Physics-Informed Neural Compression

• Physics-Informed Neural Compression (PINC) is a methodology that uses neural networks integrated with physics-based loss functions to compress complex, high-dimensional scientific data while maintaining critical physical invariants.
• It exploits diverse architectures such as autoencoders, implicit neural fields, and tensor decompositions to achieve remarkable compression ratios and real-time efficiency.
• PINC leverages specialized loss functions and advanced compression schemes, delivering up to 10^5× compression with minimal degradation of physical fidelity, making it essential for resource-constrained PDE and turbulence simulations.

Physics-Informed Neural Compression (PINC) refers to a suite of neural network-based methodologies that enable extreme compression of scientific data or neural PDE solvers while rigorously enforcing the preservation of essential physical properties. Unlike classical data-driven or lossy compression approaches that may distort or destroy critical physical invariants, PINC integrates physics-informed loss terms into the training of neural compression models and/or applies compression to physics-informed neural networks (PINNs) via specialized, mathematically controlled representations. This inclusive definition spans both (1) PINC for state field compression—such as full 5D complex-valued turbulence data, and (2) PINC for PDE-model surrogates—neural networks trained with physics constraints and then aggressively compressed with guarantees on solution quality or dynamics.

1. Neural Architectures and Model Design in PINC

PINC encompasses diverse neural architectures, each tailored to specific scientific domains and data types. Two central lines have emerged:

A. Data Field PINC (e.g., Plasma Turbulence Compression)

• Autoencoders (AE, VQ-VAE): Inputs are 5D complex plasma phase-space fields $$f \in \mathbb{C}^{v_\parallel \times \mu \times s \times x \times y}$$, optionally conditioned on physical parameters $(R/L_T, R/L_n, q, \hat s)$. Swin Transformer (nD-Swin) blocks and window-based self-attention modules enable hierarchical downsampling over all coordinates. Latent bottlenecks use either low-dimensional vectors ($D_\text{latent}=32$) or vector-quantized dictionaries ($K=8192$, dim 128).
• Implicit Neural Fields (NF): Each data snapshot $f_t$ is fitted by a separate MLP. Input indices are embedded via hash tables; five SiLU-activated layers with skip connections map from coordinates to complex values. These models are amenable to pipelined, in-situ parallelization for TB-scale data compression (Galletti et al., 4 Feb 2026).

B. PINC for Neural PDE Surrogates

• Hierarchical or Tensorized Representations: Weight matrices in PINNs are compressed adaptively using blockwise low-rank H-matrices, tensor-train (TT) decompositions, and quantized formats, with complexity $O(k n \log n)$ for H-matrices, and further speed/memory gains via TT at rank $R\ll n$. These methods preserve the trainability and domain generalization properties critical for scientific modeling (Mango et al., 2024, Lu et al., 10 Dec 2025).
• Domain-Decoupled Architectures: In PINNs for control and dynamics, decoupling time from the neural network via analytic Ansätze enables closed-form gradient evaluation, drastically accelerating training and improving stability (Krauss et al., 2024).

2. Physics-Informed Losses and Constraints

PINC's distinctiveness lies in physics-informed constraints that guarantee retention of scientifically meaningful observables and invariants after compression.

• Integral Losses: For gyrokinetic turbulence, explicit $L_1$ losses penalize error in key nonlinear integrals: the electrostatic potential

$$\phi(x,s,y) = A(x,s,y) \int J_0 f\, dv_\parallel d\mu$$

and heat flux

$$Q = \int B(x,s,y) v^2 \phi f\, dv_\parallel d\mu dx dy ds,$$

with loss terms $\mathcal{L}_\phi$ and $\mathcal{L}_Q$ enforcing fidelity in both real and imaginary components (Galletti et al., 4 Feb 2026).

• Spectral Diagnostic Losses: Wasserstein distances compare modal energy spectra (e.g., $k_y^{\mathrm{spec}}$, $Q^{\mathrm{spec}}$) between reconstructed and ground-truth fields, sensitive to phase and energy distribution—which traditional metrics neglect.
• Monotonicity and Isotonic Losses: Terms such as $\mathcal{L}_\mathrm{iso}$ enforce the monotonic decay of spectra beyond the injection peak, reflecting physical steady-state behavior.
• Composite Loss Formulation: The total PINC loss for field compression is

$$\mathcal{L}_\mathrm{PINC} = \mathcal{L}_\mathrm{recon} + (\mathcal{L}_Q+\mathcal{L}_\phi) + (\mathcal{L}_{k_y} + \mathcal{L}_{Q^{\mathrm{spec}}}) + (\mathcal{L}_\mathrm{iso}(k^{\mathrm{spec}}_y) + \mathcal{L}_\mathrm{iso}(Q^{\mathrm{spec}})),$$

with $\mathcal{L}_\mathrm{recon}$ the base complex MSE.

For surrogate PINNs, loss functionals typically combine initial condition, physics (residual), and optional data misfit components—with architecture-specific modifications to enforce exact physical constraints or allow for analytic derivative calculation (Krauss et al., 2024).

3. Compression Algorithms and Memory/Compute Scaling

PINC leverages advanced compression schemes at the algorithmic and hardware level:

• Hierarchical Matrix (H-matrix) Compression: Dense PINN weight matrices are recursively partitioned and approximated by blockwise low-rank factorizations, with adaptive rank selection per block to enforce a global error bound $\|W-\tilde W\|_F \leq \varepsilon$. This yields $O(k n \log n)$ memory and flops, with near-dense accuracy even at $>90\%$ compression, and empirically preserves the spectral and convergence properties controlled by the neural tangent kernel (NTK) (Mango et al., 2024).
• Tensor-Train (TT) Decomposition: Weight tensors are expressed as chains of third-order "cores," reducing parameters from $O(n^2)$ to $O(d n R^2)$ for depth-$d$ with rank $R$. Partial-reconstruction schemes (PRS) reduce quantization error when used under low-precision training (Lu et al., 10 Dec 2025).
• Quantization and Difference-Based Schemes: PINC for edge-device PINN training uses bidirectional INT8/INT12 square-block (SMX) quantization formats and difference-based quantization for Stein's estimator to preserve fine-resolution residual signals during forward and backward passes.
• Entropy Coding for Data Fields: After vector quantization, additional Huffman encoding leverages non-uniform codeword frequency to extend compression rates to $>10^5\times$ (Galletti et al., 4 Feb 2026).

A summary table of compression effectiveness:

Method	Memory (relative)	Test Accuracy	Speedup
Dense	1.0	0.983	1×
SVD	0.10	0.976	1.5×
Pruning	0.12	0.960	1.8×
Quantization	0.25	0.930	1.0×
PINC H-matrix	0.08	0.981	2.8×

4. Evaluation Pipelines and Physical Diagnostics

Comprehensive PINC evaluation employs a domain-specific, multi-dimensional fidelity pipeline:

For Plasma Data Compression (Galletti et al., 4 Feb 2026):

• Spatial Fidelity: PSNR and $L_1$ errors for $f$; errors in physical invariants $\phi$ and $Q$; spectral diagnostics (Wasserstein distance, peak correlation) for $k_y$ and $Q^{\mathrm{spec}}$.
• Temporal Fidelity: Energy cascade error, quantified by timestep-summed Wasserstein distances; optical-flow endpoint error (EPE) based on 5D Horn–Schunck flow in phase space, tracking dynamical transport and turbulence evolution.

For Surrogate PINNs (Krauss et al., 2024, Lu et al., 10 Dec 2025):

• Physics-Loss and Test Error: Direct minimization and reporting of physics-loss components and test-time MSEs on out-of-sample control or system dynamics.
• Convergence & Speed: Empirical wall-clock reduction in training and inference time through compression, often by up to two orders of magnitude.

5. Quantitative Results and Empirical Performance

Representative PINC results:

A. Gyrokinetic Turbulence Compression (Galletti et al., 4 Feb 2026):

• Baseline (traditional) approaches (ZFP, Wavelet, PCA, JPEG2000): $\sim 10^3\times$ compression, PSNR($f$) = 29–34 dB, but poor physical fidelity (e.g., $L_1(Q) \approx 60$–110).
• PINC (Neural Field + Physics Loss) at $\sim 10^3\times$: $L_1(Q)$ drops to $2.18$, $PSNR(\phi)=13.5$ dB, $WD(k_y)=0.006$, indicating >10× fidelity improvement in physical invariants.
• VQ-VAE + EVA + Entropy coding: up to $1.2 \times 10^5$ compression, with physical errors only marginally increased over neural baselines.

B. PINC for Edge/Resource-Constrained PDE Solvers (Lu et al., 10 Dec 2025):

• PINC achieves MSE within $10–30\%$ of full-precision baseline on 2D Poisson, 20D HJB, and 100D Heat equations.
• Hardware-accelerated (PINTA) implementations yield $5.5-83.5\times$ speedup and $159.6-2324.1\times$ energy savings relative to uncompressed, full-precision training.

C. PINC for Model Compression and Control (Krauss et al., 2024, Mango et al., 2024):

• H-matrix PINC matches or exceeds the test accuracy of SVD, pruning, and quantization at higher compression ratios, with up to $2.8\times$ speedup over dense baseline.
• Domain-decoupled PINC (DD-PINN) accelerates training by $5–38\times$ on complex dynamical systems, with better or equal long-horizon stability and dramatically reduced divergence in control benchmarks.

6. Limitations and Future Directions

PINC—regardless of instantiation—bears several acknowledged limitations:

• Physics-informed loss terms are currently domain-specific and require redesign for other classes of PDEs or scientific data (e.g., conservation laws in CFD, seismic or wave-equation constraints).
• Temporal physics loss is not explicitly enforced in most PINC architectures for data field compression; movement errors (e.g., EPE) may rise, and improved fidelity will require temporally aware loss terms.
• Training stability for end-to-end multi-stage PINC (AE+EVA) remains sensitive; neural field fits can be compute-intensive at low compression ratios.

Future work targets:

• Integration of temporal PINC losses (e.g., invariants linked to time derivatives or spectral recurrence constraints).
• Generalization of PINC to new physics (mass/momentum/energy conservation, wave propagation).
• Fusion of PINC with neural-operator surrogates for in-situ, online post-processing and predictive modeling.
• Hardware/software co-design for real-time, large-scale deployment.

7. Significance Across Domains

PINC provides an enabling technology for scientific computing where both storage and analysis would otherwise be intractable. The ability to preserve spectral structure, dynamical invariants, and physically coherent evolution at compressions $>10^4\times$ facilitates full-field storage, high-throughput diagnostics, and downstream post-hoc analysis in plasma physics, complex dynamical systems, and high-dimensional PDEs (Galletti et al., 4 Feb 2026, Lu et al., 10 Dec 2025, Mango et al., 2024). Rigorous maintenance of domain-specific invariants provides confidence in compressed data representations, distinguishing PINC from generic neural compression.

A plausible implication is that PINC-style approaches will become foundational not only in plasma physics, but throughout scientific machine learning, as the ESSENTIAL methodology for scaling data storage, real-time modeling, and resource-efficient deployment, subject to the continued development of rigorous, problem-adapted physics-informed constraints.