A Few-Shot Learning Framework for Time-Varying Scientific Data Generation via Conditional Diffusion Model.

Area of Science:

• Scientific Visualization
• Deep Learning
• Computational Science

Background:

• Deep learning models require large datasets for optimal performance.
• Scientific visualization often faces data scarcity due to high computational costs and storage limitations.
• Existing methods struggle with limited data availability in scientific visualization.

Purpose of the Study:

• To develop a few-shot learning framework for scientific visualization tasks.
• To address data sparsity issues in training deep learning models.
• To ensure generalization and high performance with minimal training samples.

Main Methods:

• A conditional diffusion model framework with forward (noise injection) and reverse (denoising) processes.
• Utilized a time-aware UNet for iterative denoising.
• Introduced a noise-aware loss function for dynamic optimization weighting.

Main Results:

• The proposed method demonstrates consistent and robust performance across various few-shot scenarios (1, 3, or 5 samples).
• Achieved superior quantitative and qualitative results compared to state-of-the-art methods.
• Successfully applied to spatial super-resolution, temporal super-resolution, and variable translation tasks.

Conclusions:

• Few-shot learning with conditional diffusion models is effective for data-scarce scientific visualization.
• The framework generalizes well and provides high performance regardless of sample selection.
• Offers a viable solution for training deep learning models in resource-constrained scientific domains.