1 of 1

VDL-Surrogate: A View-Dependent Visualization Surrogate for Ensemble Simulations

Scientific Achievement

An AI-based visualization surrogate for rapid exploration of parameter space of scientific ensemble simulations

Significance and Impact

Visualization surrogate allows scientists to bypass expensive simulations but still obtain high quality and interactive visualization of simulation results generated from arbitrary simulation input.

Research Details

Ray-based Autoencoder (RAE): Uses view dependent training data collected from pre-generated simulation results to produce view dependent latent representations
VDL-Predictor: Uses view-dependent latent representations as training data to train a deep neural network based regressor that can take arbitrary simulation input to predict simulation output
RAE+VDL-Predictor： From given simulation input, VDL-Predictor will generate view-dependent latent as the simulation output for multiple pre-selected view angles. RAE will then decode the latent and perform viewpoint interpolation
Performance: VDL-surrogate can speed up the generation of simulation output by 100 times while maintaining high accuracy for cosmology Nyx and MAS-Ocean applications
Publication: Neng Shi, Jiayi Xu, Haoyu Li, Hanqi Guo, Jonathan Woodring, Han-Wei Shen, VDL-Surrogate: A View Dependent Latent Based Model For Parameter Space Exporation of Ensemble Simulation, IEEE Vis2022 Best Paper Honorable Mention

Visualization Surrogate

RAE/VDL Training

VDL-Predictor

VDL Output

Han-Wei Shen (OSU), Hanqi Guo (ANL/OSU), and Jonathan Woodring (LANL)

Technical Contacts:

�Han-Wei Shen (OSU) shen.94@osu.edu

�Supported by:

RAPIDS2 SciDAC Institute for Computer Science, Data, and Artificial Intelligence

�Notes:

To assist scientist to explore the immense parameter space for ensemble simulations, which is often prohibitively computationally expensive, we develop a visualization surrogate, VDL-Surrogate, using a regression-based deep neural network (bottom figure). The idea of the visualization surrogate is to collect training data from ensemble simulations with preselected input parameters to train a deep neural network regression model (top left figure). With the trained model, scientists can specify any arbitrary input, and the neural network model will generate the corresponding visualization, bypassing the expensive simulation runs (top right figure). To minimize the GPU memory cost, we employ a latent representation called Ray Based Latent, generated by a Ray Based Autoencoder (RAE) (first bullet under the research detail). The RAE sill work with a view dependent latent predictor (VDL-Predictor) to predict the simulation output for the given input parameter (third bullet under the Research Details). We show that using our visualization surrogate, we can help scientists in Cosmology (left most image, the second row in the figure) and Climate/Ocean sciences (the second and third images to the left, second row) to generate high quality visualization with 1/100 computation cost.

Our paper was selected as a best paper honoralbe mention by IEEE Vis’22 conference.