ARC-STAR: Auditable Post-Hoc Correction for PDE Foundation Models

Chengze Li, Lingwei Wei, Li Sun, Hongbo Lv, Jie Yang, Hanrong Zhang, Kening Zheng, Wei-Chieh Huang, Enze Ma, Philip S. YuarXiv 2026 · arXiv

TL;DR

ARC-STAR introduces an auditable, budget-aware post-hoc correction framework for pretrained PDE foundation models that dramatically reduces simulation drift without altering the base model. By combining a global bias corrector with a targeted, blockwise local refiner, it routes compute exclusively to high-risk spatial regions, cutting velocity rollout errors by at least 36× across diverse fluid flow benchmarks.

What Problem Does This Paper Address?

Pretrained Partial Differential Equation (PDE) foundation models act as general-purpose neural solvers for physical fields like fluid velocity and pressure. However, when deployed on unfamiliar flow regimes, small one-step errors compound autoregressively over long simulation horizons, causing massive "rollout drift."

Fine-tuning the base model for every new scenario is computationally expensive and often destabilizes the network. Existing post-hoc correction methods apply uniform fixes across the entire spatial domain, which wastes compute because simulation errors are rarely distributed evenly—they naturally concentrate around turbulent structures like vortex cores and shear layers.

Figure 1: Post-broad-correction residual is spatially concentrated, not uniform. Each cell is partitioned into 64 blocks of 16×16 pixels; tile color is per-block error share, with per-panel annotation reporting the share carried by the highest-error 20% of blocks (mean per-cell Gini coefficient 0.48, between the uniform null and the sparse-activation regime). Block ranking is an offline diagnostic; the deployed selector is label-free.

Methods

The authors propose ARC-STAR (Adaptive Risk-Calibrated Spatial Triage for Auditable Refinement), a multi-stage framework that keeps the base model completely frozen:

  1. Stage I: Global Corrector ($G_\phi$): A full-field residual operator (built using a lightweight Tensor Fourier Neural Operator, or TFNO-lite) that eliminates broad, systematic solver bias.
  2. Stage II: Local Refiner ($L_\theta$): A blockwise neural network (ConvNeXt-style) that cleans up the remaining localized errors. It uses a "halo-read, center-write" contract, meaning it looks at a $32\times32$ pixel neighborhood (context) but only updates the central $16\times16$ block. Crucially, it uses an outer-product Hann window to blend block boundaries and prevent artificial discontinuities.
  3. Stage III: Spatial Triage: At deployment, a label-free risk score (combining velocity innovation and kinetic-energy gradients) ranks spatial blocks. Compute can then be dynamically routed to only the top-$k$ highest-risk blocks, making the framework highly budget-aware.

Because the stages are decoupled, the framework is fully auditable, allowing exact mathematical tracking of how much error was removed by the global vs. local components.

Overview of the ARC-STAR framework showing global training, local training, and inference triage stages.

Datasets

The framework was rigorously evaluated on five 2D incompressible flow families from the public Poseidon benchmark suite:

  • NS-G: Smooth fields with Gaussian initial conditions.
  • KF: Mixed-scale stationary turbulence with Kolmogorov forcing.
  • NS-SL: Shear layers developing Kelvin-Helmholtz instabilities.
  • NS-PwC: Piecewise-constant initial vorticity creating errors at vortex boundaries.
  • NS-Sines: Sinusoidal initial conditions testing wavenumber leakage.

Each family was tested under both moderate (m) and extreme (x) viscosity regimes (10 distinct evaluation cells). Additionally, the authors cross-validated the method using a structurally distinct DPOT-Ti host model on the PDEArena NS-Gauss dataset.

Key Insights and Ideas

  • Errors Don't Homogenize: Even after a powerful global correction layer cleans up broad biases, the remaining residual error stays highly concentrated in specific physical structures rather than dissolving into uniform statistical noise.
  • Host Independence: This spatial error concentration tracks physical turbulent mechanisms (like vortex cores) rather than artifacts of a specific model's pretraining. The exact same concentration patterns appear whether using Poseidon or DPOT as the base solver, allowing a single hand-designed routing rule to transfer across models without per-cell tuning.
  • The Power of Blending: Independent block processing introduces subtle boundary mismatches. Without Hann-window blending to taper block edges, these tiny discontinuities explode under autoregressive rollouts, making the corrected model perform significantly worse than the raw uncorrected host.

Key Results

  • Massive Error Reduction: ARC-STAR was the only evaluated method to cut velocity rollout error by at least 36× over raw Poseidon across all ten benchmark cells.
  • Stage-wise Performance: The global corrector alone wiped out 91–99% of the raw host error. The local refiner then stepped in to reduce the remaining post-global residual by up to 94.4%.
  • Budget Efficiency: Under tight compute constraints (e.g., refining only 25% of the spatial blocks), ARC-STAR's label-free routing policy consistently traced the lowest error frontier compared to nine external routing baselines (such as uncertainty- or saliency-based methods).

Limitations and Future Work

Currently, ARC-STAR is restricted to two-dimensional, incompressible Navier-Stokes equations on periodic grids, focusing exclusively on velocity-channel corrections. The authors note that extending the framework to compressible flows, three-dimensional spaces, non-periodic boundary conditions, and longer rollout horizons represents the next logical frontier for post-hoc foundation model correction.

Summary updated

Abstract

Partial differential equation (PDE) foundation models are pretrained networks that forecast how physical fields like velocity and pressure evolve from a single reusable solver. On unfamiliar flows their predictions drift step by step, errors concentrate in a few regions, yet retraining destabilizes the network and uniform post-hoc correction overlooks this spatial concentration. To address this, we propose a frozen-solver post-hoc correction framework, Adaptive Risk-Calibrated Spatial Triage for Auditable Refinement (ARC-STAR). ARC-STAR organizes correction into three stages: a global corrector removes broad solver bias, a blockwise local refiner cleans the post-global residual, and, at deployment, a label-free score routes refinement to high-risk blocks under a compute budget. The framework is designed to be (i) frozen-host, preserving the pretrained solver without fine-tuning; (ii) auditable, with global and local stages trained and evaluated separately for measurable contributions; and (iii) budget-aware, using a blockwise interface that either refines the full field or routes limited compute to high-risk regions. Across five flow benchmarks spanning ten regime cells, ARC-STAR is the only method that cuts velocity rollout error by at least 36x over raw Poseidon on every cell. The global stage reduces raw host error by 91-99%, and the local stage further reduces the remaining post-global residual by up to 94.4%.
BibTeX
@misc{li2026arc,
  author = {Chengze Li and Lingwei Wei and Li Sun and Hongbo Lv and Jie Yang and Hanrong Zhang and Kening Zheng and Wei-Chieh Huang and Enze Ma and Philip S. Yu},
  title = {{ARC-STAR: Auditable Post-Hoc Correction for PDE Foundation Models}},
  year = {2026},
  eprint = {2605.22222},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2605.22222},
}
Figure