A Fluid Flow Data Set for Machine Learning and its Application to Neural Flow Map Interpolation.

Area of Science:

• Scientific Visualization
• Fluid Dynamics
• Deep Learning

Background:

• Deep learning (DL) is increasingly used for neural network exploration.
• Applying DL to visualization problems requires open data sharing for data-driven research.
• Current methods for analyzing time-dependent fluid transport are memory-intensive.

Purpose of the Study:

• To construct and release a large-scale fluid flow dataset for scientific visualization research.
• To apply deep learning techniques to improve post-hoc Lagrangian fluid flow analysis.
• To enhance the accuracy and efficiency of flow map interpolation for fluid transport assessment.

Main Methods:

• Simulated a comprehensive fluid flow dataset (8000 time-dependent 2D vector fields, >16 TB) across laminar and turbulent regimes, parameterized by Reynolds number.
• Trained deep convolutional neural networks (CNNs) on the public dataset to benchmark improved Lagrangian fluid flow analysis.
• Developed a DL-based approach to improve the accuracy of flow map interpolations for in-situ analysis.

Main Results:

• Established a benchmark for post-hoc Lagrangian fluid flow analysis using deep convolutional neural networks.
• Demonstrated improved accuracy in flow map interpolations compared to traditional methods.
• Achieved a reduced memory Input/Output (IO) footprint for time-dependent fluid analysis.

Conclusions:

• The developed fluid flow dataset and deep learning methods enable more precise and efficient scientific visualization.
• Deep learning significantly enhances the analysis of transport characteristics in time-dependent fluid flows.
• Openly sharing large datasets is crucial for advancing data-driven research at the intersection of deep learning and scientific visualization.