🌊 Physics-Informed Super-Resolution for Fluid Dynamics

Bridging the gap between coarse simulations and high-fidelity reality using Deep Learning.

🚀 Project Overview and Goal

Simulating high-resolution fluid dynamics is computationally expensive. This project explores a Deep Learning approach to Super-Resolution (SR), enabling the reconstruction of fine-grained turbulent details from low-resolution, "blocky" inputs.

Our Goal: To develop a model that doesn't just "upscale" images, but understands physics. By integrating Navier-Stokes constraints directly into the loss function, we ensure that our generated high-resolution fields ($256 \times 256$) are not only visually sharp but physically valid divergence-free flows.

---

🛠️ Project Pipeline & Methodology

Our pipeline treats fluid simulation as a "Sim-to-Real" problem. We generate data using the PhiFlow toolkit, creating pairs of Low-Res ($64^2$) inputs and High-Res ($256^2$) ground truth targets across 4 physical channels: Velocity (u, v), Pressure (p), and Smoke Density (s).

We investigated two distinct architectural approaches to solve this challenge:

1. ResUNet

• Input: Stack of 3 consecutive frames [t-1, t, t+1] to capture temporal flow dynamics.
• Architecture: A Residual U-Net with deep encoder-decoder paths and skip connections.
• Focus: Smoothness and temporal consistency. The model takes pre-upsampled coarse inputs and refines them to match the target.

2. SRGAN

• Input: Single-frame physics fields (4 channels).
• Architecture: A Super-Resolution GAN composed of:
  • Generator: Modified SRResNet (16 Residual Blocks, PReLU) for 4x spatial upsampling.
  • Discriminator: VGG-style network to differentiate real vs. generated turbulence.
• Physics-Informed Training:
  • Phase 1 (Pre-training): Supervised learning on mathematically downscaled inputs.
  • Phase 2 (Fine-tuning): Finetuning with a dataset where inputs are generated natively at low resolution (64x64)

Left: High-Resolution Ground Truth | Right: Low-Resolution Input