PaperlayerView exact revision on arXiv

Source-linked AI summary

ANTIC: Adaptive Neural Temporal In-situ Compressor

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

arXiv:2604.09543v3cs.LG

TL;DR

Transient PDE simulations generate petabyte-to-exabyte data volumes, while existing in-situ compression may miss physics-driven temporal behavior and non-stationary spatial structure. ANTIC combines physics-aware temporal selection with continual neural-field compression of adjacent-snapshot residuals. It reports up to 435× compression for 2D Kolmogorov flow and 6807× for a 3D BBH trajectory.

  • Problem

    Transient PDE simulations produce petabyte-to-exabyte data, while in-situ methods may ignore physics-driven temporal behavior and fixed spatial representations struggle with non-stationary multiscale dynamics.

    Source: §1 · p. 1

  • Method

    ANTIC combines a physics-aware temporal selector with continual fine-tuning of neural fields to encode residual updates between adjacent snapshots.

    Source: §1 · p. 2

  • Results

    Up to 435× compression is reported for 2D Kolmogorov turbulence and 6807× for a 3D BBH trajectory.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Varying residual-update rank exposes an accuracy–memory Pareto front that users can traverse according to their needs.

    Source: §1 · p. 2

    PATS uses PDE-specific physics metrics, and generalizable PDE-agnostic temporal selectors that retain sensitivity to critical transients remain an open problem.

    Source: §6 · p. 9

Abstract

from arXiv · show

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.

1. Introduction

ANTIC addresses the storage burden of large transient PDE simulations by combining physics-aware temporal selection with neural spatial compression. It reports large spatiotemporal reductions on 2D Kolmogorov flows and a 3D binary black hole merger.

Motivation

  • Petabyte-to-exabyte transient datasets strain HPC storage, while offline compression is infeasible and fixed sampling or spatial representations can miss multiscale dynamics.PDE-aware online compression is needed to avoid archiving full trajectories while addressing temporal and spatial reduction together.Source: §1 · p. 1

Contribution

  • ANTIC combines a physics-aware selector that accepts or rejects snapshots using physics-of-interest and preceding context with neural compression of selected snapshots.The pipeline operates in situ across temporal and spatial axes.Source: §1 · pp. 1–2, Figure 1 · p. 2
  • Continual fine-tuning encodes residuals between adjacent snapshots in neural-field weights, with update rank exposing an accuracy–memory Pareto front.The residual updates can use full fine-tuning or low-rank updates.Source: §1 · p. 2

Results

  • Up to 435× compression on 2D Kolmogorov turbulence and 6807× on a 3D BBH trajectory are reported.For Kolmogorov flow, 62% fewer snapshots are retained alongside 47× spatial compression per snapshot; for BBH, temporal reduction is ∼45% and spatial compression reaches 3744× per snapshot.Source: §1 · p. 2

2. Background

The background frames in-situ compression as a response to trajectory-scale storage demands and reviews temporal selection, spatial compression, parameter-efficient updates, and video codecs. ANTIC is distinguished by physics-aware non-uniform selection paired with continual neural-field compression.

2.1. Temporal Frame Selection Methods

  • Heuristic temporal selectors identify salient snapshots but can miss system-specific physical behavior because they ignore governing physics.ANTIC instead uses measures sensitive to the physics of interest for each system.Source: §2.1 · p. 2

2.2. Spatial Compression Techniques

  • Traditional transform, tensor, and early autoencoder methods reduce volume, while neural fields represent grid functions in compact differentiable weights and support continuous queries.Neural-field compression has been reported across climate, relativistic, and plasma simulations.Source: §2.2 · pp. 2–3

2.3. Parameter-efficient Fine-tuning (PEFTs)

  • LoRA uses trainable low-rank factors to reduce full-update costs, and ANTIC uses LoRA-style fine-tuning to offer an accuracy–complexity trade-off.Its background describes the low-rank update assumption and cites prior evidence of accuracy comparable to full fine-tuning [Xin et al. (2024)].Source: §2.3 · p. 3

2.4. Codecs & (Neural) Video Compression

  • Offline codecs exploit temporal redundancy, but their storage demands make offline compression infeasible for trajectories reaching hundreds of terabytes or beyond.Video codecs and neural video compression use inter-frame dependencies, while computer-vision keyframe selectors do not consider physics of interest.Source: §2.4 · p. 3
  • ANTIC selects snapshots non-uniformly using physical saliency and pairs this with continual neural-field fine-tuning, unlike MGARD’s uniform temporal output frequency.The passage reports higher compression ratios than classical lossy error-bound compressors at equivalent fidelity for time-evolved simulations.Source: §2.4 · p. 3

2.5. In-situ Compression

  • Existing in-situ frameworks may lack adaptive temporal selection or neural compression, whereas ANTIC combines physics-tailored selection with a user-adjustable spatial accuracy–memory trade-off.The cited methods are described as offering modest compression rates or using fixed-sized latents that are not resolution invariant.Source: §2.5 · p. 4

3. Methodology

ANTIC combines a physics-aware temporal selector with spatial neural compression. Its selector uses a rolling context and adaptive window, while continual fine-tuning encodes inter-snapshot residual updates with adjustable rank.

3. Methodology

  • PATS comprises a Metric, Regulator, Queue, and Gate, and is paired with spatial neural compression that traverses a rate-distortion trade-off.The Queue stores metric context; the Regulator adapts its window, and the Gate decides whether to compress a new snapshot.Source: §3 · p. 4, §3.1 · p. 4

3.1. Physics-aware Temporal Selector

  • PATS computes a PDE-specific physics metric, maintains a sliding window of metric values, and uses a dynamic threshold to decide whether to retain a snapshot.After selection, the regulator truncates the window so the selected snapshot becomes the new reference anchor; the Gate also feeds back a signal to adjust the next window size.Source: §3.1 · p. 4

3.2. Spatial Neural Compression

  • Continual fine-tuning initializes a neural field from the preceding snapshot and learns the residual for the next snapshot as an incremental weight update.LoRA factorizes the update at rank r; restricting rank reduces memory but may increase distortion, while full fine-tuning reduces distortion at greater memory cost.Source: §3.2 · pp. 4–5

4. Experiment Setup

The experiments use 2D Kolmogorov flow to stress-test temporal selection and a high-resolution 3D BBH merger to stress-test spatial compression. They evaluate sampling and compression variants across both simulations.

4. Experiment Setup

  • The 2D Kolmogorov benchmark has 1000 time steps, 17 MiB per snapshot, and a 16.8 GiB trajectory; enstrophy is its physics metric for PATS.The flow transitions from high-frequency initial conditions to decaying turbulence, where small vortices coalesce; enstrophy serves as a proxy for turbulence magnitude and fast/slow evolution.Source: §4 · p. 5
  • The 3D BBH trajectory contains 5,966 snapshots of 18 evolved variables and 4.2 TiB of raw data, with the Weyl scalar used to select salient snapshots.The study uses a GPU-accelerated numerical-relativity solver and treats the simulation as a stress test for extreme spatial compression.Source: §4 · p. 5

4.1. NF Implementation Details

  • The neural-field base model is a 256 × 6 SiLU coordinate MLP with ∼0.4M parameters; Fourier features are used for Kolmogorov flow and LayerNorm plus weight decay address continual-update instability.Training uses FLOAT32 by default, and LoRA ranks are swept.Source: §4.1 · pp. 5–6

4. Experiment Setup

  • Table 1 compares sparse and dense sampling combined with ZFP, continual fine-tuning, or LoRA, tracking temporal retention, physics-awareness, spatial compression, and total compression.For ZFP, reconstruction quality is matched to ANTIC when comparing compression ratios.Source: Table 1 · p. 6

5. Results

ANTIC combines physics-aware temporal selection with continual neural-field fine-tuning to compress both snapshot retention and spatial data. Results cover 2D turbulent flows, 3D binary black hole mergers, and training-time feasibility.

5.1. 2D Kolmogorov flow simulations

  • ANTIC variants achieve the highest temporal coherence in 2D Kolmogorov flows, with 4× greater spatial compression than ZFP and 37% temporal retention.Dense sampling retains 100% of snapshots but has lower total compression; Sparse retains 20% but misses informative transients.Source: §5.1 · p. 6
  • PATS reduces 2D temporal retention to 37% while preserving physical fidelity, sparsifying decaying turbulence after T > 10 and retaining non-stationary transients.The authors report that physics-agnostic selectors fail to identify dynamically critical transients in these PDE trajectories.Source: §5.1 · pp. 6–7
  • For 2D neural compression, CFT is 30% more accurate than training from scratch; LoRA at r = 32 matches CFT accuracy with 4× less parameter storage.LoRA at r = 64 improves Mean Rel. ℓ2 and MAE by 30% over CFT and reaches a 23× compression factor relative to the 17 MiB snapshot.Source: §5.1 · p. 7, Table 2 · p. 7

5.2. 3D BBH merger simulations

  • For the BBH case, ANTIC retains activity near the merger with 55% temporal retention and reports 17×–65× data-volume reduction compared with PATS+ZFP.PATS uses the Weyl-scalar metric and sparsifies pre- and post-merger phases.Source: §5.2 · p. 7

5.3. Neural Compression Scalability and In-Situ Feasibility

  • CFT reduces BBH per-snapshot training time from approximately 45 s to 4.5 s, a speedup exceeding 10×, by learning residual updates between successive snapshots.The authors identify resolution-decoupled neural-field scaling and inter-snapshot efficiency as support for in-situ use with large-scale mesh-intensive PDE simulations.Source: §5.3 · p. 9

6. Limitations.

The authors identify latency, derivative fidelity, temporal-metric generality, and off-grid evaluation as limitations or open areas.

6. Limitations.

  • Despite reducing compression time by more than 10× versus cold-start training, the solver remains about an order of magnitude faster; real-time in-situ deployment needs lower training latency.Standard MLPs also lack the inductive bias to preserve high-order derivatives, and derivative supervision adds training overhead.Source: §6 · p. 9
  • PATS metrics are tailored to the PDE, generalizable PDE-agnostic selectors remain open, and off-grid neural-field accuracy has not been systematically validated.The reported evaluation optimizes relative ℓ2 error at grid-coincident coordinates.Source: §6 · p. 9

7. Conclusion

ANTIC combines physics-aware temporal selection with continual fine-tuning of neural fields, achieving high compression ratios while maintaining spatial fidelity in two simulation cases. The authors identify HPC-scale deployment and learned temporal selection as future directions.

7. Conclusion

  • ANTIC achieves compression ratios exceeding 400× for turbulent 2D Kolmogorov flows and 10,000× for 3D BSSN binary black hole simulations while maintaining spatial fidelity.The conclusion also reports a 10× reduction in training wall-clock time relative to cold-start methods.Source: §7 · p. 9
  • Continual fine-tuning uses perturbative weight updates between neural fields, reducing per-snapshot encoding cost and training wall-clock time by 10× versus cold starts.The framework combines spatial neural compression with a physics-aware temporal selector tailored to PDE trajectories.Source: §7 · p. 9
  • Future work targets HPC-scale deployment with data and pipeline parallelism, and reinforcement learning for a more universal temporal-selection policy.The proposed deployment settings include computational fluid dynamics, gyrokinetics, and astrophysics.Source: §7 · p. 9

Impact Statement

ANTIC addresses storage and data-management burdens in large-scale PDE simulations through in-situ temporal selection and spatial compression. Its selectors use physics-specific indicators and adaptive rules to retain salient snapshots while controlling operational overhead.

A. Physics-aware Temporal Selector suite

  • The temporal-selector suite uses domain-specific physical invariants and a regulator–decision-maker feedback loop to adapt sliding-window size and select salient snapshots online.The selectors are designed for 2D Kolmogorov flows and 3D BSSN binary black hole mergers; the framework states that the decision engine is not learned.Source: A. Physics-aware Temporal Selector suite · p. 15

A.1. 2D Kolmogorov flows

  • The 2D regulator adjusts its stability factor using enstrophy variation, extending selection windows in stable regions and increasing subsampling during physically violent transitions.A rolling history buffer and median baseline are used to estimate quiescent activity in the 3D merger selector.Source: §A.1 · p. 16, §A.2 · p. 16
  • For 2D flows, snapshots are bypassed only when enstrophy divergence remains bounded and structural correlation with the reference frame meets its threshold.The correlation criterion is intended to preserve spatially complex, temporally persistent features even when temporal variance is marginal.Source: §A.1 · pp. 15–16

Impact Statement

  • In-situ neural compression has O(T) time complexity and O(Q) constant memory overhead, with operational constraints for volumetric simulations containing billions or trillions of mesh cells.The impact statement notes that compression-induced errors must remain acceptable for downstream analysis.Source: Impact Statement · p. 10, §A.1 · p. 16

A.2. 3D BSSN BBH Merger Simulations

  • For BBH mergers, look-ahead checks clamp temporal jumps at detected surges and switch to unit-stride sampling before the transient.A gated history update avoids incorporating extreme merger transients into the baseline, retaining sensitivity to salient physics.Source: §A.2 · p. 16
  • ANTIC’s streaming loop applies temporal selection, fetches selected snapshots, and stores either full fine-tuning or low-rank adapter weight updates.The algorithm returns both the selection set and compressed updates.Source: §A.2 · p. 18

B.1.1. EXPERIMENTAL SETUP, DATA PREPARATION AND NF TRAINING SPECIFICS

The experimental setup uses decaying 2D turbulence as a multiscale benchmark and describes the numerical-relativity formulation used for binary black hole simulations. It also explains how the Weyl scalar Ψ4 is extracted and used as a physics-aware activity metric.

B.1.1. EXPERIMENTAL SETUP, DATA PREPARATION AND NF TRAINING SPECIFICS

  • The 2D Kolmogorov benchmark uses a 2048 × 2048 periodic domain and 1,000 time steps, producing a 16.8 GiB trajectory of decaying turbulence.The flow evolves without external forcing, and its fine-scale eddies and rapid transients challenge neural-field compression.Source: §B.1.1 · p. 18

C. Numerical Relativity and the BSSN Formulation

  • BSSN evolves conformal variables and uses conformal transformations to improve numerical stability and constraint preservation in binary black hole simulations.The setup presents the 3+1 ADM split into lapse, shift, and induced spatial metric.Source: C. Numerical Relativity and the BSSN… · p. 19

C.1. Extraction of Gravitational Radiation: The Weyl Scalar Ψ4

  • The Weyl scalar Ψ4 is reconstructed from electric and magnetic Weyl components and represents outgoing gravitational radiation at the extraction boundary.ANTIC uses the magnitude of Ψ4 as its BBH physics-aware metric to identify high-frequency merger and ringdown transients.Source: §C.1 · p. 19

C.2. Experimental Configuration

The BBH simulation spans 250M over 5,966 steps on a Cartesian grid, with 18 tensor components per selected snapshot used for neural compression.

C.2. Experimental Configuration

  • The BBH setup uses a 2133 Cartesian grid over [−15M, 15M], evolves for 250M across 5,966 steps, and trains on 18 tensor components per selected snapshot.The setup excludes gauge-dependent shift vectors but retains the lapse function for visualization.Source: §C.2 · p. 19

D. Extended Pareto-Frontier over Model Size Sweeps

The extended Pareto-frontiers compare cold-start, continual fine-tuning, and LoRA-based continual fine-tuning across model sizes for 2D Kolmogorov flows and 3D BBH simulations. They report comparable reconstruction accuracy with improved compression from LoRA variants.

D. Extended Pareto-Frontier over Model Size Sweeps

  • Across model-size sweeps from 32 × 2 to 512 × 8, continual fine-tuning reaches accuracy comparable to cold-start training for high-resolution BSSN simulations.The extended sweep expands on the 256 × 6 model presented in the main paper.Source: D. Extended Pareto-Frontier over Model… · p. 20
  • Across the highly-turbulent window and near-merger regime, LoRA attains error comparable to continual fine-tuning while significantly improving compression, at ranks r > 32 and r ≥ 8, respectively.The reported metric is averaged relative ℓ2 error over 5 subsequent frames.Source: D. Extended Pareto-Frontier over Model… · p. 20

E. Benchmarking Continual Learning Against Cold-Start Methodologies

ANTIC uses continual fine-tuning to encode changes between adjacent snapshots, compares neural compression with classical compressors, and evaluates reconstruction of flow and merger diagnostics. Across these evaluations, the reported results cover compression, training cost, and preservation of physical quantities.

E. Benchmarking Continual Learning Against Cold-Start Methodologies

F. Comparison against Other Classical Compressors in Scientific Computing

  • Unlike snapshot-wise classical compressors, ANTIC's neural-field representation supports point-wise queries without decompressing a full block or subgrid.The comparison discusses SZ3, ZFP, and MGARD as snapshot-wise compressors; Blosc-2 is described as lossless.Source: F. Comparison against Other Classical… · p. 21
  • 3745× compression for tensor-valued BSSN variables is reported for neural compression with CFT, with comparable reconstruction fidelity of ∼10−4.The reported ratio is a two-order-of-magnitude improvement over SZ3 for the large-scale premerger data.Source: Table 5 · p. 22

G. Global Reconstruction Quality and Physics Preservation

  • ANTIC maintains per-snapshot training MSE of 10−7–10−9 across long-horizon runs and evaluates derived quantities in both Kolmogorov-flow and BBH simulations.The reported diagnostics include normalized enstrophy flux, constraint violations, and geometric quantities.Source: G. Global Reconstruction Quality and… · p. 22

G.1. 2D Kolmogorov flows Enstrophy Flux Reconstruction

  • The reconstructed enstrophy flux follows ground truth through the turbulent phase, while relative ℓ2 error remains below 4 × 10−3 in the highest-turbulence region.The paper identifies enstrophy flux as a sensitive probe of preservation of the flow's multiscale vortical structure.Source: §G.1 · p. 22, Table 6 · p. 22, Figure 6 · p. 23

G.2. 3D BBH Merger Weyl Scalar Magnitude Reconstruction

  • ANTIC reproduces Weyl-scalar amplitude and phase through the BBH merger, and its lapse reconstructions capture sharp gradients near apparent horizons without concentrated spatial error.The Weyl-scalar assessment covers t/M ∈ [136, 210]; the lapse examples span inspiral through post-merger ringdown.Source: Figure 9 · p. 25, §G.2 · p. 25, Figure 11 · p. 26

H. Comparing Temporal Retention of Physics-Agnostic selectors against PATS

The section compares physics-aware temporal selection with physics-agnostic alternatives, neural-compression scaling with solver latency, and dynamic with binary stride regulation. It also documents the computational hardware and software used.

H. Comparing Temporal Retention of Physics-Agnostic selectors against PATS

  • PATS retains 37.5%–42.9% of snapshots across resolutions, while physics-agnostic selectors either discard most snapshots, retain all of them, or become unstable at high resolution.The enstrophy-based selection concentrates retained frames in high-turbulence transients; the residual selector reaches full retention at 2048².Source: Table 8 · p. 29, H. Comparing Temporal Retention of… · p. 30

J. Dynamic Adaptive Regulator vs Binary Adaptive Regulator

  • The dynamic regulator retains 37% of snapshots versus 53% for the binary regulator, whose lack of intermediate stride choices leads to oversampling in moderate-activity regions.The dynamic regulator maps saliency to strides from 1 to 5; the binary option switches between strides 1 and 5.Source: Figure 19 · p. 31, J. Dynamic Adaptive Regulator vs… · p. 31

K. Hardware and Licenses

  • Experiments used dual-socket Intel Xeon systems for CPU tasks and NVIDIA H200 or B300 GPUs for simulation and neural-field workloads.The implementation used JAX with CUDA support and Flax NNX, Orbax, and Optax.Source: K. Hardware and Licenses · p. 31

Related papers

Loading 2604.09543v3…