Compressive Modeling and Visualization of Multivariate Scientific Data using Implicit Neural Representation

TL;DR

This work addresses efficient compression and visualization of high-dimensional multivariate scientific data by employing MVNet, a residual SIREN-based implicit neural representation that jointly models tens to hundreds of variables with a single network. By learning a mapping $F( heta): \mathbb{R}^d \rightarrow \mathbb{R}^v$ that predicts all variables from coordinates, MVNet achieves superior data, feature, and visualization quality compared to LERP, TTHRESH, ZFP, and Copula methods, while maintaining compact storage. The method demonstrates strong preservation of linear and non-linear dependencies (via Pearson correlation and mutual information) and supports accurate multivariate query-driven visualization, with robust performance across varying variable counts and training data availability. The results suggest practical impact for scientific visualization workflows, enabling efficient analysis and discovery, and the authors outline future work on extending MVNet to time-varying and spatio-temporal datasets.

Abstract

The extensive adoption of Deep Neural Networks has led to their increased utilization in challenging scientific visualization tasks. Recent advancements in building compressed data models using implicit neural representations have shown promising results for tasks like spatiotemporal volume visualization and super-resolution. Inspired by these successes, we develop compressed neural representations for multivariate datasets containing tens to hundreds of variables. Our approach utilizes a single network to learn representations for all data variables simultaneously through parameter sharing. This allows us to achieve state-of-the-art data compression. Through comprehensive evaluations, we demonstrate superior performance in terms of reconstructed data quality, rendering and visualization quality, preservation of dependency information among variables, and storage efficiency.

Figures (19)

• Figure 1: Schematic architecture of MVNet. It uses a residual SIREN architecture to enhance its learning capability.
• Figure 2: Comparison of reconstruction quality for MVNet, LERP, and TTHRESH. We provide a boxplot for each method for each dataset. We observe that MVNet consistently delivers superior PSNR compared to LERP and TTHRESH.
• Figure 3: Visualization of isosurface of Velocity Magnitude at isovalue=41 of Isabel dataset with ground truth (GT) shown in Fig. \ref{['isabel_isosurf_41_GT']}. Isosurface produced by LERP is least accurate while MVNet and TTHRESH produce visually comparable results.
• Figure 4: Reconstructed data visualization for Pressure (P) variable of Isabel dataset. Images generated by ground truth (GT), MVNet, TTHRESH, and LERP are shown. We observe that MVNet produces the most accurate result.
• Figure 5: Reconstructed data visualization for the U-Velocity (U10) variable of the Climate100 dataset. Images generated by ground truth data, MVNet, TTHRESH, and LERP methods are shown from left to right with a region zoomed below to show the differences. We observe that MVNet produces the most accurate result, while the result from TTHRESH shows artifacts as highlighted by the red dotted box in Fig. \ref{['climate100_u10_TTHRESH_comb']}. The LERP method-generated image is the least accurate.