Stream functions are 3D scalar fields of which isosurfaces of visualize stream surfaces for a flow field and are difficult to compute directly. We use implicit neural representation (INRs) to learn stream functions for flow fields without any stream function given as ground truth.
Significance and Impact
Our approach learns a stream function for a given flow field implicitly with a lightweight neural network, which takes as input spatial coordinates and outputs the stream function value at that point. The stream function is learned in less than 15 minutes (depending on data size), with the option to align surfaces with flow structure or with seeding rakes of interest, enabling new visualizations for exploring fluid data with the assistance of implicit neural representations.
Pipeline for our neural stream function method, which learns a stream function for a given flow to visualize flow features via isosurfaces or volume rendering.
Technical Approach
A single 3D flow field is used for training the neural network.
The neural network maps input spatial coordinates to output stream function value, and the gradient of the output with respect to the input creates a 3D vector.
During training, we minimize the inner product of the vector field and the gradient of the neural network output with respect to the input, implicitly learning a scalar stream function as a result.
The stream function can be sampled to a grid and visualized in ParaView for flow analysis.
Volume renders of learned neural stream functions with streamlines inlayed.
Skylar W Wurster, Hanqi Guo, Tom Peterka, Han-Wei Shen. "Neural Stream Functions." In Proceedings of PacificVis 2023.