CD-TVD: Contrastive Diffusion for 3D Super-Resolution with Scarce High-Resolution Time-Varying Data.

Area of Science:

• Scientific Computing
• Data Augmentation
• Computational Science

Background:

• Large-scale scientific simulations generate high-resolution time-varying data (TVD), demanding significant computational resources.
• Existing super-resolution methods require extensive high-resolution (HR) training data, limiting their use in diverse simulation contexts.

Purpose of the Study:

• To develop a novel framework, CD-TVD, for accurate 3D super-resolution of scientific simulation data.
• To reduce the dependency on large HR datasets by combining contrastive learning and diffusion models.

Main Methods:

• Proposed CD-TVD framework combining contrastive learning and an improved diffusion-based super-resolution model.
• Pre-training modules on historical simulation data to learn degradation patterns and sample features.
• Fine-tuning the diffusion model with a local attention mechanism using minimal new HR data.

Main Results:

• Achieved accurate 3D super-resolution from limited time-step HR data.
• Demonstrated resource-efficient enhancement of fluid and atmospheric simulation datasets.
• Successfully recovered fine-grained details despite reduced training data requirements.

Conclusions:

• CD-TVD offers an effective solution for 3D super-resolution in scientific simulations.
• The framework significantly advances data augmentation strategies for large-scale simulations.
• Minimizes reliance on extensive HR datasets while maintaining high accuracy.