Refined Grid Sampling for Efficient Implicit Neural Representations of Large-Scale Scientific Data
Authors: Guang Zhao, Xihaier Luo, Sean McCue, Luke P. Van Roekel, Seungjun Lee, Yihui Ren, Shinjae Yoo, Wei Xu
Organizations: Brookhaven National Laboratory Upton, NY 11973, USA · University of California San Diego La Jolla, CA 92093, USA · Los Alamos National Laboratory Los Alamos, NM 87545, USA
Implicit neural representations (INRs) offer a powerful framework for modeling high-dimensional physical fields, but their training remains inefficient because evaluating dense coordinate grids is computationally expensive. Existing coordinate-pruning strategies rely on local error functions that can oversample correlated points and incur substantial overhead from full-grid error evaluation. We introduce Refined Grid Sampling (RGS), which combines stratified grid sampling, error-based refinement, and dynamic grid-resolution adjustment. This design reduces spatial redundancy without evaluating errors over the full grid. On high-resolution Navier--Stokes fields, RGS achieves higher reconstruction accuracy within a fixed wall-clock budget and reaches the accuracy of full-grid training up to 6× faster than the evaluated baselines.
Figures & tables
Figure 1: Evolution of gradient cosine correlation during INR training. Left: Heatmaps of the correlation between a reference coordinate and its neighbors within a 60×60 region at training epochs t=10 , 200, and 1,600. The high-correlation area contracts during training, indicating reduced spatial redundancy. Right: Center-row correlation profiles, showing sharper decay at later epochs.
Figure 2: Refined Grid Sampling (RGS) framework. RGS combines stratified grid sampling with error-ranked selection to balance spatial coverage and informative points, reducing redundancy and forward-pass cost. Sampling starts with coarse grids and progressively refines as gradient correlations decay, improving reconstruction accuracy.
Figure 3: PSNR and error-map comparison of sampling strategies on Navier–Stokes fields.
Sampling strategy
PSNR
SSIM
Rel. RMSE
Time/epoch (s)
Full Training
37.19 ± 2.29
0.978 ± 0.024
0.092 ± 0.031
0.190
Random
43.72 ± 1.20
0.982 ± 0.004
0.042 ± 0.006
0.003
EVOS
37.41 ± 1.70
0.930 ± 0.017
0.087 ± 0.017
0.006
INT
18.38 ± 0.85
0.212 ± 0.094
0.773 ± 0.109
0.036
RGS (ours)
56.24 ± 2.73
0.998 ± 0.001
0.010 ± 0.003
0.005
Table 1: Performance after a 20-second wall-clock budget, reported as mean ± standard deviation over six independently reconstructed time frames. Best values are shown in bold.
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 4: PSNR versus training epochs for different sampling strategies. RGS converges faster and attains higher final accuracy than the error-driven baselines, while random sampling converges more slowly.
Figure 5: Training time versus epochs for different sampling strategies. RGS substantially reduces training time relative to full training and INT, while remaining competitive with random sampling and EVOS.
School of Computer Science and Engineering, Sun Yat-sen University · Department of Computer Science, Hong Kong Baptist University · School of Computer Science and Engineering, Sun Yat-sen University, Guangdong Province Key Laboratory of Computational Science, and National Supercomputer Center in Guangzhou