VQ-ANR: Attention Neural Representation with Vector Quantization for Time-Varying Ensemble Data Exploration
Jump to Discussion
Presentation
- Session
- Making things visual in science and engineering
- Time
- Tuesday, Nov 10, 13:24 – 13:36 (US/Eastern) · session 13:00 – 14:30
- Room
- Hall Essex center
- Presenting from
- Boston
Links
Sign in to access the Paper PDF.
Abstract
Neural representation–based surrogate models have demonstrated remarkable performance in synthesizing ensemble data. However, existing approaches primarily focus on steady ensemble data and face challenges in extending to time-varying cases due to the limited expression of interpolable feature grids in parameter space. To address this limitation, we propose VQ-ANR, an attention-based neural representation with vector quantization for exploring time-varying ensemble data. Our framework consists of two stages: codebook construction and ensemble exploration. During codebook construction, we employ an autoencoder-decoder with vector quantization to map continuous data embeddings to a discrete codebook. In the ensemble exploration stage, given a parameter input, we develop an attention mechanism to query relevant data embeddings from the learned codebook. The retrieved embeddings, together with spatial coordinates, are then fed into an implicit neural representation for voxel prediction. We evaluate VQ-ANR across diverse time-varying ensemble datasets spanning astrophysics, cosmology, hydrodynamics, and ocean domains. Both quantitative and qualitative results demonstrate that VQ-ANR can significantly outperform state-of-the-art approaches, predict high-fidelity data, and effectively capture the underlying relationships among simulation parameters over time. The source code is available on https://github.com/huayuan4396/VQ-ANR.
For Practitioners
simulation scientists, data scientists, etc.