PaperlayerView exact revision on arXiv

Source-linked AI summary

Structure-preserving reduced-order models of thin panels in supersonic and hypersonic flow

Miguel A. Aguilo

arXiv:2610.31504v1cs.CEcs.LGcs.NAmath.NA

TL;DR

The AePW-4 challenge packages provide no time-resolved response or repeated runs, limiting what can be learned directly from data. The paper builds reference and learned reduced-order models that preserve energy structure, then tests them against available RC-19 and HyMAX evidence. The learned model matches several measured or published response findings, while some comparisons remain constrained by missing data.

  • Problem

    The challenge packages lack time histories and repeated runs, although the panels’ nonlinear flutter responses require models that can represent behavior beyond linear reduced-order models.

    Source: §1 · p. 2

  • Method

    A Ritz-Galerkin reference model supplies the exact potential, while a learned model fits generalized forces and trains only dissipation constrained not to add energy.

    Source: §1 · pp. 2, 4

  • Results

    The learned model reproduces the published RC-19 response regimes without its unsupported 113 Hz peak, places mean dominant frequency within 1.0% of measured 236 Hz, and tracks HyMAX reference displacement with 3.67% relative error over 64.9 ms.

    Source: §1 · p. 5, §12 · pp. 62–63

  • Takeaways & Limitations

    The construction matches available flutter evidence while keeping energy accounting exact, but experimental trajectory grading and training require time-series data the packages do not contain.

    Source: §7.7 · p. 54, §12 · pp. 62–63

    The packages provide no response time history or repeated runs, so learned-model training records come from the reference model and experimental comparisons rely on statistics.

    Source: §11 · p. 61

Abstract

from arXiv · show

We build structure-preserving reduced-order models for both challenge problems of the AePW-4 High-Speed Working Group. They are a clamped thin panel at Mach 1.92 (RC-19) and a cantilever plate at Mach 5.85 (HyMAX). The published data for each case hold no time-resolved response and no repeated runs. The physics must therefore be carried by the form of the model. One construction gives two models per problem. A Ritz-Galerkin reference model describes the plate. A learned model keeps the reference model's exact potential, fits the remaining generalized forces, and trains only a dissipation that cannot add energy. Its energy balance holds for any weights by construction. On the trained model the worst relative residual is 2.2e-15. On RC-19, shock relations predict a pressure ratio of 1.521 across the panel. Pressure-sensitive paint measured the same value. Thirty runs of the learned model place the mean dominant frequency within 1.0% of the measured 236 Hz. At one periodic and one chaotic condition, the learned model reproduces the response regime of a published reduced-order model in all sixteen runs. It does so without that model's 113 Hz spectral peak, which the measurements do not show. Measured displacements can also correct the chaotic response through one more force that respects the energy balance. We tested this on synthetic data at the experiment's 5 kHz sampling rate. The state error is 1.25% of the panel thickness without noise and 3.5% with 10 micrometers of noise. On HyMAX the learned model follows the reference model's trailing-edge displacement for 65 ms with a relative error of 3.7%. These results show that a model whose structure carries the physics can match measured flutter behavior when the available data cannot train it.

1. Introduction

The challenge packages lack time histories and repeated runs, so the model must carry the physics in its structure. The paper tests a structure-preserving learned model against reference models and available experimental statistics.

1. Introduction

  • The paper’s hypothesis combines measured-mode elastic energy, fixed least-squares non-conservative forces, trainable non-energy-adding dissipation, and discrete-gradient time stepping.This modal Hamiltonian reduction retains the earlier series’ commitments to exact energy accounting and dissipation that cannot add energy.Source: §1 · pp. 3–4
  • The RC-19 reference model uses seven measured mode shapes, membrane stiffness from in-plane elasticity, and a temperature field derived from rigid-panel pressures and calibrated thermocouples.The measured shapes are orthogonalized by Löwdin’s symmetric construction.Source: §1 · p. 4
  • The learned model preserves exact energy accounting for any weights, with a worst trained-model balance residual of 2.2 × 10^-15.Its learned dissipation is positive definite and can only absorb energy; the generalized aerodynamic force is fitted by least squares and frozen.Source: §1 · p. 4
  • The learned model uses the same discrete-gradient step as its training trajectories, preventing integration mismatch drift and preserving exact stepwise energy balance.The reported trained models were integrated with generalized-α; the discrete-gradient step applies unchanged.Source: §1 · p. 4
  • Assimilation adds a displacement-and-rate error force while keeping each time step’s energy balance exact; in a 5 kHz twin experiment, an offline smoother holds state error to 0.0125 thicknesses without noise.The nine measurement points observe all seven modes.Source: §1 · pp. 4–5
  • At the periodic case’s operating condition, a thirty-member trained-model ensemble averaged 238.4 Hz, 1.0% from measured 236 Hz; at Brouwer et al.’s simulation conditions, it reproduced the reduced-order model’s regimes without its 113 Hz peak.P2 was tested at two simulation operating conditions, with sixteen runs at each condition, eight from each well.Source: §1 · pp. 3, 5
  • The attached-shock displacement comparison remains open: the reference model gives 0.06 thicknesses against a measured 0.943, and testing the proposed remaining mechanism needs deformed-panel pressure data.The package lacks time-resolved surface pressure or a computation on the deformed panel.Source: §1 · p. 5
  • 3.67% relative error over 64.9 ms measures HyMAX tracking of the reference model’s trailing-edge displacement without feedback.The construction detects that a six-mode basis is unsupported because its whitened coordinates retain only eleven of twelve state directions.Source: §1 · p. 5

2. The challenge problems and what their packages hold

The two challenge packages describe different panel configurations and provide limited response data. RC-19 supplies measured modes and pressure information but no response statistics or time history in the package, while HyMAX provides pressure measurements and a first frequency but no displacement.

2. The challenge problems and what their packages hold

  • Both packages are used as distributed, and no number in the report is fitted to a measured flutter response.The stated structural calibrations are thermal effectiveness on the buckling temperature rise and HyMAX bending stiffness on the first frequency.Source: §2 · p. 6

2.1. AFRL/SSC RC-19 clamped thin panel

  • RC-19 is a clamped steel panel at Mach 1.92 with periodic, chaotic, and attached-shock cases; its package supplies seven measured modes, time-averaged rigid-panel pressures, and 5 kHz response measurement setup, but no response time history or published displacement statistics.The attached-shock case includes schlieren and pressure-sensitive paint; the planform marks nine extraction points.Source: §2.1 · p. 6, Figure 1 · p. 7, Figure 2 · p. 8

2.2. UNSW HyMAX cantilever plate

  • HyMAX is a Mach 5.85 cantilever plate with a shock generator; its package supplies one run, four pressure transducers, and the measured first frequency, but no displacement.The figure shows the two shock positions used for training trajectories and the two clamped-free modes of the reference model.Source: §2.2 · pp. 8–9, Figure 3 · p. 9

3. Background and related work

The background places the work among nonlinear panel-flutter models and structure-preserving learned models. The paper distinguishes its modal Hamiltonian construction from prior mesh-based reductions and penalty-based physical constraints.

3. Background and related work

  • Prior work identifies heated buckled panels as capable of chaotic flutter, while first-order piston theory is a standard aerodynamic model for panel flutter.The background also points to reviews of nonlinear panel flutter and aerothermoelasticity.Source: §3 · p. 10
  • Brouwer et al.’s RC-19 model reproduces published regimes but has a 113 Hz spectral peak absent from experiment and requires offsets in periodic-case cavity pressure and temperature rise.Their model uses finite-element modes, boundary-layer-corrected piston theory, static loading referenced to computed surface pressure, and Rayleigh damping.Source: §3 · p. 10
  • Other working-group analyses include RC-19 shock correlations and computational HyMAX modeling; a separate Mach 2.11 cantilevered panel study reports frequency agreement within 1 Hz.The Mach 2.11 panel is over a backward-facing step and is not the RC-19 panel.Source: §3 · p. 10
  • Unlike structure imposed by a penalty, the paper builds structure into the model parameterization so it holds for every parameter value; an average-vector-field discrete gradient carries exact energy accounting through each step.The port-Hamiltonian form accounts for flow work, which can have either sign.Source: §3 · p. 10

4. Notation

The notation table defines vector, matrix, time-derivative, and state conventions used throughout the report. It also names the variables for the learned model, assimilation, and reported statistics.

4. Notation

  • Bold lowercase letters denote vectors over retained modes, bold uppercase letters denote matrices, and calligraphic letters denote fourth-order tensors.A dot denotes a physical-time derivative; primes denote nondimensional-time derivatives except on the HyMAX beam, where they denote derivatives in x.Source: §4 · pp. 11–12
  • The learned-model notation distinguishes modal state and rate, conditioning parameters, fitted forces, trained dissipation, and rollout saturation quantities.The state is (W, W′); W is nondimensional modal displacement in thicknesses, and τ = ω1t.Source: §4 · p. 13
  • The statistics notation includes ensemble size, HyMAX relative root-mean-square error, trailing-edge displacement, pooled standard deviation, and equivalence tolerance.Assimilation notation identifies observed displacement readings, observation timing, and observer gains.Source: §4 · pp. 13–14

5. Methods

The methods use a physical reference model and a trained model whose force structure preserves the reference potential. The RC-19 reference model reduces plate mechanics to measured modes and accounts for thermal, membrane, and pressure effects in its Hamiltonian.

5. Methods

  • The reference model is assembled without training, while the learned model fits non-conservative forces and trains dissipation on generated reference-model trajectories.Agreement between the two levels tests the fitted structure-preserving model and integrator, not dynamics learned from experimental training data.Source: §5 · p. 14

5.1. The RC-19 reference model

  • The RC-19 Ritz–Galerkin reduction restricts plate deflection to seven measured modes and evaluates the continuum energies and virtual work in modal coordinates.The modes are interpolated from the package grid and orthogonalized in the mass inner product; the Hamiltonian represents the continuum total energy restricted to those modes.Source: §5.1 · p. 16
  • RC-19 potential energy combines measured-frequency bending stiffness, thermal preload, quartic membrane hardening, and static pressure work.The package supplies no material constants; thermal effectiveness is calibrated so the flat panel buckles at the reported 5.6 K temperature rise.Source: §5.1 · pp. 18–20
  • The reference model places damping and aerodynamic pressure among generalized forces, with positive-semidefinite damping able to remove energy but not add it.The aerodynamic force can supply or remove energy, while the records omit turbulent boundary-layer pressure forcing.Source: §5.1 · pp. 20–21
  • The learned model uses W = q/ws and τ = ω1t, with ws chosen so a unit first-mode peak deflection equals one panel thickness.The scale makes each component of the post-flutter state order one.Source: §5.1 · p. 21

5.2. The generalized aerodynamic force

The RC-19 aerodynamic force separates a flow-energy input from dissipative aerodynamic damping. The HyMAX reference model uses a clamped-free beam with time-varying flow conditions and shock-dependent pressure loading.

5.2. The generalized aerodynamic force

  • For RC-19, piston theory yields f = −(Aq + Bp), with skew-symmetric A supplying the only term that can feed energy from the flow into the panel.B is symmetric positive definite and dissipative; the computed skew-symmetry residual for A is about 10^-15.Source: §5.2 · p. 21

5.3. The HyMAX reference model

  • The HyMAX reference model treats the plate as an Euler–Bernoulli clamped-free beam using its first two modes and a purely quadratic potential.The measured linear response motivates omitting membrane and thermal terms; damping is set to ζ = 0.01 because the package supplies no displacement measurement.Source: §5.3 · pp. 22–23
  • HyMAX static pressure is a prescribed drive, while motion-dependent pressure follows piston theory with local flow coefficients that change across the measured shock.The freestream pressure varies over the first 90 ms, and the unmeasured back pressure and shock position are sampled at three and two values, respectively.Source: §5.3 · pp. 23–24

5.4. Time integration of the reference model

The reference model uses an average-vector-field discrete-gradient step that exactly balances the stepwise energy change against midpoint generalized-force power. This identity applies to any non-conservative force, including damping and aerodynamic forcing.

5.4. Time integration of the reference model

  • The average-vector-field step evaluates the potential-gradient integral exactly for the model’s potential, whose degree is at most four.Its gradient is at most cubic along the straight segment between states, so Simpson’s rule is exact for the integral.Source: §5.4 · pp. 24–25
  • For any midpoint force fm, the discrete step satisfies H1 − H0 = ∆t pm · fm exactly.The midpoint force includes damping, aerodynamic force, and any prescribed load.Source: §5.4 · pp. 24–25
  • With damping alone, the energy change is nonpositive, and with no force the step conserves energy.The reported balance is exact up to the implicit solver tolerance and roundoff.Source: §5.4 · p. 25

5.5. The record

The record varies measured inputs rather than experimental outcomes, while the learned model preserves the reference potential and energy balance. Its dissipation is constrained to remove energy, and its fitted force uses the exact reference-model rate.

5.5. The record

  • RC-19 records three pressure variants and HyMAX records six back-pressure/shock-position variants; neither family is an ensemble of experiments.RC-19 members share a buckled initial shape, while HyMAX members share the measured freestream and trailing-edge pressures.Source: §5.5 · p. 26

5.6. The learned model

  • The learned model retains the reference potential and fits remaining generalized forces while enforcing an energy identity for every parameter value.The identity holds with fixed conditioning parameters and follows from the model form, not from the fitted force or dissipation values.Source: §5.6 · pp. 26–27

5.7. The learned dissipation

  • The dissipation is symmetric positive definite for every finite parameter vector, so it cannot supply energy to the panel.Its diagonal term damps each modal velocity, while the low-rank term couples modes.Source: §5.7 · p. 28

5.8. The fitted force term

  • The fitted force is affine in displacement, rate, and conditioning features, and is fit by ridge-regularized least squares before training.It remains fixed during training; the reference model's exact stored rate is used to form the fit target.Source: §5.8 · pp. 29–30
  • At the experiment’s 5 kHz sampling rate, finite-difference rates produced an error twelve times the aerodynamic-force scale, so fitting requires a finer rate record.For HyMAX, constant fitted coefficients represent time-varying operators only on average.Source: §5.8 · p. 30

5.9. Time integration of the learned model

The learned model is integrated with second-order schemes, and the AVF step preserves a discrete energy identity for any fitted and trained parameters. The objective fits rollout data without trading away structure; measurement correction enters as a generalized force whose work remains in the energy budget.

5.9. Time integration of the learned model

  • Every AVF step satisfies the discrete energy identity for every value of the trained and fitted parameters.The trained RC-19 and HyMAX models were trained and evaluated with generalized-α; AVF is used to establish the exact discrete identity.Source: §5.9 · p. 31

5.10. The objective

  • The objective fits sixteen-step rollouts and adds a one-sided bound, but contains no conservation or dissipation penalty.The energy identity and admissible dissipation follow from the parameterization, so the optimizer cannot trade structure for fit.Source: §5.10 · pp. 32–33

5.11. Assimilation as a generalized force

  • The observer feeds back measurement and rate errors as a generalized force, with the correction’s work reported separately in the exact discrete energy balance.Synthetic-observation tests are twin experiments, not comparisons with experiment; the smoother reconstruction is non-causal.Source: §5.11 · pp. 33–35
  • At the chaotic condition, position gain alone did not make the closed-loop exponent negative, while rate feedback contracted the error.The open-loop exponent was 285 s−1.Source: §5.11 · p. 34

6. Experimental setup and verification protocol

Training uses the challenge-problem record sets, while RC-19 temperature reaches the learned model through its potential. The RC-19 protocol preregisters acceptance conditions and uses repeat sets and ensembles; the HyMAX protocol measures free-rollout displacement error over specified windows.

6.1. Training

  • RC-19 training uses three record members at 12.76 K, and HyMAX training uses the six-member record.Because RC-19 training contains only one temperature, the fitted temperature feature is zero and temperature enters through the potential alone.Source: §6.1 · p. 35

6.2. RC-19 protocol

  • RC-19 predictions and acceptance conditions were fixed before runs, and a condition with tolerance below statistic noise was reported as unusable.Response statistics are read at three-quarter chord; a repeat set comprises eight kicked releases per well.Source: §6.2 · pp. 35–36

6.3. HyMAX protocol

  • The HyMAX comparison is a free rollout from recorded sample 100, assessed by trailing-edge displacement over an early 1.9 ms window and the full 64.9 ms window.No recorded state is fed back after rollout begins.Source: §6.3 · pp. 37–38

7. RC-19: the clamped panel at Mach 1.92

This section examines structure-preserving modeling of RC-19, including energy constraints, experimental response statistics, published response regimes, aerodynamic-model sensitivity, and displacement-based assimilation. It also addresses the limits of synthetic assimilation and of testing a structural explanation for measured fluctuations without time-resolved data.

7.1. Exact structure, and self-consistency

  • 2.2 × 10^-15 was the worst relative residual of the trained model’s energy identity, while the fitted force reproduced the reference model’s non-conservative operators to 3.6 × 10^-7 relative error.The AVF energy-balance residual was 3.8×10^-11 per step relative to energy, and learned dissipation remained positive definite; on this record, the fitted term also carried structural damping, leaving little for dissipation to carry.Source: §7.1 · p. 38, Figure 4 · p. 39, Figure 5 · p. 40

7.2. Agreement with the experimental measurements

  • 1.521 was both the shock-relation prediction and the measured pressure ratio; the two pressure instruments also agreed on wall temperature to 0.1 K, and model mean frequencies were within 1.1% of the measured 236 Hz.The chaotic-case wall temperatures were 369.4 K from Kulite pressures and 369.5 K from scanner pressures; both differed from the 377 K simulation by about 7.5 K. The frequency statistics depend on the kick interval sampled by the ensemble.Source: §7.2 · pp. 39, 41, Table 6 · p. 41, Figure 6 · p. 42

7.3. The regime map at the simulation operating conditions of Brouwer et al.

  • Across sixteen repeats, the trained model reproduced the periodic and chaotic regimes of Brouwer et al.’s model at its simulation inputs and showed no 113 Hz peak.At the periodic condition, every trained-model repeat returned the same 253.9 Hz bin; its amplitude was 0.7164 panel thicknesses versus 0.7161 for the reference model. The trained-model predictions held at conditions outside its training temperature.Source: §7.3 · pp. 43–44, Table 7 · p. 43, Figure 8 · p. 44

7.4. What the aerodynamic model contributes

  • 2.4% amplitude change accompanied a 5.4% aerodynamic-coefficient change, while the mean frequency was equivalent within 2% in both wells.The spectral change was 3.09 and 3.74 times the within-configuration sampling difference, so the registered joint criterion for frequency equivalence and spectral stability was not met; the study characterizes the enrichment as a percent-level refinement, not a determinant of regime.Source: Figure 9 · p. 46, §7.4 · pp. 47–48, Figure 10 · p. 47

7.5. Chaos and the tracking horizon

  • 104 ms was the estimated tracking ceiling at double-precision roundoff accuracy for the chaotic case, and each decade of improved starting accuracy added only 6.6 ms.The trained and reference models agreed in Lyapunov-exponent sign at all six swept temperatures, but the 8.7 K single-trajectory comparison differed by 17.6%; eight-trajectory ranges overlapped and were not statistically distinguished (p = 0.19).Source: §7.5 · pp. 48–50

7.6. Tracking by assimilation

  • The assimilation test used nine measured displacement points to correct a reference-model copy initialized 10^-3 panel thicknesses from the truth.The correction acted like springs and dampers on displacement and rate mismatch, with kp = 105 N m−1 and kd = 5 × 10^3 N s m−1; this passage describes the test setup, not its outcome.Source: §7.6 · p. 50
  • The nine published observation points observe all seven modes, with an observability-Gramian condition number of 45 and the weakest mode combination about 6.7 times less visible in amplitude.The open-loop copy leaves the truth after about 22 ms, whereas the smoother’s late error is measured over 60–120 ms.Source: §7.6 · pp. 50–51
  • With position gain kp = 102, rate feedback yields closed-loop error exponents of −10.6 s−1 at kd = 2 × 103 and −2.6 s−1 at kd = 2 × 104.Position feedback alone did not contract error, even up to kp = 1014; with feedback, the copy stays on the truth while observations arrive.Source: §7.6 · p. 50
  • At the experiment’s 5 kHz rate, the offline smoother’s state error is 0.0125 panel thicknesses without noise and 0.0352 with 10 micrometers of noise.The 10-micrometer noise is Gaussian and independent across points and samples; the causal hold gives 0.325 thicknesses at 5 kHz.Source: §7.6 · pp. 51–52, Figure 12 · p. 51

7.7. The open comparison, and the measurement that closes it

  • In the attached-shock case, the reference model’s fluctuation standard deviation is 0.057–0.062 thicknesses against the measured 0.943, a deficit of about sixteenfold.Its mean chordwise shape matches to an RMS deviation of 0.22 thicknesses, while the digitized profile’s reading error is about 0.11; published statistics do not determine the panel’s motion regime.Source: §7.7 · pp. 52–53, Figure 13 · p. 54
  • All tested parametric candidate causes were eliminated, and closing the fluctuation deficit through piston-theory enrichment would require scaling its coefficients by about 102.7.A learned conservative correction to the potential was registered as a prediction and falsified; the remaining issue lies upstream of learning.Source: §7.7 · p. 53, Table 9 · p. 53
  • The remaining structural hypothesis is that holding the rigid-panel pressure field fixed misses its dependence on deformed shape, which changes effective aeroelastic stiffness.Testing it requires time-resolved displacement histories at the nine points or a computation on the deformed geometry; neither is in the public package, and the assimilation test assumes an exact model and model-generated observations.Source: §7.6 · p. 52, §7.7 · pp. 53–54

8. HyMAX: the cantilever plate at Mach 5.85

The HyMAX model applies the RC-19 structure-preserving construction to a cantilever plate at Mach 5.85. In free rollout it follows the reference model’s trailing-edge displacement for 64.9 ms, while the retained-mode results expose what the record can support.

8. HyMAX: the cantilever plate at Mach 5.85

  • The HyMAX construction changes the reference model and generalized forces for a cantilever plate at Mach 5.85, while retaining the RC-19 energy identity, learned dissipation, objective, and integrator.No term was added to the objective.Source: §8 · p. 54
  • 3.67% relative root-mean-square error over the 64.9 ms rollout; the model follows the reference model’s trailing edge for about 5.4 first-mode cycles without feedback.The error grows linearly rather than exponentially because the HyMAX response is not chaotic.Source: §8 · pp. 54–55, Table 10 · p. 55
  • The fitted and exact generalized forces correlate at 0.998 across both modes, but time-varying aerodynamic operators are represented only on average.The residual acts at the first-mode frequency and builds linearly over times short compared with about 190 ms.Source: §8 · p. 55
  • Two and three modes yield 3.67% full-rollout error, while adding a third mode does not reduce it; a six-mode basis is rejected because its sixth direction carries no representable energy under the measured drive.The record’s eleven independent whitened coordinates cannot represent the twelve-component six-mode state.Source: §8 · p. 55, Figure 15 · p. 57
  • The HyMAX results are graded against the reference model, not experiment; its 83 Hz first frequency is measured by construction, while ζ = 0.01 is a setting.The reference model’s 520 Hz second frequency is within 2% of Xu and Dowell’s 530.4 Hz computation.Source: §8 · p. 55

9. Preregistered predictions and their outcomes

The registered RC-19 studies report four broken predictions, retain them without changing thresholds, and identify lessons about calibration, noise-aware acceptance criteria, and single-trajectory Lyapunov estimates. Four trained-model predictions held, supporting specific claims about regimes, statistics, and fitted forces rather than learned dynamics.

9. Preregistered predictions and their outcomes

  • Four of nine registered predictions broke—A1, R2, R3, and T5—and all outcomes were retained without moving thresholds.Twenty-one earlier predictions for superseded configurations are not listed.Source: §9 · p. 58
  • R2’s 2% frequency tolerance was smaller than the estimator’s 4.88 Hz grid interval and the 16.3 Hz pooled standard deviation, so the criterion was below its own noise.R3 changed the statistic and sized its sample from the measured spread, but broke its spectral clause even though its mean-frequency clause held.Source: §9 · p. 58
  • T5 broke at 8.7 K, where eight trajectories per model found the 17.6% difference inside the scatter of single-trajectory exponents; that diagnosis did not re-grade the prediction.The follow-up was registered before it ran.Source: §9 · p. 58
  • Four trained-model predictions held: outside its training temperature, it returned the reference model’s regimes, ensemble frequency and amplitude, and neither model showed the experiment-absent 113 Hz peak.The fitted force matched the reference model’s non-conservative forces to relative root-mean-square error 3.6 × 10−7; these results do not show that training learned the dynamics.Source: §9 · p. 58

10. Discussion

The discussion argues that structure lets the models report both supported behavior and unsupported directions, despite records that lack measured trajectories. The reported fits and comparisons establish self-consistency and bounded experimental agreement, not that training learned the dynamics.

10. Discussion

  • The challenge packages provide neither an ensemble nor time-resolved responses, so the models rely on physics encoded in their form rather than data-learned physics.On RC-19, 35 trained and 119 fitted numbers sit alongside 2464 fixed potential numbers; the energy identity holds at initialization and every epoch.Source: §10 · pp. 58, 60
  • The RC-19 ridge-regularized least-squares fit reaches relative error 3.6 × 10−7 before training, while HyMAX fitted and exact forces correlate at 0.998 across both modes.The HyMAX fit is limited by time-varying aerodynamic operators.Source: §10 · p. 60
  • On RC-19, fitted forces absorb the reference model’s linear damping, leaving trained-dissipation eigenvalues of 2 × 10−8 to 10−7 versus damping of 0.010 to 0.020; its added value needs a record with nonlinear dissipation.The authors report this as the expected outcome for linear non-conservative forces, not as optimizer success.Source: §10 · p. 60
  • Rejecting HyMAX’s unsupported sixth mode and closing RC-19’s work budget to roundoff illustrate structure as a diagnostic: failures appear in the budget term where they occur, rather than later as drift.The RC-19 budget includes work by the flow and by the assimilation correction.Source: §10 · p. 60
  • Chaotic trajectories cannot be forecast beyond tens of milliseconds at practical starting accuracy, so reproducible comparisons are statistics and longer tracking needs measurement.The twin experiment shows that offline reconstruction could use nine published points sampled at 5 kHz, but requires time series absent from the package.Source: §10 · p. 60
  • The four broken preregistered predictions show that corrections can be calibration, acceptance criteria must account for statistic noise, and a single-trajectory Lyapunov exponent is a range rather than a value.A preregistration with every prediction holding would have been less informative and credible.Source: §10 · pp. 60–61

11. Limitations

The limitations stem mainly from what the public packages measure and omit: they support statistical grading but generally not training or trajectory tracking. Consequently, several results are self-consistency tests, and HyMAX has no comparison with experiment in this report.

11. Limitations

  • The public packages were assembled to grade computations against statistics, not to train or track models; their gaps constrain the reported results.RC-19 has no response time history or repeat runs, and HyMAX has no displacement or damping estimate.Source: §11 · p. 61
  • RC-19 inputs include database material constants, an inferred temperature-field shape based on two thermocouples, and pressures measured on a rigid rather than deflecting panel.The chaotic response is reported from single runs without repeats, so one statistic’s sampling spread cannot be estimated from the package.Source: §11 · p. 61
  • Because records are generated by reference models, learned-model comparisons are self-consistency tests; the linear non-conservative forces are fitted to relative error 3.6 × 10−7 before training.A measured record with non-linear dissipation would be needed to test whether learned dissipation adds accuracy.Source: §11 · p. 61
  • The assimilation result is a twin experiment, not agreement with experiment, because no measured time series is available to assimilate.It assesses what the published sampling rate and layout can support.Source: §11 · p. 61
  • HyMAX back pressure and shock position were unmeasured, its damping ratio ζ = 0.01 is a setting, and results are graded against its own reference model rather than experiment.A measured trailing-edge displacement is stated as forthcoming but is not used here.Source: §11 · p. 61
  • RC-19 comparisons with experiment fall short at some conditions: the periodic simulation frequency is 7.6% above measurement, and the attached-shock fluctuation level is about sixteen times below it.The measured pressure field is from a rigid panel, leaving dependence on deflected shape untested; ensemble means and single-trajectory Lyapunov exponents also have sampling limits.Source: §11 · p. 62
  • No computational cost comparison was made, so the paper claims no speedup.Source: §11 · p. 62

12. Conclusions and outlook

The paper reports structure-preserving learned models for both challenge problems, with energy balance maintained by construction. Results include RC-19 agreement with measured shock pressure and response frequency, synthetic-data assimilation, and a HyMAX rollout; the outlook identifies data needs and possible extension to finite-element models.

12. Conclusions and outlook

  • The construction covers both challenge problems, and its energy identity holds to roundoff while learned dissipation cannot add energy and the discrete-gradient step balances energy at each step, including under assimilation.Source: §12 · p. 62
  • Assimilation at the published 5 kHz rate keeps the twin-experiment state error near 1% of a thickness, while the HyMAX model follows its reference trailing edge for 65 ms with 3.67% relative error.Source: §12 · p. 63
  • A live digital twin remains untested for model-error correction, and causal tracking at 5 kHz holds state to 0.325 thicknesses versus 0.0125 for the smoother.The correction has so far been run only on the reference model.Source: §12 · p. 63
  • Further data are needed to test deformed-geometry pressure effects, compare HyMAX displacement with experiment, and assess learned dissipation and fitted-term temperature dependence.The proposed tests require records with nonlinear non-conservative forces and records spanning temperature.Source: §12 · p. 63
  • The construction can extend beyond these panels: finite-element models can supply the operators, but parameter-range tests require a valid modal basis and data exciting every retained mode.Source: §12 · p. 63

Reproducibility

The challenge-problem packages are distributed by the AePW-4 High-Speed Working Group. This report's records are generated from those packages using the reference models.

Reproducibility

  • The AePW-4 High-Speed Working Group distributes the challenge-problem packages.Source: Reproducibility · p. 63
  • The report's records are generated from the distributed packages using the reference models described in Section 5.Source: Reproducibility · p. 63
  • Inquiries about the records may be directed to the corresponding author.Source: Reproducibility · p. 63

Related papers

Loading 2610.31504v1…