SWAP-INR: Implicit Neural Representation with Switchable Attention Paths for Time-varying Multivariate Scientific Data Compression
Organizations: Korea University · University of Notre Dame
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.
Figures & tables
| Grid type | Dataset | # of points ( ) | # of variables |
| Unstructured | ROCOM [ 3 ] | 12,440,366 5 | 8 |
| CUPID | 1,279,375 30 | 11 | |
| Synthetic | 5,253,750 30 | 4 | |
| Structured | SMR Simulation | 1,048,576 10 | 8 |
| Combustion [ 10 ] | 648,000 25 | 4 | |
| Ionization [ 36 ] | 576,600 50 | 4 |
| Dataset | Method | PSNR | NRMSE | CR | Size(KB) | Dataset | Method | PSNR | NRMSE | CR | Size(KB) | ||
| ROCOM | SIREN | 42.34 | 0.0114 | 0.9323 | 3,311 | 587 | CUPID | SIREN | 21.46 | 0.1657 | 0.5585 | 2,634 | 626 |
| NeurComp | 42.66 | 0.0105 | 0.9398 | 3,346 | 581 | NeurComp | 24.63 | 0.0850 | 0.7002 | 2,626 | 628 | ||
| CoordNet | 43.49 | 0.0105 | 0.9392 | 3,351 | 580 | CoordNet | 23.87 | 0.0914 | 0.6540 | 2,609 | 632 | ||
| Instant-NGP | 30.28 | 0.0356 | 0.4391 | 3,346 | 581 | Instant-NGP | 19.61 | 0.1791 | 0.3948 | 2,561 | 644 | ||
| fV-SRN ∗ | 40.24 | 0.0125 | 0.8764 | 3,351 | 580 | fV-SRN ∗ | 13.99 | 0.2362 | -13.4032 | 2,618 | 630 | ||
| WIRE | 25.09 | 0.0966 | -0.0004 | 3,351 | 580 | WIRE | 16.92 | 0.2595 | 0.0000 | 2,630 | 627 |
| Dataset | Method | PSNR | NRMSE | |
| ROCOM | w/o | 46.29 | 0.0082 | 0.9665 |
| w/ | 47.41 | 0.0073 | 0.9782 | |
| SMR | w/o | 37.79 | 0.0158 | 0.4195 |
| Simulation | w/ | 38.17 | 0.0153 | 0.4494 |
| Ionization | w/o | 50.09 | 0.0033 | 0.9899 |
| w/ | 50.25 | 0.0032 | 0.9905 |
| Dataset | (a) | (b) | (c) | PSNR | NRMSE | CR | Size(KB) | |
| ROCOM | 46.20 | 0.0080 | 0.9742 | 1,967 | 988 | |||
| ✓ | 47.23 | 0.0073 | 0.9782 | 1,967 | 988 | |||
| ✓ | ✓ | 47.41 | 0.0073 | 0.9782 | 1,967 | 988 | ||
| ✓ | ✓ | ✓ | 47.41 | 0.0073 | 0.9782 | 3,351 | 580 | |
| SMR Simulation | 37.25 | 0.0161 | 0.4448 | 486 | 674 | |||
| ✓ | 38.14 | 0.0154 | 0.4395 | 486 | 674 |
| Dataset | Bit-widths | PSNR | NRMSE | Size(KB) | |
| ROCOM | 32-bit | 47.89 | 0.0072 | 0.9788 | 1,101 |
| 16-bit | 47.41 | 0.0073 | 0.9782 | 580 | |
| 8-bit | 24.41 | 0.0642 | -6.5984 | 319 | |
| CUPID | 32-bit | 50.91 | 0.0033 | 0.9912 | 1,167 |
| 16-bit | 50.29 | 0.0035 | 0.9904 | 626 | |
| 8-bit | 16.19 | 0.1707 | -22.2139 | 351 |
Appendix figures & tables16 assets
Supplementary material from the paper’s appendix.
Appendix
| Methods | Multi. | Grid-agn. | No-Interf. | Corr. | Impl. |
| ZFP [ 14 ] | Yes | No | Yes | No | Official |
| TTHRESH [ 1 ] | Yes | No | Yes | No | Official |
| SZ3 [ 12 ] | Yes | Yes | No | No | Official |
| SIREN [ 27 ] | No | Yes | No | No | Official |
| NeurComp [ 16 ] | No | No | No | No | Official |
| CoordNet [ 7 ] | No | Yes | No | No | Official |
| Dataset | Method | PSNR | CR | Size(KB) |
| ROCOM | SZ3 | 47.41 | 14 | 141,163 |
| ZFP ‡ | 11 | 171,308 | ||
| TTHRESH ‡ | 169 | 11,510 | ||
| Ours | 3,351 | 580 | ||
| CUPID | SZ3 | 50.29 | 26 | 63,001 |
| ZFP ‡ | 15 | 108,906 |
| Dataset | Method | (sec) | (hr) | Inf. (sec) | Dataset | Method | (sec) | (hr) | Inf. (sec) | Dataset | Method | (sec) | (hr) | Inf. (sec) |
| ROCOM | SIREN | 216 | 36.2 | 177 | CUPID | SIREN | 140 | 4.7 | 136 | Synthetic | SIREN | 600 | 20.3 | 483 |
| NeurComp | 223 | 13.8 | 176 | NeurComp | 138 | 7.0 | 109 | NeurComp | 533 | 26.5 | 466 | |||
| CoordNet | 227 | 45.4 | 182 | CoordNet | 134 | 15.3 | 150 | CoordNet | 686 | 31.5 | 494 | |||
| Instant-NGP | 199 | 137.6 | 205 | Instant-NGP | 133 | 85.1 | 124 | Instant-NGP | 524 | 185.5 | 494 | |||
| fV-SRN | 230 | 132.7 | 200 | fV-SRN | 240 | 109.7 | 210 | fV-SRN | 600 | 338.3 | 656 | |||
| WIRE | 246 | 8.2 | 198 | WIRE | 132 | 6.7 | 142 | WIRE | 598 | 19.9 | 542 |
| Variables | PSNR | ||
| (Concentration of Boron) | ✓ | 67.12 | |
| (Turbulent viscosity) | ✓ | 32.64 | |
| (Pressure) | ✓ | 53.35 | |
| (Density of liquid) | ✓ | 53.40 | |
| (Liquid velocity vector) | ✓ | 43.14 | |
| (Liquid velocity vector) | ✓ | 46.65 |
| Variables | PSNR | ||
| (Pressure) | ✓ | 50.37 | |
| (Axial core power) | ✓ | 45.39 | |
| (Temperature of liquid) | ✓ | 45.28 | |
| (User-defined) | ✓ | 52.76 | |
| (User-defined) | ✓ | 45.74 | |
| (User-defined) | ✓ | 60.48 |
| Variables | PSNR | ||
| (Dissipation rate of turbulence) | ✓ | 79.58 | |
| (Turbulent kinetic energy) | ✓ | 75.10 | |
| (Turbulent viscosity) | ✓ | 68.70 | |
| (Pressure) | ✓ | 76.60 |
| Variables | PSNR | ||
| (Vortex identifier) | ✓ | 48.36 | |
| (Pressure) | ✓ | 47.90 | |
| (Velocity components) | ✓ | 32.62 | |
| (Velocity components) | ✓ | 31.11 | |
| (Velocity component) | ✓ | 32.22 | |
| (Vorticity components) | ✓ | 37.01 |
| Variables | PSNR | ||
| (Heat release rate) | ✓ | 30.80 | |
| (Mixture fraction) | ✓ | 30.35 | |
| (Vorticity) | ✓ | 32.80 | |
| (OH mass fraction) | ✓ | 30.74 |
| Variables | PSNR | ||
| (Molecular hydrogen) | ✓ | 52.97 | |
| (Ionized hydrogen) | ✓ | 48.21 | |
| (Neutral helium) | ✓ | 46.80 | |
| (Photon/Particle density) | ✓ | 53.02 |
| Dataset | Method | PSNR | NRMSE | CR | Size(KB) | Dataset | Method | PSNR | NRMSE | CR | Size(KB) | ||
| ROCOM | SIREN | 42.84 | 0.0105 | 0.9417 | 3,345 | 581 | CUPID | SIREN | 23.62 | 0.0098 | 0.6511 | 2,622 | 629 |
| NeurComp | 42.59 | 0.0105 | 0.9399 | 3,328 | 584 | NeurComp | 23.15 | 0.1163 | 0.6457 | 2,630 | 627 | ||
| CoordNet | 42.88 | 0.0108 | 0.9382 | 3,323 | 585 | CoordNet | 22.19 | 0.0164 | 0.6110 | 2,626 | 628 | ||
| Instant-NGP | 32.49 | 0.0266 | 0.5243 | 3,351 | 580 | Instant-NGP | 19.98 | 0.1759 | 0.4335 | 2,561 | 644 | ||
| fV-SRN ∗ | 41.45 | 0.0112 | 0.9145 | 3,351 | 580 | fV-SRN ∗ | 21.52 | 0.1616 | 0.4151 | 2,618 | 630 | ||
| WIRE | – | – | – | – | – | WIRE | – | – | – | – | – |
Implicit Neural Representations
Implicit neural representations encode signals and fields as continuous functions that can be queried at arbitrary coordinates and resolutions.
See all 200 papers