PaperlayerView exact revision on arXiv

Source-linked AI summary

Structured Extrema Errors in Classical Surrogates for Viscous Burgers: A Physics-Consistent Interpretation

Youssef Oubari

arXiv:2609.07952v1cs.LGphysics.flu-dyn

TL;DR

Global error measures do not show where surrogate errors occur or whether they have repeatable local structure. The paper analyzes raw-grid Burgers surrogates through local residual geometry and tests links to PDE terms. It finds extrema-local branches across models, but evidence for insufficient viscous smoothing is strongest for KRR and Ridge at moderate and high viscosity.

  • Problem

    Global MSE and relative L2 measures summarize prediction error over the domain but do not identify where the largest errors occur or whether they have repeatable local structure.

    Source: §1 · p. 1

  • Method

    The study models local discrepancies through amplitude, shift, and smoothing, analyzes curvature and fold geometry, and tests Burgers-term explanations on held-out trajectories.

    Source: §1 · p. 2

  • Results

    Extrema-fold geometry organizes residual branches across models, while KRR and Ridge at moderate and high viscosity show evidence consistent with insufficient smoothing and tree diffusion attribution is weaker.

    Source: §1 · p. 2, §5.7 · p. 13

  • Takeaways & Limitations

    The diagnosed structure carries predictive information and illustrates a route from an observed surrogate failure to a testable physical interpretation.

    Source: §6 · p. 14

    The study covers one one-dimensional PDE and one raw-grid representation, so the geometric pattern should not yet be claimed as universal.

    Source: §5.7 · p. 13

Abstract

from arXiv · show

We study the local errors of classical machine-learning surrogate models, which approximate the time evolution of the one-dimensional viscous Burgers equation. Four models are compared on the same prediction task, using the spatial grid values directly: radial basis function (RBF) kernel ridge regression (KRR), linear Ridge, ExtraTrees, and Random Forests. Across all four models, the one-step residual, defined here as the true value minus the predicted value at each grid point, forms clear curved branches near predicted maxima and minima. A more detailed analysis of KRR shows that these errors are much more strongly related to the second spatial derivative, which measures local curvature, than to the first spatial derivative. Near a smooth extremum, predicted value and curvature form a local two-branch fold. Under our local curvature-based model of the residual, this fold predicts a leading-order near-parabolic relation between predicted value and residual. This geometric result motivates a direct test of the Burgers advection (transport) and diffusion (smoothing) terms. For KRR and Ridge, regression tests on held-out trajectories, a control that breaks the spatial alignment of the diffusion term, and a spectral test of high-frequency content are consistent with insufficient viscous smoothing at moderate and high viscosity. In this case, the surrogate retains more small-scale structure than the true future state. The same physical explanation is much weaker for the tree models. Finally, a correction that uses only predicted quantities reduces both one-step error and error during recursive rollout, where each prediction is used as the next input.

1 Introduction

The introduction asks whether surrogate prediction errors have repeatable local structure that can be linked to governing PDE terms. It studies raw-grid classical models for viscous Burgers, then connects extrema-local residual patterns to curvature, diffusion tests, and prediction correction.

1 Introduction

  • The study asks whether local prediction-error structure can be linked to governing PDE terms, since global MSE and relative L2 measures do not show where errors occur or whether they repeat locally.The motivating question builds on links between numerical error and differential operators.Source: §1 · p. 1
  • The comparison uses RBF KRR, linear Ridge, ExtraTrees, and Random Forests on the same discretized grid values and viscosity.This shared input setting provides a controlled comparison of error structure.Source: §1 · p. 1
  • For KRR, curvature is more strongly associated with residuals than slope, and a local fold near smooth extrema predicts a leading-order near-parabolic value–residual relation.The fold combines with the curvature-based residual model to explain the predicted geometry.Source: §1 · pp. 2–3
  • For KRR and Ridge, held-out regression, a spatial negative control, and spectral tests are consistent with insufficient viscous smoothing at moderate and high viscosity; the tree models show weaker diffusion attribution.In this regime, KRR and Ridge retain more small-scale structure than the true future state.Source: §1 · pp. 2–3
  • Across four classical surrogate families, one-step residuals form curved branches near predicted maxima and minima, with branch size and orientation varying by model.The pointwise residual is the true future value minus the prediction.Source: §1 · pp. 1–3
  • A correction trained only on predicted-state quantities reduces KRR error in both one-step prediction and recursive rollout.The test includes held-out one-step predictions and rollout, where each prediction is used as the next input.Source: §1 · pp. 2–3

2 Related Work

Related work uses PDE and numerical-analysis structure to interpret or correct model errors. This paper instead begins with the prediction residual’s local structure and tests its connection to governing-equation terms.

2 Related Work

  • Prior approaches connect numerical errors to differential operators and use physical structure to estimate or correct surrogate errors.Modified-equation analysis and PDE-residual-based correction are cited as relevant lines of work.Source: §2 · p. 3
  • Scientific interpretability work emphasizes quantitative tests and checks against domain knowledge.The paper applies this framing to residual structure and governing-PDE terms.Source: §2 · p. 3
  • The study differs by starting from prediction residuals, analyzing their local structure, and testing which governing-PDE terms explain it across regimes and held-out trajectories.The analysis also uses negative controls and error correction.Source: §2 · p. 3

3 Problem Setting and Method

The study evaluates raw-grid surrogates for viscous Burgers prediction and defines procedures for examining their residuals, local geometry, and predictive errors. It compares four models and uses KRR as the reference for detailed analyses.

3.1 Burgers equation and prediction task

  • The prediction task uses a numerical flow map for the one-dimensional viscous Burgers equation, with grid-sampled reference states as inputs to raw-grid surrogates.One-step predictions start from true input states.Source: §3.1 · p. 3

3.2 Data and surrogate models

  • The evaluation compares RBF KRR, Ridge, ExtraTrees, and Random Forests; KRR is strongest in the canonical comparison and anchors the detailed analyses.The equation is posed on a periodic domain, and the canonical data set has 30 trajectories split into 24 training and 6 validation trajectories.Source: §3.1 · p. 3, §3.2 · p. 4

3.3 Prediction residual and predicted extrema

  • The analysis defines pointwise residuals, detects predicted extrema on the grid, and uses radius-3 masks and 11-point radius-5 windows for local geometry.Active windows follow a 90th-percentile rule.Source: §3.3 · p. 4

3.4 Local geometry

  • Local windows are summarized by PCA branch angle and elongation, then conic fits are labeled ellipse-like, hyperbola-like, or parabola-like using a normalized discriminant.The conic fits use predicted value and residual as their two coordinates.Source: §3.4 · pp. 4–5

3.5 Evaluation protocol

  • Prediction error is measured with mean squared error and relative L2 error; physics attribution uses trajectory-wise cross-fitting and bootstrap resampling of whole trajectories.The attribution analysis cross-fits over all 30 trajectories, so each residual comes from a surrogate not trained on that trajectory.Source: §3.5 · p. 5

4 Geometric Analysis

The geometric analysis models local surrogate error through amplitude, spatial shift, and smoothing differences. Near a smooth, non-degenerate extremum, the model yields a leading-order fold and a quadratic predicted-value–residual relation, while finite-window fits can differ from that local shape.

4.1 A first-order residual model

  • The local error model separates amplitude, spatial shift, and smoothing differences, and is intended only as a local description of those possible errors.The analysis applies a first-order approximation near detected extrema to motivate testing curvature there.Source: §4.1 · pp. 5–6

4.2 Fold geometry near an extremum

  • Under smoothness and non-degeneracy assumptions, predicted value and curvature form a two-branch local fold near an extremum.The fold relation is a leading-order result.Source: §4.2 · p. 6
  • Under the curvature-based residual model, the fold gives a leading-order quadratic, parabolic relation between predicted value and residual; finite-window fits may instead appear ellipse-like or hyperbola-like.Higher-order terms, finite windows, discretization, and fitting error can affect empirical conic fits, which do not always recover the local shape.Source: §4.2 · p. 6

5 Experiments and Interpretation

Across four classical surrogates, residuals form structured branches near extrema, while KRR and Ridge show evidence consistent with insufficient viscous smoothing at moderate and high viscosity. A prediction-only correction reduces KRR error in one-step prediction and recursive rollout.

5.1 Structured residuals appear across model families

  • Across all four model families, residuals form off-axis branches; KRR and Ridge stay nearer r = 0, while tree models have larger residuals and more r-axis-oriented branches.Active-branch counts were 67.3 ± 2.1 for ExtraTrees and 66.7 ± 1.5 for Random Forests across three seeds.Source: §5.1 · pp. 6–7, Figure 2 · p. 7, Table 1 · p. 7

5.2 The largest residuals are concentrated near extrema

  • The radius-3 near-extrema mask covers 23.9% of grid points but contains 39.6% of the largest 5% and 40.9% of the largest 1% of KRR residuals.Thus, extrema occupy a minority of points but contain a disproportionate share of the largest residuals.Source: §5.2 · pp. 7–8, Table 2 · p. 8, Figure 3 · p. 8

5.3 Testing the first-order residual model

  • In the KRR diagnostic regression, curvature was the strongest single first-order correlate of residuals, while the first derivative explained almost none; the full model explained more variance at extrema.The predicted value–curvature relation near extrema was consistent with the local fold geometry.Source: §5.3 · pp. 8–9, Table 3 · p. 9, Figure 4 · p. 9

5.4 Connecting the residual to the Burgers equation

  • For KRR and Ridge, held-out attribution showed a large contribution from aligned diffusion and a large spatial-shift gap, while advection contributed much less.ExtraTrees and Random Forests had much smaller diffusion contributions and shift gaps; the shifted control preserved diffusion values but broke pointwise alignment.Source: §5.4 · p. 10, Table 4 · p. 10

5.5 Effective diffusion and spectral roughness

  • At H = 0.025, the moderate/high-viscosity diagnostics were consistent with insufficient viscous smoothing: νeff/ν < 1 and QR > 1.As the horizon increased, νeff/ν moved toward one while the roughness ratio grew; the study assigns no specific cause to this divergence.Source: Table 5 · p. 10, §5.5 · p. 11

5.6 Can the identified structure help reduce the error?

  • The oracle local fit removed a median 85% of active-window residual energy, an upper bound because it uses the true residual during fitting.97% of active windows had more than half their residual energy removed.Source: §5.6 · p. 11, Figure 5 · p. 11
  • The prediction-only KRR correction reduced all-point MSE by 31.24% and 20-step relative L2 error from 0.935 to 0.616.Near-extrema MSE fell by 49.54% and active-window MSE by 45.22%; the largest peak errors remained the hardest cases.Source: §5.6 · pp. 11–13, Table 6 · p. 11, Table 7 · p. 13

5.7 Discussion and limitations

  • The interpretation is limited to one one-dimensional PDE and a raw-grid representation; low-viscosity results are less uniform, and directly transferred local conic fits generalize poorly.The authors treat the fold/extrema mechanism as the main geometric result and finite-window conics as descriptive diagnostics.Source: §5.7 · p. 13

6 Conclusion

The paper identifies structured extrema errors in classical raw-grid Burgers surrogates and links the KRR geometry to curvature, while physical tests indicate insufficient smoothing for KRR and Ridge at moderate and high viscosity. Non-oracle corrections reduce one-step and rollout error.

6 Conclusion

  • Across KRR, Ridge, ExtraTrees, and Random Forests, one-step residuals form coherent branches near predicted extrema; KRR curvature dominates transport, and physics tests indicate insufficient smoothing for KRR and Ridge at moderate and high viscosity.The tree ensembles show much weaker diffusion attribution.Source: §6 · p. 13
  • Non-oracle corrections using predicted quantities reduce both one-step error and error during recursive rollout.The diagnosed structure carries predictive information.Source: §6 · pp. 13–14
  • The results illustrate a practical route from an observed surrogate failure to a testable physical interpretation.Source: §6 · p. 14

A Full Experimental Setup

The study uses three distinct experimental protocols for canonical model comparison, physical diagnostics, and held-out conic tests. The canonical setup compares four models on a shared discretized grid, while the conic replication uses a fresh trajectory draw and split.

A Full Experimental Setup

  • Three protocols serve different purposes, so their absolute errors should not be compared across protocols.Source: A Full Experimental Setup · p. 15, Table 8 · p. 15
  • The large diagnostic protocol uses a one-frame saved-time horizon only for Sec. 5.3 regression, and its absolute MSE is not compared with the canonical MSE.Source: A Full Experimental Setup · p. 15
  • The canonical protocol uses 30 trajectories split 24/6, a one-step horizon H = 0.025, and a shared raw-grid input across KRR, Ridge, ExtraTrees, and Random Forests.It uses the periodic domain [0, 1), N = 128, and ∆x = 1/128.Source: A Full Experimental Setup · p. 15
  • The replication protocol uses a fresh draw of 30 trajectories and a fresh 24/6 split only for held-out conic tests in Appendix C.Its KRR validation MSE is 5.221 × 10^-3 and relative L2 error is 0.2157.Source: A Full Experimental Setup · p. 15
  • Active windows are radius-5 windows whose peak absolute residual exceeds the extremum-specific 90th-percentile threshold from radius-3 near-extrema points.Source: A Full Experimental Setup · p. 15
  • Within each active radius-5 window, the conic fit standardizes X = predicted value and Y = residual, then selects the smallest-singular-value right singular vector.The design rows use [X^2, XY, Y^2, X, Y, 1], with coefficients normalized before Eq. (11).Source: A Full Experimental Setup · p. 15

B Derivations and Supporting Calculations

The supporting calculations derive first-order operator and extremum-fold relations, justify comparing residual rates with Burgers terms, and explain the diffusion and roughness diagnostics. Appendices also distinguish the leading parabolic relation from finite-window conic behavior and test conic transfer.

B.1 First-order operator perturbation

  • First-order operator perturbation combines translation and heat-operator expansions, retaining terms linear in the small changes.Source: §B.1 · pp. 15–16

B.2 Fold relation and residual relation

  • At a smooth interior extremum, Taylor expansions yield the local value–curvature fold; under the first-order residual model, it implies the leading relation between predicted value and residual.If c3 = 0, the displayed fold is not the leading asymmetric relation, so the expansion must continue to the first nonzero odd derivative.Source: §B.2 · p. 16

B.3 Why the residual rate is compared with Burgers terms

  • The residual rate r/H compares reference and surrogate average rates of change, motivating endpoint regression features for Burgers advection and diffusion.Source: §B.3 · p. 17

B.4 Effective-diffusion diagnostic

  • The effective-diffusion diagnostic matches the leading diffusion coefficient under a local heat model and short-horizon regularity assumptions.Source: §B.4 · p. 17

B.5 Why the roughness diagnostic uses fourth-power frequency weighting

  • The roughness measure weights Fourier mode k by k^4, up to a common constant, because it normalizes second-derivative energy by total Fourier energy.Source: §B.5 · p. 18

C.1 Finite-window conic classes

  • Empirical finite-window conic classes include higher-order effects, whereas the leading-order relation is parabolic; representative fits can follow a local branch or select a mismatched conic type.The classifications are summarized in Table 9.Source: §C.1 · p. 18, Table 9 · p. 18, Figure 8 · p. 19

C.2 Held-out conic transfer

  • In the independent replication test, conics fit on 6 of 11 points and evaluated on the remaining 5 show a large held-out gap, indicating poor transfer as prediction rules.Source: §C.2 · p. 19, Table 10 · p. 20

D Full Physical Sweep

The physical sweep supports an insufficient-smoothing interpretation only in moderate- and high-viscosity regimes; low-viscosity results are less coherent. The pooled regression is distinct from per-viscosity estimates and can yield an effective-diffusion value outside their range.

D Full Physical Sweep

  • The pooled effective-diffusion value can lie outside the per-viscosity range because pooled coefficients are jointly fitted rather than averaged from separate regressions.The main-text physical interpretation uses separate moderate- and high-viscosity results.Source: D Full Physical Sweep · p. 20
  • At ν = 0.01 and H = 0.025, KRR gives νeff/ν > 1, so the study restricts its insufficient-smoothing interpretation to moderate and high viscosity.The low-viscosity rows are described as less coherent.Source: D Full Physical Sweep · p. 20

E Correction Details and Ablations

The profile–spectral method is the main KRR correction, with the larger tail-aware ensemble giving almost the same result. The recursive-rollout analysis gives a first-order account of how one-step residuals propagate, while its figures show residual structure during rollout.

E Correction Details and Ablations

  • The correction removes a fraction E of active-window residual energy, with E = 1 meaning all energy is removed and E = 0 meaning squared energy is unchanged.Source: E Correction Details and Ablations · p. 21
  • The non-oracle correction uses sign-canonicalized radius-5 predicted-state profiles, local summaries, and viscosity, with separate models for maxima and minima.Its profile predictor uses gradient boosting; overlapping windows are center-weighted, and spectral calibration is learned from training data only.Source: E Correction Details and Ablations · p. 21
  • The profile–spectral method is the main correction because the larger tail-aware ensemble gives almost the same result.The correction metrics are evaluated over the full canonical validation set; the single-state corrected snapshot is qualitative.Source: E Correction Details and Ablations · p. 21

F Recursive Rollout Details

  • The first-order rollout error accumulates propagated earlier errors and new one-step residuals, with second-order remainders omitted.The displayed three-step form is e3 ≈ J2J1r1(û0) + J2r1(û1) + r1(û2).Source: F Recursive Rollout Details · pp. 22–23
  • During rollout, the near-zero residual cloud becomes less dominant while structured off-axis branches remain visible.Source: Figure 11 · p. 23
  • Figure 12 compares plain and corrected residuals on the fixed active-window set during recursive rollout.The supplied caption specifies the comparison but not its outcome.Source: Figure 12 · p. 23

G Additional Robustness Checks

The cross-model conclusion is not driven by a single random tree seed. Finite-window conic labels are mostly stable to trimmed refits, but the transfer results support treating fitted conics as descriptive local diagnostics.

G Additional Robustness Checks

Related papers

Loading 2609.07952v1…