ANTIC: Adaptive Neural Temporal In-situ Compressor

Sandeep S. Cranganore; Andrei Bodnar; Gianluca Galletti; Fabian Paischer; Johannes Brandstetter

cs.LGApr 10, 2026

ANTIC: Adaptive Neural Temporal In-situ Compressor

Authors: Sandeep S. Cranganore, Andrei Bodnar, Gianluca Galletti, Fabian Paischer, Johannes Brandstetter

Organizations: Institute for Machine Learning, Ellis Unit, JKU Linz, Austria · University of Manchester, United Kingdom · EMMI AI, Linz, Austria

Abstract

The persistent storage requirements for high-resolution, spatiotemporally evolving fields governed by large-scale and high-dimensional partial differential equations (PDEs) have reached the petabyte-to-exabyte scale. Transient simulations modeling Navier-Stokes equations, magnetohydrodynamics, plasma physics, or binary black hole mergers generate data volumes that are prohibitive for modern high-performance computing (HPC) infrastructures. To address this bottleneck, we introduce ANTIC (Adaptive Neural Temporal in situ Compressor), an end-to-end in situ compression pipeline. ANTIC consists of an adaptive temporal selector tailored to high-dimensional physics that identifies and filters informative snapshots at simulation time, combined with a spatial neural compression module based on continual fine-tuning that learns residual updates between adjacent snapshots using neural fields. By operating in a single streaming pass, ANTIC enables a combined compression of temporal and spatial components and effectively alleviates the need for explicit on-disk storage of entire time-evolved trajectories. Experimental results demonstrate how storage reductions of several orders of magnitude relate to physics accuracy.

Figures & tables

Figure 1: Figure 1: Schematic Overview of $\mathtt{ANTIC}$ . $\mathtt{ANTIC}$ facilitates high-fidelity neural compression via a dual-stage asynchronous architecture. Physics-aware Temporal Selector serves as a physics-aware filter that identifies salient physical transients. Spatial Neural Compression encodes the selected snapshots into compact NN weights, stored in the persistent storage. $\mathtt{ANTIC}$ enables a combined in-situ compression resulting in multi-fold reduction in data volume while preserving the temporal coherence of the entire simulation.

Figure 2: Workflow of $\mathtt{ANTIC}$ . A new snapshot at time $t+1$ is passed over by the simulator. The Metric extracts physics of interest for the new snapshot $\phi_{t+1}$ , which is passed to the Regulator. Based on $\phi_{t+1}$ , the regulator truncates the context queue and adds $\phi_{t+1}$ . Finally, the truncated context is passed to the gate along with $\phi_{t+1}$ to determine whether or not to compress the new snapshot.

Method	2D Kolmogorov flows (16.8 GiB)				3D BSSN simulations (4.2 TiB)
	TR( $\uparrow$ ) (%)	PA ( ✓ / ✗ )	SC( $\uparrow$ ) (×)	TC( $\uparrow$ ) (×)	TR( $\uparrow$ ) (×)	PA ( ✓ / ✗ )	SC( $\uparrow$ ) (×)	TC( $\uparrow$ ) (×)
sparse sampling + $\mathtt{ZFP}$ [tol= $1e\text{-1}$ ]	$20\%$	✗	$13\times$	$65\times$	$20\%$	✗	$27\times$	$141\times$
Dense sampling + $\mathtt{ZFP}$ [tol= $1e\text{-1}$ ]	$100\%$	✓	$13\times$	$13\times$	$100\%$	✓	$27\times$	$28\times$
PATS + $\mathtt{ZFP}$ [tol= $1e\text{-1}$ ]	$37\%$	✓	$13\times$	$120\times$	$55\%$	✓	$27\times$	$52\times$
Sparse sampling + FT	$20\%$	✗	$12\times$	$60\times$	$20\%$	✗	$471\times$	$2457\times$
Sparse sampling + LoRA	$20\%$	✗	$47\times$ [r $:32$ ]	$235\times$	$20\%$	✗	$3744\times$ [r $:16$ ]	$18720\times$

Table 1: Performance of

\mathtt{ANTIC}

compared to baselines on 2D Kolmogorov flows and 3D BSSN simulation trajectories . We compare (i) Uniform sparse sampling (every 5th step), (ii) Dense sampling (every timestep), combined with (i) traditional compression (

\mathtt{ZFP}

), (ii) CFT, or (iii) CFT via LoRA fine-tuning. For comparison to

\mathtt{ZFP}

we match reconstruction quality to

\mathtt{ANTIC}

and compare for compression ratio. TR (Temporal Retention) denotes the percentage of the simulation duration preserved and Physics-awareness (PA) denotes whether or not the temporal selector uses a PDE specific metric. SC (Spatial Compression) and TC (Total Compression) represent the memory reduction factors.

\mathtt{ANTIC}

consistently outperforms ZFP and different temporal sampling schemes. In Sec. F , we report the performance of neural compression (CFT, CFT

+

LoRA) against well-known scientific computing classical compressors like SZ3 , MGARD , Blosc-2 zstd , clearly indicating the extreme compression obtained via neural fields for 3D high-resolution tensor-valued BSSN evolution simulations.

Panel of Figure 3 — Figure 3: Qualitative results for Physics-aware Adaptive Temporal Selection. (Left) The enstrophy flux (top) identifies turbulent regions in 2D Kolomogorov flows, hence is well suited as metric for our PATS. As a consequence, PATS densely samples at the start of the trajectory that exhibits strong chaotic behavior, whereas sampling more sparsely in non-turbulent regions. (Right) The Weyl scalar (top) isolates the non-linear gravitational wave (GW), resulting in dense sampling around the merger region and sparse sampling otherwise.

Method	Mean Rel. $\ell_{2}$ (5 snapshots)	Mean MAE (5 snapshots)	Mem. /snapshot
Scratch	$1.7\text{e{-3}}\pm 8.5\text{e{-5}}$	$0.20\pm 0.026$	1.50 MiB ( $12\times$ )
Full FT	$6.1\text{e{-4}}\pm 8.0\text{e{-6}}$	$0.061\pm 0.011$	1.50 MiB ( $12\times$ )
LoRA 8	$6.9\text{e{-2}}\pm 6.3\text{e{-3}}$	$0.63\pm 0.09$	98 KiB $(188\times)$
LoRA 16	$2.1\text{e{-2}}\pm 2.5\text{e{-3}}$	$0.13\pm 0.02$	196 KiB $(94\times)$
LoRA 32	$6.8\text{e{-4}}\pm 7.0\text{e{-5}}$	$0.063\pm 0.011$	392 KiB ( $47\times$ )
LoRA 64	$\pagecolor{background}\mathbf{4.3\text{e{-4}}\pm 2.4\text{e{-5}}}$	$\pagecolor{background}0.041\pm 0.006$	784 KiB ( $23\times$ )

Table 2: Accuracy metrics vs memory footprints for 2D Kolmogorov flows. Different spatial compression schemes evaluated across five sequential snapshots in a highly-turbulent regime. The last column reports the memory compression per snapshot relative to the 17 MiB ground truth. Continual fine-tuning using LoRA establishes an accuracy-memory pareto front.

Method	Mean Rel. $\ell_{2}$ (5 snapshots)	Mem. / snapshot	Time / snapshot
Scratch	$6.2\text{e{-4}}\pm 4.2\text{e{-5}}$	1.52 MiB ( $471\times$ )	$\sim 162$ (s)
Full FT	$\mathbf{4.6\textbf{e{-4}}\pm 7.6\textbf{e{-5}}}$	1.52 MiB ( $471\times$ )	$\sim 162$ (s)
LoRA 8	$6.0\text{e{-4}}\pm 4.1\text{e{-5}}$	$\mathbf{98\ KiB\ (7489\times)}$	$\sim 131$ (s)
LoRA 16	$5.7\text{e{-4}}\pm 3.9\text{e{-5}}$	196 KiB ( $3744\times$ )	$\sim 133$ (s)
LoRA 32	$5.3\text{e{-4}}\pm 3.7\text{e{-5}}$	392 KiB ( $1872\times$ )	$\sim 137$ (s)
LoRA 64	$4.6\text{e{-4}}\pm 3.2\text{e{-5}}$	784 KiB ( $936\times$ )	$\sim 149$ (s)

Table 3: Accuracy metrics vs memory footprints for 3D BBH merger simulations. Different spatial compression schemes evaluated across five sequential snapshots within the black hole merger regime. For completenees, we report wallclock-time per snapshot in the last column for the different training schemes.

Appendix figures & tables21 assets

Supplementary material from the paper’s appendix.

Appendix

Hardware	Method	Wallclock (s)	Epochs	MSE
H200	BSSN Solver (GPU)	$\sim 0.25-0.27$	–	–
	Cold Start	$\sim 183-187$	35	$\sim 3\times 10^{-8}$
	CFT (Ours)	$\mathbf{\sim 21-24}$	3	$\mathbf{\sim 3\times 10^{-8}}$
B300	BSSN Solver (GPU)	$\sim 0.27-0.44$	–	–
	Cold Start	$\sim 52-54$	35	$\sim 3\times 10^{-8}$
	CFT (Ours)	$\mathbf{\sim 4.5-4.7}$	3	$\mathbf{\sim 3\times 10^{-8}}$

Appendix

Table 4: Hardware Benchmark: Continual Fine-Tuning (CFT) vs. Solver vs. Cold-Start. Comparison of wall-clock convergence times across NVIDIA H200 and B300 (Blackwell) architectures for the 3D BSSN system. CFT updates perturbative changes between snapshots, achieving an order-of-magnitude speedup over cold-Starts while maintaining a comparable MSE of

3\times 10^{-8}

. The BSSN solver is a bit slower on B300s due to skinny batched GEMM — millions of tiny einsum-based contraction operations that are lowered to dot_general on cuBLAS for which below the Tensor Core tile threshold, rendering the solver memory-bandwidth-bound rather than compute-bound.

Compressor	2D Kolmogorov (16.8 MiB/Snapshot)			3D BSSN (0.7 GiB/Snapshot)
	CR ( $\uparrow$ )	Time ( $\downarrow$ )	Rel $\ell_{2}$ ( $\downarrow$ )	CR ( $\uparrow$ )	Time ( $\downarrow$ )	Rel $\ell_{2}$ ( $\downarrow$ )
ZFP (tol=1e-03)	$8\times$	$\mathbf{0.06}$ (s)	$\mathbf{1.02}$ e-4	$4\times$	$\mathbf{2.82}$ (s)	$\mathbf{1.27}$ e-4
MGARD [CPU] (abs=1e-03)	$19\times$	$0.37$ (s)	$4.02$ e-4	$5\times$	$38.67$ (s)	$7.97$ e-4
SZ3 (abs=1e-03)	$\mathbf{154\times}$	$0.07$ (s)	$8.31$ e-4	$16\times$	$3.42$ (s)	$1.02$ e-4
BLOSC-2 zstd (lossless)	$1.7\times$	$1.38$ (s)	$0.0$ e+00	$1.3\times$	$54.48$ (s)	$0.0$ e+00
BLOSC-2 zstd (lossy trunc=10)	$5\times$	$1.03$ (s)	$4.23$ e-4	$2.8\times$	$37.82$ (s)	$3.93$ e-4

Appendix

Table 5: Performance Benchmarks of Traditional vs. Neural Compression . Comparison of Compression Ratio (CR), Wall-clock Time, and Relative

\ell_{2}

Error across 2D Kolmogorov flows vorticity field (

2048^{2}

) in the turbulent regime and large-scale 3D BSSN evolved variables (

\sim 0.7

GiB/snapshot,

(213^{3},18)

shaped array) in the premerger phase. While traditional compressors (SZ3, ZFP, MGARD and Blosc-2) maintain low latency for small-scale 2D data, our NC with (CFT) particularly the LoRA variant achieves an unprecedented

3745\times

compression ratio for tensor-valued BSSN variables. This represents a two-order-of-magnitude improvement over state-of-the-art error-bounded compressors (SZ3) while maintaining comparable reconstruction fidelity (

\sim 10^{-4}

), effectively enabling the storage of high-resolution multi-scal simulations with a minimal memory footprint on the persistent storages. The neural field compression numbers reported are conducted on the B300 GPU.

Figure 6: Enstrophy flux reconstruction for 2D Kolmogorov flow. ( Left ) Enstrophy flux extracted from ground truth snapshots (solid blue) and

\mathtt{ANTIC}

reconstructions (dashed orange) over the peak turbulent phase

\tau\in[0,15]

(timesteps

[0,400]

\mathtt{ANTIC}

faithfully reproduces the temporal evolution of the enstrophy flux, including sharp transient features associated with turbulent intermittency. ( Right ) Pointwise absolute error over the same interval, remaining uniformly small with no systematic drift or error accumulation, confirming the temporal stability of the CFT-based continual learning scheme over long-horizon trajectories.

Figure 7: CFT reconstruction fidelity over long-horizon 2D Kolmogorov flow trajectories. Relative

\ell_{2}

error of

\mathtt{ANTIC}

reconstructions plotted against timestep for the full 2D Kolmogorov flow simulation. The trajectory spans two dynamically distinct regimes: a peak turbulence phase over timesteps

[0,100]

, characterized by rapid vorticity production, strong enstrophy cascade, and sharp small-scale features that place maximal demand on the neural field parametrization, followed by a gradual decay phase over timesteps

[100,400]

as the flow relaxes towards a quasi-steady turbulent state. Despite the elevated reconstruction difficulty during the high-turbulence regime — where inter-snapshot variability

\Delta u:=\|u(\mathbf{x},t+\Delta t)-u(\mathbf{x},t)\|

is largest — the CFT-based continual learning scheme maintains a stable relative

\ell_{2}

error

<4\times 10^{-3}

in the highest turbulence region and

<2\times 10^{-3}

, with no systematic degradation across the trajectory. This confirms that perturbative weight editing between subsequent neural field parametrizations remains well-conditioned even under large dynamical variability, demonstrating the robustness of

\mathtt{ANTIC}

for in-situ compression over long-horizon simulation runs.

Quantity	Metric	Value
Enstrophy Flux	Mean Absolute Error	$1.830\times 10^{-4}$
Enstrophy Flux	Max Absolute Error	$2.870\times 10^{-4}$
Vorticity Field	Mean Relative $\ell_{2}$ Error	$1.587\times 10^{-3}$
Vorticity Field	Max Relative $\ell_{2}$ Error	$3.669\times 10^{-3}$

Appendix

Table 6: Reconstruction Fidelity Metrics. Error analysis for Enstrophy Flux (scalar) and the Vorticity Field (2D grid). The low maximum relative

L_{2}

error (

0.37\%

) demonstrates that PATS successfully captures the most physically informative transients in the Kolmogorov flow.

Figure 9: Weyl scalar magnitude reconstruction during the BBH merger phase. The gravitational wave signal

|\Psi_{4}(t/M)|

extracted from ground truth snapshots (solid) and

\mathtt{ANTIC}

neural field reconstructions (dashed) at extraction radius

r=14.3\,M

, over the merger phase

t/M\in[136,210]

. This interval encompasses the peak gravitational wave emission, ringdown onset, and the most dynamically complex regime of the spacetime evolution, placing maximal demand on the fidelity of the neural field parametrization.

\mathtt{ANTIC}

accurately reproduces the amplitude and phase evolution of the Weyl scalar throughout, including the characteristic amplitude peak at merger.

Figure 10: Absolute reconstruction error in Weyl scalar magnitude. Pointwise absolute error

\big||\Psi_{4}^{\text{model}}(t/M)|-|\Psi_{4}^{\text{GT}}(t/M)|\big|

at extraction radius

r=14.3\,M

over

t/M\in[136,210]

. Errors remain uniformly small across the full merger phase, with no systematic growth near the amplitude peak, confirming that

\mathtt{ANTIC}

preserves the gravitational wave content of the spacetime to high fidelity under the CFT-based continual learning scheme.

Panel of Figure 11 — Figure 11: Lapse function reconstruction quality across the BBH merger trajectory. Each row corresponds to a representative timestep $t/M\in\{136.30,\ldots,209.22\}$ , spanning the inspiral-to-merger transition and post-merger ringdown phase of the 3D BBH simulation at spatial resolution $213^{3}$ . First column: Ground truth lapse function $\alpha(\mathbf{x},t)$ evolved under the BSSN framework (Sec. B.2 ), exhibiting the characteristic collapse toward zero in the black hole interior regions as the merger proceeds. Second column: $\mathtt{ANTIC}$ neural field reconstructions at the corresponding timesteps, obtained by querying the continually fine-tuned implicit parametrization. Third column: Pointwise relative $\ell_{2}$ error between the ground truth and reconstructed lapse fields, confirming that $\mathtt{ANTIC}$ faithfully captures both the smooth exterior spacetime geometry and the sharp gradients near the apparent horizons, with no systematic spatial concentration of errors in the dynamically critical merger region.

Category	Metric	NeF (Ours)	Ground Truth
$\|\Psi_{4}\|$ Range	Minimum Value	$2.825\times 10^{-5}$	$3.381\times 10^{-5}$
$\|\Psi_{4}\|$ Range	Maximum Value	$6.267\times 10^{-4}$	$6.741\times 10^{-4}$
Metric	Quantity		Value
Error	Mean Absolute Error		$1.835\times 10^{-5}$
Error	Max Absolute Error (at $t/M=170.49$ )		$4.838\times 10^{-5}$

Appendix

Table 7: BSSN Reconstruction Performance. Error analysis for the Weyl Scalar magnitude

|\Psi_{4}|

during the Binary Black Hole (BBH) merger and initial post-merger regime (

t/M\in[136.30,209.22]

). The results compare 146 snapshots across the peak gravitational wave emission phase.

Temporal Selector	Resolution	Retained Frames	Retention (%)
Enstrophy (PATS)	256	429	42.90
	512	427	42.70
	1024	427	42.70
	2048	375	37.50
JSD	256	3	0.30
	512	3	0.30

Appendix

Table 8: Temporal selector retention analysis across resolutions for 2D Kolmogorov flows. Comparison of snapshot retention rates across temporal selectors and spatial resolutions

\{256^{2},512^{2},1024^{2},2048^{2}\}

for a trajectory of 1000 snapshots. Retention is computed as

(selectedframes/1000)×\times×100\%

. The enstrophy-based PATS selector retains 37.5%–42.9% of snapshots consistently across all resolutions, reflecting a physically meaningful and resolution-stable selection that concentrates retained frames in high-turbulence transient regimes while discarding redundant snapshots during quasi-static evolution. In contrast, the physics-agnostic selectors exhibit two distinct failure modes: under-retention , where JSD (0.30% across all resolutions), Spectral (2.0%–2.7%), and Momentum (2.8%–5.5%) selectors discard the vast majority of snapshots indiscriminately, failing to identify dynamically critical transients; and over-retention , where MI retains all 1000 snapshots at every resolution (100%), providing no compression of the temporal dimension whatsoever. The Residual selector exhibits resolution-dependent instability, retaining 31.4%–39.4% at coarser resolutions but collapsing to full retention (100%) at

2048^{2}

, indicating sensitivity to the spectral content of high-resolution fields that renders it unreliable for in-situ deployment. Collectively, these results demonstrate that physics-agnostic selectors fail to generalize across resolutions and dynamical regimes, whereas PATS provides stable, physically grounded temporal compression that adapts to the underlying simulation dynamics.

Figure 18: Solver latency vs. neural compression scaling for 2D Kolmogorov flows. Wall-clock time per snapshot plotted against spatial resolution $N^{2}$ for the traditional Navier-Stokes solver and $\mathtt{ANTIC}$ neural compression (NC) on a single NVIDIA H200 GPU. Solver latency scales super-linearly with resolution, consistent with $\mathcal{O}(N^{2})$ mesh-point complexity under CFL constraints, while NC training time grows sub-linearly due to the resolution-invariant capacity of the underlying implicit neural field. The narrowing NC-to-solver ratio.

Resolution ( $N^{2}$ )	Total Pixels	Solver ( $dt$ )	NC (per snapshot)	NC/Solver Ratio
$256^{2}$	65,536	3.4 ms	24.90 s $\pm$ 1 ms	$\sim$ 7323 $\times$
$512^{2}$	262,144	7.3 ms	28.37 s $\pm$ 4 ms	$\sim$ 3886 $\times$
$1024^{2}$	1,048,576	28 ms	42.02 s $\pm$ 16 ms	$\sim$ 1500 $\times$
$2048^{2}$	4,194,304	0.2 s	62.19 s $\pm$ 22 ms	$\sim$ 310 $\times$
$2800^{2}$	7,840,000	1.3 s	75.67 s $\pm$ 41 ms	$\sim$ 58 $\times$

Appendix

Table 9: Hardware Benchmarks: Solver Latency vs. Neural Compression (NC). Wall-clock training time per snapshot (

dt

) for 2D Navier-Stokes Kolmogorov flows. All benchmarks were performed on a single NVIDIA H200 CUDA GPU. The scaling ratio (

\text{NC}/\text{Solver}

) demonstrates that while NC has a higher absolute latency, its complexity scales sub-linearly relative to the exponential growth of the traditional solver.

Partial Differential Equations

See all 52 papers

EVOLVE: Efficient Learned Volume Compression with Variable-Rate Encoding on a Cross-Domain Database
Jul 20, 2026Kaiyuan Tang, Maizhe Yang, Chaoli WangAutoencoders
Department of Computer Science and Engineering, University of Notre Dame, Notre Dame, IN 46556, USA
Physics-Preserving Latent Compression for Zero-Shot Resolution Transfer in 3D Turbulence
Jun 19, 2026Yilong Dai, Yiming Sun, Yiheng Chen +4Variational Autoencoders Physics-Informed ML
The University of Alabama · University of Pittsburgh · University of Maryland, College Park
NeuralZip: Reusable Setup for Fast Lossless Compression
Oct 7, 2026Martín Bravo, Samuel Horváth, Gonzalo Navarro +1Lossless Data Compression Neural Network Compression
Mohamed bin Zayed University of Artificial Intelligence, Abu Dhabi, UAE · Department of Computer Science, University of Chile, Santiago, Chile · National Center for Artificial Intelligence (CENIA), Santiago, Chile

Jul 20, 2026J/K move · Enter open · S save

EVOLVE: Efficient Learned Volume Compression with Variable-Rate Encoding on a Cross-Domain Database

Kaiyuan Tang, Maizhe Yang, Chaoli Wang

Department of Computer Science and Engineering, University of Notre Dame, Notre Dame, IN 46556, USA

Autoencoders

Open paper AskPDF arXiv

Large-scale scientific simulations generate volumetric data at rates that far outpace advances in storage and network bandwidth, making effective lossy compression increasingly critical. However, conventional compressors often struggle to preserve fine structural details at high compression ratios (CRs), and implicit neural representations (INRs) require costly per-volume optimization and produce models with fixed CRs. To respond, we present EVOLVE, an autoencoder (AE)-based volume-compression framework that targets high CRs for offline compression, with three key contributions. First, we construct a large-scale cross-domain database of 6,376 volumes from 21 scientific simulations, curated via perceptual hashing to ensure diversity, enabling the optimized model to extract features that generalize across volumes within the covered scientific simulation domains. Second, we reexamine the design space of AE-based compressors and incorporate several macro- and micro-designs into a vanilla AE to develop EVOLVE, which substantially improves the expressive power and compression capability. Third, we develop a learnable gain mechanism with a three-stage training strategy to enable variable-rate encoding, allowing a single model to support continuous CR adjustment at inference time. Experiments on multiple unseen scientific simulation datasets demonstrate that EVOLVE achieves substantially higher CRs than conventional compressors at comparable reconstruction quality, while delivering compression speeds that are orders of magnitude faster than INR-based methods, highlighting its promise as a strong alternative for compressing scientific data. The code, model weights, and results are available on our project page at https://evolve-vis.github.io.

Abstract

Partial Differential Equations

Explore similar work

EVOLVE: Efficient Learned Volume Compression with Variable-Rate Encoding on a Cross-Domain Database

Physics-Preserving Latent Compression for Zero-Shot Resolution Transfer in 3D Turbulence

NeuralZip: Reusable Setup for Fast Lossless Compression