SINCPS: Semantic-aware Implicit Neural Compression for Physics Simulations

Machine learning surrogates and data-driven scientific discovery require efficient access to simulation data, yet physics simulations generate terabyte-scale datasets. Traditional compression either achieves insufficient ratios or corrupts physics-critical features like conservation laws.

SINCPS leverages wafer-scale computing to train implicit neural representations in 2-3 hours each. Across 22 datasets from The Well benchmark, we achieve 150x to 25,000x compression while preserving domain-specific conservation laws.

Key Results

Physics Domain	Compression Ratio	PSNR (dB)	L2 Error
Astrophysics	4,213x	32 dB	3.4%
Compressible Flow	25,348x	21 dB	5.5%
Wave Phenomena	6,759x	27 dB	3.1%
Turbulence	152x	14 dB	13.6%

All 22 models compress to 37.6 MB each from original sizes of 5.4 GB to 1.9 TB.

Installation

pip install torch numpy h5py pyyaml

Quick Start

Decompress a Checkpoint

from sincps import SINCPSDecompressor

# Load a trained model
decompressor = SINCPSDecompressor(
    checkpoint_path='checkpoints/shear_flow/checkpoint_50000.mdl',
    config={
        'input_dim': 3,      # time + 2D spatial
        'output_dim': 4,     # number of fields
        'hidden_dim': 1024,
        'num_hidden_layers': 4,
    }
)

# Reconstruct specific timesteps
data = decompressor.reconstruct(
    spatial_shape=(256, 256),
    num_timesteps=100,
    timestep_indices=[0, 50, 99],  # reconstruct 3 timesteps
)
# Returns shape: (3, 256, 256, 4)

Low-Level API

from sincps import load_model, reconstruct

# Load model
model = load_model(
    'checkpoints/mhd_64/checkpoint_50000.mdl',
    config={'input_dim': 4, 'output_dim': 6, 'hidden_dim': 1024}
)

# Reconstruct full 3D volume at middle timestep
data = reconstruct(
    model,
    spatial_shape=(64, 64, 64),
    num_timesteps=100,
    timestep_indices=[50],
)

Model Architecture

SINCPS uses a SIREN network with Fourier positional encoding:

1. Input: Normalized spatiotemporal coordinates [0, 1]
2. Fourier Encoding: Multi-scale frequencies from π to 512π
3. SIREN Layers: 4 hidden layers × 1024 units with sinusoidal activations
4. Output: Z-score normalized field values

Coordinates → Fourier Encoding → SIREN Layers → Field Values
   (t,x,y)        (63-dim)        (4×1024)        (n_fields)

Available Checkpoints

Trained models for 22 physics datasets from The Well:

Dataset	Dimensions	Fields	Checkpoint
shear_flow	2D+T	4	checkpoints/shear_flow/
rayleigh_taylor_instability	2D+T	4	checkpoints/rayleigh_taylor_instability/
mhd_64	3D+T	6	checkpoints/mhd_64/
supernova_explosion_64	3D+T	6	checkpoints/supernova_explosion_64/
acoustic_scattering_maze	2D+T	2	checkpoints/acoustic_scattering_maze/
gray_scott_reaction_diffusion	2D+T	2	checkpoints/gray_scott_reaction_diffusion/
...	...	...	...

See checkpoints/ for the complete list and configs/ for model configurations.

Training

Models were trained on the Cerebras CS-3 wafer-scale engine:

• Training time: 2-3 hours per dataset (vs 6-11 hours on GPU/CPU)
• Batch size: 16,384
• Steps: 50,000
• Loss: MSE
• Optimizer: Adam

Training code is available in the training/ directory for reference.

Citation

@inproceedings{sincps2025,
  title={SINCPS: Semantic-aware Implicit Neural Compression for Physics Simulations},
  author={...},
  booktitle={...},
  year={2025}
}

Acknowledgments

This work was made possible by:

• ByteBoost Cybertraining Program (NSF Awards: 2320990, 2320991, 2320992)
• Neocortex Project (NSF Award: 2005597)
• ACES Platform (NSF Award: 2112356)
• Ookami Cluster (NSF Award: 1927880)
• Cerebras Systems for CS-3 access and technical support

License

MIT License