficial. The K-Omega model, for instance, required a higher learning rate to converge, indicating potential challenges with more complex equation sets.
Summary updated