PaperlayerView exact revision on arXiv

Source-linked AI summary

QoI-Aware Provisional Rollout and Retrospective Reconciliation for Reduced-State Scientific Twins

Liangji Zhu, Scott Klasky, Jaemoon Lee, Qian Gong, Anand Rangarajan, Sanjay Ranka

arXiv:2608.29633v1cs.CE

TL;DR

Scientific twins need an immediate trajectory during synchronization gaps and a way to revise that history once future boundary data arrives. This paper introduces deterministic temporal bridging and analytic energy matching for reduced-state twins, improving field and gradient-sensitive fidelity within the evaluated turbulence regime.

  • Problem

    Synchronization gaps require scientific twins to provide provisional outputs and later revise their history, while field and derived-quantity errors can degrade at different rates.

  • Method

    The method uses a closed-form residual bridge followed by calibration-based energy matching with smooth regional gains and global rescaling, without an additional correction network or reference fields inside the gap.

  • Results

    At S = 8, full reconciliation reduces window-averaged NRMSE by about 60% for both components and lowers vx global gradient-intensity error from 4.21% to 2.91%, versus 20.40% for future-aware physical interpolation.

  • Takeaways & Limitations

    Provisional continuity is available online, while future-conditioned reconciliation improves historical fidelity and enforces the boundary-inferred global energy target within the evaluated turbulence setting.

  • Takeaways & Limitations

    Reduced state is an input assumption rather than an evaluated systems benefit because runtime, memory, hardware, and energy measurements are unavailable.

Abstract

from arXiv · show

Scientific twins may need to continue operating when updates from an authoritative primary system are temporarily unavailable. Once synchronization resumes, the new boundary can also be used to revise the intervening history. We distinguish an immediately available causal provisional trajectory from a delayed, future-conditioned reconciled trajectory. For reduced-state twins, we introduce a deterministic, calibration-based reconciliation method. A smooth temporal bridge carries the residual observed at the next synchronization block backward through the provisional interval. An analytic energy-matching stage then applies smooth regional gains and a global rescaling to match a component-energy trajectory estimated by cubic regression in log-energy space from synchronized frames on both sides of the gap. The method uses no additional correction network and revises decoded history without changing the latent state used for later rollouts. We evaluate 64 spatial patches from 16 JHTDB isotropic-turbulence slices for both velocity components and gaps S in {4, 6, 8}. During the longest gap, field error and gradient-sensitive QoI error degrade at markedly different rates, so field error alone does not characterize provisional fidelity. At S = 8, full reconciliation reduces window-averaged NRMSE by about 60% for both components and global gradient-intensity error from 4.21% to 2.91% for vx, whereas future-aware physical interpolation reaches 20.40% on the same metric. Energy matching additionally makes the reconciled history match its boundary-inferred global energy trajectory exactly. Future boundary information therefore substantially improves scientifically relevant properties within the evaluated regime.

I. INTRODUCTION

The paper separates immediate causal operation from delayed history revision in reduced-state scientific twins, then develops a deterministic reconciliation method and evaluates QoI-aware recovery across synchronization gaps.

  • Operating problem: Scientific twins must provide a provisional trajectory during unavailable primary updates and may revise that history after synchronization resumes.Online consumers use provisional states, while buffered or archival consumers can replace them with reconciled states.
  • Research focus: The study asks how finite synchronization gaps affect scientifically important quantities and how later primary boundaries improve decoded history.It focuses on trajectory fidelity rather than compression or deployment cost.
  • Evaluation: Field error alone does not characterize fidelity because gradient-sensitive QoI error degrades at a markedly different rate within JHTDB turbulence gaps.The evaluation considers both velocity components and synchronization gaps S ∈ {4, 6, 8}.
  • Method: The method uses a closed-form residual bridge followed by analytic energy matching, without hidden primary fields from the repaired interval or an additional correction network.The bridge corrects temporal drift, while regional gains and global rescaling control component energy.
  • Positioning: The paper positions retrospective reconciliation as distinct from forward rollout remedies and future-aware interpolation because it first emits a causal trajectory, then revises its finite history.The broader study does not evaluate a complete deployed twin or decision loop.

C. Finite-Interval Smoothing and Adjacent Systems Axes

The paper frames reconciliation as deterministic finite-interval smoothing under block synchronization, distinct from stochastic data assimilation and separate from compression or delayed-state replacement.

  • Finite-interval smoothing: The bridge uses future boundary residuals to smooth earlier decoded fields through the gap with a precomputable deterministic linear operator.It assumes the correction varies smoothly in time and does not estimate uncertainty, process noise, or a Bayesian posterior.
  • Adjacent systems axes: Compression addresses retained-field bit usage, not the trajectory-fidelity objective or evaluation criterion studied here.The paper treats compression as a separate systems axis.
  • Block synchronization: The synchronization schedule supplies K = 4 consecutive states before each withheld interval, followed by another primary block.The evaluated gaps are S ∈ {4, 6, 8}, and the four-state block follows the predictor’s context length rather than an optimality claim.
  • Scope of repair: Retrospective repair requires a later authoritative boundary; if missed states were buffered and later delivered, they should directly replace the provisional interval instead.The method targets permanently unrecorded interior states whose boundary blocks survive.
  • State provenance: The synchronization interface retains latent states rather than corresponding full-resolution authoritative fields, while reconciliation uses representation-consistent decoded boundaries.Authoritative provenance refers to the primary state, whereas the decoded boundary is the physical interface available to reconciliation.

B. Reduced State and Two Trajectory Products

A frozen reduced-state surrogate produces an immediate causal rollout, while decoded-field reconciliation later revises that history without changing the latent state used for subsequent rollouts.

  • Reduced state: A frozen CAESAR variational autoencoder maps synchronized primary fields to reduced latent states and decodes them back to physical fields.The representation has encoder E and decoder D, with latent dimension d < N.
  • Provisional trajectory: The latent predictor recursively advances four-state contexts over the gap to produce a provisional trajectory available without waiting for synchronization.The predictor maps four consecutive latent states to the next four.
  • Reconciliation: After the trailing block arrives, its residual relative to the rollout defines a retrospective correction for each position in the provisional interval.This correction is applied after resynchronization and therefore cannot change actions already taken from provisional states.
  • Two trajectory products: Reconciliation operates only on decoded physical fields and neither produces corrected latent states nor alters the state used for the next autonomous rollout.The paper therefore calls this trajectory reconciliation rather than twin-state assimilation.
  • Error scope: The repair targets temporal prediction drift relative to the fixed representation, not VAE reconstruction error.Reported errors still include reconstruction effects because evaluation uses authoritative fields.
  • Evaluation boundary: Because no operational error threshold is defined, the evaluation reports error versus interval position and gap length rather than a certified safe horizon.The analysis asks how resynchronization changes QoI error and how results depend on S.

C. Physics-Based Quantities of Interest

The study evaluates velocity-component energy and longitudinal gradient intensity as complementary QoIs, using regional maps and global statistics to capture both intensity and fine-scale variation.

  • Study setting: The QoIs are defined separately for vx and vy on JHTDB isotropic turbulence.The representative figure pairs the vx field with its energy and gradient maps.
  • Energy QoI: Component-wise velocity energy measures preservation of each velocity component’s intensity in global and regional forms.Regional quantities average over non-overlapping 16×16 blocks.
  • Gradient QoI: Longitudinal gradient intensity measures small-scale velocity variation and serves as a proxy for dissipative activity under isotropy.Its regional maps describe spatial distribution rather than local dissipation estimates.
  • QoI computation: Spatial derivatives use a second-order centered stencil without periodic wrap-around, with unit index spacing and complete regional blocks retained.The derivative support is 254×256 for Dx and 256×254 for Dy on a 256×256 patch.
  • Spatial fidelity: The gradient maps emphasize fine-scale filamentary structures that aggregate metrics do not resolve.Global fidelity uses relative error of the global statistic, while regional fidelity uses relative L2 error between maps.

IV. METHODOLOGY

The method combines a frozen VAE-based reduced-state predictor with a downstream reconciliation procedure that corrects decoded history using trailing synchronization residuals. The causal predictor operates in latent space, while reconciliation applies a deterministic temporal bridge in physical space without altering later latent rollouts.

  • Latent predictor: The predictor uses learned per-position, per-channel trend coefficients and horizon-specific learned residual heads to jointly predict the next four latent states.Longer forecasts recursively feed each predicted four-state block back into the same predictor.
  • Operational separation: The causal stage remains strictly future-blind, whereas reconciliation operates entirely downstream of the decoder and never feeds repaired fields back into the predictor or latent sequence.Thus provisional output is the only available output during synchronization loss, while reconciliation revises decoded history later.
  • Boundary residual: Reconciliation compares decoded rollout fields with frozen-VAE reconstructions of resumed synchronized latent states to form a representation-consistent synchronization residual.This residual measures temporal prediction drift while excluding the VAE reconstruction residual.
  • Boundary residual: A closed-form quadratic smoother maps trailing boundary residuals backward through the provisional interval, penalizing unnecessary correction and temporal curvature.The resulting matrix Wλ depends only on repair geometry and regularization, so it can be precomputed and applied as a temporal matrix product.
  • Latent predictor: Four input frames are independently encoded by a frozen VAE, and the latent history plus first- and second-order differences feed trend and correction paths.The paths produce four predicted latent states that the frozen decoder maps to physical fields.

C. QoI-Aware Energy Matching and Validation Selection

The QoI-aware correction estimates component-energy trajectories from synchronized frames, applies smooth regional amplitude gains, and globally rescales reconciled fields. Calibration fixes the smoother and correction strength before test inference, with no additional learned repair network.

  • Energy matching: The energy-matching stage directly targets component energy, while improvements in field and gradient error remain empirical consequences rather than enforced properties.Calibration choices are fixed using training and held-out validation data, and targets use frozen-VAE reconstructions.
  • Calibration and inference: Validation selects one gap-specific regularization and correction-strength pair shared across windows and both velocity components, but it cannot protect any individual test interval.Inference uses the precomputed parameters and closed-form equations without a learned correction network.
  • Energy target estimation: Cubic regression in log-energy space uses synchronized energy samples before and after the gap to estimate positive regional and global target trajectories.The regression is fit independently for each 16 × 16 region and uses all selected synchronized samples rather than only two endpoints.
  • Spatial modulation: Smooth regional gains are represented on a 16 × 16 coarse map, bilinearly upsampled in log space, and exponentiated into a continuous spatial gain field.This avoids artificial discontinuities from independently scaling spatial blocks.
  • Global energy lock: A final frame-wise scalar corrects the remaining global mismatch after regional modulation, making the final field match the safeguarded boundary-inferred global target up to numerical precision.Reported global-energy error instead measures the accuracy of that inferred target against withheld authoritative energy.

D. Operational Semantics

Reconciliation is a delayed, future-conditioned product whose availability depends on the trailing synchronization block. It revises decoded history while preserving the newly received primary-derived latent block as the seed for the next rollout.

  • Availability: The reconciled version of gap position ℓ becomes available after S+K−ℓ time steps when the trailing synchronization block ends at τj + S + K.The method uses all K states in that trailing block, so version and provenance should be labeled explicitly.
  • State continuity: Reconciliation does not change the latent state that seeds the next autonomous interval, which begins directly from the newly received primary-derived K-state latent block.Under block synchronization, this avoids an operational penalty from repaired latent feedback.
  • Scope: Corrected-state feedback becomes relevant only during extended outages spanning multiple gaps without resynchronization.The stated operational semantics therefore apply to intervals followed by an authoritative trailing block.

E. Representation Boundary

The evaluation separates reduced latent-state rollout from decoded-field reconciliation and tests both across JHTDB turbulence patches, overlapping windows, and gaps of varying length. Metrics compare predictions with original physical references and include derived QoIs beyond NRMSE.

  • Representation boundary: The retrospective stage uses decoded fields and analytic operations without Fourier transforms, additional correction networks, or trained parameters.Runtime and peak-memory measurements are unavailable, so the study makes no system-level efficiency claims.
  • Evaluation setting: The dataset contains 64 spatial sequences from 16 JHTDB 512 × 512 isotropic-turbulence slices, with four patches per slice sharing the same timestamps.The four patches from each slice are not independent realizations.
  • Evaluation setting: Overlapping test windows use a four-frame stride, span the full gap with L = S, and are immediately followed by K = 4 synchronized states.Repair-window starts are defined over the test interval, while calibration excludes test windows and hidden physical references.
  • Metrics: All reported metrics compare decoded predictions with original physical references, including range-normalized RMSE and relative global and regional energy errors.Metrics are averaged per patch sequence and frame before position-wise and scalar aggregation.
  • Comparisons and questions: The study compares causal provisional rollout, future-aware physical linear interpolation, temporal smoothing, and full reconciliation under a common synchronization schedule.The evaluation asks how provisional QoI fidelity degrades, how much reconciliation removes that degradation, and how outputs vary with gap length S.

B. RQ1: Provisional QoI Drift Within a Synchronization Gap

During the longest tested gap, provisional decoded and QoI errors grow with position, with regional and gradient-sensitive quantities degrading more strongly than field NRMSE.

  • Provisional error growth: 0.002 to 0.023: decoded NRMSE rises across the eight positions of the S = 8 provisional interval.The evaluation averages 14 overlapping test intervals and 64 spatial patch sequences.
  • Provisional error growth: Below 3% to approximately 10%: global component-energy error stays lower than regional component-energy error by the interval’s final position.
  • Provisional error growth: 6% and 27%: global and regional longitudinal-gradient-intensity errors reach these approximate levels at the eighth position.
  • Provisional error growth: Distinct position-wise growth means NRMSE alone does not characterize component-energy and gradient-intensity fidelity.The reported metrics use different normalizations, so their numerical ratio has no direct physical meaning.

C. RQ2: Retrospective Reconciliation

Retrospective reconciliation revises provisional histories after synchronization returns, improving window-averaged fidelity across gaps and QoIs while retaining a distinct temporal pattern near the trailing boundary.

  • Temporal behavior: Reconciled mean error rises initially and falls toward the trailing synchronization boundary, where the observed residual provides strongest information.This is a mean-curve pattern, not a bound or monotonicity guarantee for individual intervals.
  • Temporal behavior: Provisional error grows through each unshaded interval, then the reconciled state replaces it after primary synchronization blocks return.The complete schedules show this repeating temporal pattern for vx across S = 4, 6, 8; vy mirrors the pattern in supplementary results.
  • Window-averaged recovery: At S = 8, full reconciliation records NRMSE of 0.0039 versus 0.0097 for provisional rollout and 0.0104 for interpolation.
  • Window-averaged recovery: At S = 8, global gradient error is 2.91% for reconciliation versus 4.21% for provisional rollout and 20.40% for interpolation.For vx, these are window-averaged means across the evaluated patch sequences and windows.
  • Window-averaged recovery: 38.5–60.3%: full reconciliation reduces vy NRMSE relative to provisional rollout as the gap grows from S = 4 to S = 8.The corresponding reductions increase monotonically with the gap.
  • Window-averaged recovery: 90.3–94.1% and 78.0–82.4%: full reconciliation reduces vy global- and regional-energy errors relative to provisional rollout across S = 4, 6, 8.

E. Effect of Energy Matching

Energy matching improves component-energy consistency and generally preserves the gains from temporal smoothing, while regional gradient error remains an exception. The evaluation separates online continuity from delayed archival revision and reports scope boundaries for the tested regime.

  • Global energy: 93.3%, 90.6%, and 88.3% reductions in authoritative-reference global-energy error for vx occur at S = 4, 6, and 8, respectively.Similar reductions are reported for vy; the comparison tests the boundary-derived target rather than constraint satisfaction alone.
  • Regional energy: At S = 8 for vx, full reconciliation changes regional energy error from 1.43% to 0.82%.Energy matching also improves NRMSE and global gradient error.
  • Uncertainty: At S = 8 for vx, reconciliation improves NRMSE by 0.00585 and global-gradient error by 1.30 percentage points.The paired patch-bootstrap 95% intervals are [0.00557, 0.00615] and [1.08, 1.53], respectively; only 2 of 64 patches degrade on global gradient.
  • Uncertainty: Cluster-bootstrap intervals for S = 8 vx exclude zero for global energy, NRMSE, regional energy, and global gradient error, but not the regional-gradient exception.The resampling accounts for dependence across the 16 physical z-slices.
  • System scope: The online provisional product provides causal continuity, while the reconciled product is a delayed finite-interval estimate for buffered and archival consumers.Reconciliation revises decoded history without changing the latent state used for subsequent autonomous rollouts.
Loading 2608.29633v1…