Abstract
High-fidelity, high-resolution (HFHR) data is essential for scientific discovery but often computationally inaccessible. Low-fidelity, low-resolution (LFLR) alternatives, while common, suffer from systematic biases and noise introduced by coarse-grid solvers or sparse sensors. We present DCSR (Diffusion-based Correction and Super-Resolution), a two-stage framework that utilizes a novel Imbalanced Perturbing and Denoising (IPD) technique to correct source-agnostic biases before upscaling. By decoupling perturbation and denoising times, DCSR effectively "forgets" input errors and reconstructs high-fidelity structures, outperforming standard super-resolution baselines across fluid dynamics, climate science, and solid mechanics.
Introduction
In fields like climate science and fluid dynamics, the demand for high-fidelity, high-resolution (HFHR) data is constant. However, obtaining such data is often computationally prohibitive due to the curse of dimensionality in high-resolution simulations. Consequently, researchers often rely on low-fidelity, low-resolution (LFLR) data.
Unlike typical image super-resolution in computer vision, scientific "low resolution" data is not merely a blurry version of the truth. Coarse-grid solvers introduce systematic biases, such as artificial diffusion or incorrect wave speeds, and observational data is frequently corrupted by sensor noise. Standard Super-Resolution (SR) techniques fail in these contexts because they assume the input is a clean downsampling of the target. DCSR addresses this by first correcting these complex, source-agnostic biases before performing the resolution boost.
Methodology
The DCSR framework consists of two primary stages designed to bridge the gap between LFLR and HFHR data distributions.
1. Correction via Imbalanced Perturbing and Denoising (IPD)
The core innovation of this work is the Imbalanced Perturbing and Denoising (IPD) technique. Traditional diffusion-based editing (like SDEdit) uses balanced timing, where noise is added for a time $t$ and removed for the same duration. The authors found this approach lacks the robustness required for scientific data correction.
IPD introduces two independent time parameters: $t_1$ for perturbation and $t_2$ for denoising. This decoupling allows the model to use $t_1$ to effectively "scrub" the specific biases and noise of the input data, while $t_2$ guides the reconstruction of the underlying structure using an unconditional diffusion model trained on high-fidelity reference data. This provides a mathematical "knob" to balance bias removal against structure preservation.
2. Cascaded Super-Resolution (SR3)
Following the correction step, the cleaned low-resolution data is passed through a series of conditional diffusion models (SR3). These models iteratively refine the data in stages (e.g., from $32 \times 32$ to $256 \times 256$), ensuring the final output matches the statistical and spectral properties of the high-fidelity target.
Key Insights and Results
DCSR was evaluated across several demanding domains, including the 1D Advection Equation, Linear Elasticity, Navier-Stokes fluid flow, and ERA5 climate data.
- Bias Removal: In 1D advection tests, IPD consistently outperformed balanced methods, showing lower Total Variation Distance (TVD) and higher robustness across different numerical solver errors (Godunov, Lax-Wendroff, etc.).
- Spectral Fidelity: For Navier-Stokes simulations, DCSR's energy spectrum closely matched reference HFHR data. The authors utilized the Mean Energy Log Ratio (MELR)—a spectral domain metric—to optimize denoising parameters, finding it far more effective than standard pixel-wise metrics for scientific accuracy.
- Physics-Agnostic Versatility: Notably, DCSR does not require governing equations. It learns the distribution of "good" data from a reference pool and aligns corrupted data to that distribution, making it applicable to a wide range of physical systems.
Conclusion and Future Work
DCSR provides a robust, physics-agnostic pipeline for purifying and upscaling scientific data. While the iterative nature of diffusion models presents a challenge for sampling speed, the framework's ability to correct systematic biases represents a significant step forward. Future work involves "chaining" these models to bridge even larger resolution gaps (e.g., from 210 km to 2 km) and using DCSR as a high-quality initial guess to accelerate traditional high-resolution solvers.
Summary updated