High fidelity compression method improves scientific simulation data storage
SWAP-INR: Implicit Neural Representation with Switchable Attention Paths for Time-varying Multivariate Scientific Data Compression
Computer Vision and Pattern RecognitionMachine Learning
Summary
Scientific simulations produce huge amounts of complex data that are hard to store efficiently. The authors introduce SWAP-INR, a new approach that uses smart neural network designs to better compress data involving multiple physical variables changing over time. Their method pays special attention to how these variables affect each other and adapts to different data features. It achieves strong compression while keeping data quality high and storage needs low, tested on several reactor and physics datasets.
What this means in practice
- •For high performance computing teams: Compress large multivariate simulation outputs to reduce data storage while preserving quality in high-performance computing environments.
- •For climate modeling engineers: Store and transmit complex climate simulation variables more efficiently with techniques that handle variable interactions and time evolution.
Authors
Hyunsoo Son, Jeonghyun Noh, Chaoli Wang, Won-Ki Jeong
Abstract
Visualizing multivariate scientific simulation data is essential for understanding complex physical phenomena, yet the massive scale of such data imposes a severe storage bottleneck in high-performance computing environments. Implicit neural representations (INRs) have emerged as a promising continuous alternative to traditional lossy compressors, but existing methods struggle to explicitly capture cross-variable interactions and to adapt to the diverse spectral characteristics of physical variables. To address these limitations, we propose SWAP-INR, an architecture for high-fidelity compression of multivariate, time-varying simulation data. SWAP-INR introduces two key components: (1) The switchable attention path (SWAP) module explicitly models cross-variable interactions through a cross-variable attention layer, then routes each variable to an attention or non-attention path by a deterministic, loss-driven criterion, and prunes the inactive layers for structural compression. (2) The adaptability amplification SIREN (A2-SIREN) replaces rigid frequency scaling with channel-wise learnable modulation, adapting to heterogeneous physical features with negligible overhead. Combined with post-training 16-bit quantization, SWAP-INR achieves state-of-the-art performance on small modular reactor (SMR) datasets and remains competitive on public physics simulation datasets, while maintaining a constrained memory footprint. Our code is https://github.com/hvcl/SWAP-INR.