Improved Conditional Generative Network for Rapid Flow Field Prediction

Technical Summary: Improved Conditional Generative Network for Rapid Flow Field Prediction

Problem Statement
Traditional Computational Fluid Dynamics (CFD) simulations, while grounded in fundamental physical laws like the Navier-Stokes equations, face significant limitations in practical engineering applications. Specifically, simulating fluid flow around complex geometric configurations involves intricate phenomena such as turbulence and vortices, requiring high-fidelity mesh generation and massive computational resources. These factors lead to prohibitively high computational costs and memory requirements, often limiting the scale, accuracy, and real-time applicability of traditional simulations. While data-driven deep learning approaches have emerged to address these challenges, existing methods like Physics-Informed Neural Networks (PINNs) can still suffer from high computational costs and boundary constraint difficulties, and purely data-driven models often struggle to capture fine details or generalize to unseen geometries without explicit physical priors.

Methodology
This study proposes a data-driven approach utilizing an Improved Conditional Generative Adversarial Network (cGAN) to predict flow fields around complex geometries efficiently. The methodology avoids embedding specialized physical priors (such as conservation laws) directly into the network, relying instead on a purely data-driven architecture trained on high-fidelity CFD-generated data.

• Input and Output: The model accepts three-channel input data: the geometric shape of the object, the Signed Distance Function (SDF) describing the geometry, and a composite parameter ($Gr/Re^2$) combining the Grashof and Reynolds numbers. The output consists of the $x$-direction velocity field ($U_x$), $y$-direction velocity field ($U_y$), and pressure field ($P$).
• Generator Architecture: The generator is based on an Improved U-Net structure. Key modifications include:
  • Removal of Skip Connections: Unlike traditional U-Nets, skip connections between encoder and decoder layers are removed. The authors found that in flow field prediction, shallow local features do not effectively enhance global mapping and can introduce noise.
  • Residual Blocks: Residual structures are introduced to improve feature transmission stability and capture complex nonlinear relationships between inputs and outputs.
  • Activation Function: The final layer uses a linear activation function instead of Tanh to allow output values to align with the actual physical distribution of flow field data, rather than being constrained to $[-1, 1]$.
• Discriminator Architecture: A Patch-based Discriminator replaces the traditional global discriminator. It divides input data into $N \times N$ patches to perform true/false discrimination locally. This approach focuses on finer local details, reduces computational load, and provides more stable adversarial signals to prevent mode collapse.
• Loss Function: The training objective combines three components:
  1. Adversarial Loss: Standard cGAN loss for the generator and discriminator.
  2. Patch Loss ($L_1$): Measures the absolute difference between real and predicted flow fields at the patch level.
  3. Smoothness Loss: A regularization term penalizing spatial gradients to ensure continuity and smoothness in the predicted flow fields, adhering to fluid dynamics principles.

Key Contributions

1. Architectural Optimization: The study identifies that traditional U-Net skip connections are detrimental to global flow field mapping and proposes their removal, replacing them with residual structures to better model global features.
2. Patch Discriminator Integration: The implementation of a patch-based discriminator enhances the model's ability to learn local structural details and stabilizes the adversarial training process.
3. Purely Data-Driven Approach: The model demonstrates that a single cGAN architecture, trained solely on CFD data without explicit physical equation constraints, can learn complete solutions to the Navier-Stokes equations for complex geometries.
4. Comprehensive Validation: The method is rigorously tested on the FlowBench dataset, covering harmonic perturbation surfaces, NURBS surfaces, and skeleton-based topologies.

Experimental Results
The model was trained and evaluated on a dataset of over 10,000 high-fidelity samples using two NVIDIA GeForce RTX 3080 Ti GPUs.

• Accuracy: The proposed model achieved a 79.14% improvement in Mean Squared Error (MSE) compared to the baseline traditional U-Net-based cGAN (T1). Specifically, the total MSE decreased from 33.75 ($\times 10^{-5}$) for the baseline to 7.04 ($\times 10^{-5}$) for the proposed model.
• Performance Comparison: The proposed model outperformed intermediate variants (T2 with smoothness loss, T3 with residual networks) across all physical quantities ($U_x$, $U_y$, and $P$).
• Generalization: The model demonstrated robust generalization capabilities, accurately predicting flow fields for unseen geometric configurations and capturing intricate nonlinear features.
• Visualization: Visual analysis confirmed high consistency between predicted and true velocity and pressure distributions, with errors primarily localized to boundary regions. Histogram analysis showed the model effectively captured the statistical characteristics of the true data, though minor discrepancies remained in extreme value regions.

Significance and Claims
The paper claims that the proposed improved cGAN offers a versatile and efficient alternative to traditional CFD simulations for predicting flow fields around complex geometries. By eliminating the need for mesh generation and time-consuming numerical iterations, the method enables rapid, real-time flow field prediction while maintaining high accuracy. The study substantiates that deep learning models can learn the complete solutions to fluid dynamics governing equations directly from data, providing significant potential for practical engineering applications in aerospace, energy, and marine engineering. The authors note that while the model performs well, future work is needed to further enhance robustness in predicting geometric edge details and handling increased flow field complexity.