EnsembleNGP: Exploring Time-Varying Ensemble Data via Parameter and Spatial Space Decomposition
Authors
Jun Han (The Hong Kong University of Science and Technology)
Presentation
- Session
- Time after time
- Time
- Thursday, Nov 12, 10:48 – 11:00 (US/Eastern) · session 10:00 – 11:30
- Location
- Hall America north
Links
Sign in to access the preprint PDF.
Sign in
Keywords
Time-varying ensemble data, volume visualization, implicit neural representation, representation learning
Abstract
Neural representations have been widely investigated as surrogate models for ensemble data prediction. However, existing approaches typically learn local (spatial) and global (ensemble and temporal) representations in a coupled manner, which limits their ability to model time-varying scenarios effectively. To overcome this limitation, we propose EnsembleNGP, a hybrid neural representation tailored for time-varying ensemble data that explicitly decomposes parameter and spatial spaces. The framework consists of two stages: parameter representation and spatial refinement. In the parameter representation stage, a parameter embedder and a decoder are jointly optimized to learn the mapping from a parameter set (i.e., ensemble and temporal parameters) to volumetric data. In the spatial refinement stage, a hash table is used to interpolate spatial features at a given spatial coordinate based on neighborhoods defined by multiple distance measurements. These interpolated features, together with the parameter embedding, are fed into a lightweight implicit neural representation that progressively leverages the learned spatial features to predict one voxel value. We evaluate our method on multiple time-varying ensemble datasets with diverse characteristics. Both quantitative and qualitative results demonstrate that EnsembleNGP significantly outperforms the state-of-the-art approaches, including CoordNet, MoE-INR, fV-SRN, APMGSRN, Explorable INR, and DRR Net. Moreover, it better captures parameter correlations and temporal evolution while achieving substantially lower storage and generation costs than traditional simulations.
For Practitioners
simulation scientistis and data scientists