acceptodds
Under review as a conference paper at ICLR 2027

Stability-Aware Parametric Neural ODEs for Bifurcating Dynamical Systems

Abstract

Parametric dynamical surrogates must recover both the stability of an equilibrium and the nonlinear oscillations that emerge when it loses stability. Near a Hopf bifurcation, weak growth or decay can be difficult to resolve alongside the dominant oscillation, while joint fitting allows nonlinear training to alter the equilibrium spectrum. We introduce a two-stage parametric neural ODE that couples shared operator identification with exact Jacobian preservation. A shared matrix polynomial in the operating parameter is identified jointly across training conditions from normalised early-window trajectory error, with stability penalties, and then frozen. A neural coefficient matrix supplies the nonlinear correction through an amplitude gate. The resulting residual is third order at the equilibrium, so the deployed model’s equilibrium Jacobian is exactly the identified operator at every parameter value, independently of residual training. We evaluate the construction on a thermoacoustic Rijke tube, a cylinder wake and a flexible aircraft wing. The method recovers the correct number of stability crossings on all three systems and accurately locates their boundaries, with errors within the reference resolution on the wake and wing and within a tenth of the grid spacing on the tube. It combines this stability recovery with accurate prediction of nonlinear trajectories and limit-cycle geometry at unseen parameter conditions. Mean per-seed median test trajectory nRMSE is 0.108, 0.169 and 0.147, respectively. Objective comparisons and architectural ablations justify the choices of trajectory fitting, polynomial operator parameterisation, freezing and gating.

Primary area:
learning on time series and dynamical systems
Keywords:
neural ODEs, parametric dynamical systems, bifurcation analysis, scientific machine learning, interpretability

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.67 over 3 reviews, above 99% of the others.
Reviewer_C3tEOpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies parametric surrogates for systems with Hopf bifurcations. The authors argue that weak growth or decay near the boundary is poorly resolved by derivative-based identification, and that jointly training a neural vector field lets nonlinear learning move the equilibrium Jacobian and hence the stability boundaries. The authors propose two-stage model. In Stage 1, a matrix polynomial is fitted jointly across training conditions to early pre-saturation trajectories by a normalised multi-step trajectory loss, with a stability hinge on stable records, and is then frozen. In Stage 2, a neural coefficient matrix is trained on free rollouts. The method is evaluated on a Rijke tube, a cylinder wake and a flexible wing. The reported results show that it is the only model recovering the correct crossing count on all three systems, with crossing errors of at most 1.26% and the lowest neural test nRMSE.

Critical strengths
  1. Well-motivated two-stage design. The key property holds by construction and is isolated by targeted ablations.

    • The motivation is stated quantitatively: the sensitivity ratio scales as ∣ω∣⊤|\omega|^\top for trajectory versus derivative fitting.
    • Gating, freezing and staged training are each shown to matter. Removing or releasing them shifts the crossings (Fig. 24, Table 17), and the objective and window comparisons support trajectory fitting (Table 5, Fig. 21, Table 13).
    • This identifies a real failure mode, nonlinear training silently moving stability boundaries, and offers a simple remedy with an explicit, inspectable spectrum at unseen parameters (Fig. 22).
  2. Thorough evaluation.

    • Crossing counts are reported next to locations, so that spurious crossings are not hidden behind small location errors.
    • Constant-predictor reference levels (App. F.2), regime-split errors (Table 9), boundedness, first-stage robustness (App. G.1), stable-only extrapolation (F.8) and a cropped-record comparison (F.7) are all included.
    • The authors also qualify their own reference quantities (App. B.5). This breadth makes the paper informative well beyond the headline numbers.
Critical weaknesses
  1. Spectrally narrow test bed, and missing examinations on numerical effects.
    • All three systems are single-pair Hopf crossings with clean oscillations. Tests on real-eigenvalue crossing, mode coalescence or multi-frequency interaction would make the manuscript even better.
    • Stage 1 uses the exact matrix exponential, while the deployed model is integrated with RK4 at the record step. On the wake this is only about 14--19 steps per period. Expanding on numerical artifacts and/or further clarifying the choice on the wake reference (which seems discretization-specific; Rec=44.11\text{Re}_c = 44.11 vs 46--47, following line 941) would be nicer.
Questions for authors
  1. Stage 1 seems noise-sensitive: z0z_0 enters exp⁡(At)z0\exp(At)z_0 when given per-record normalisation up-weights low-SNR records, and the equilibrium is assumed exact. Perhaps, would fitting z0z_0 jointly be helpful?
  2. How sensitive are crossings and nRMSE to the RK4 step per period, especially on the wake? Would an exponential integrator that treats the frozen A(p)A(p) exactly change the results? Are the reported crossings roots of the explicit A(p)A(p) or linear interpolation on the grid?
  3. Why was degree 1 not selected on the wake, given that App. F.5 (line 2425) says it's competitive (lower error on 4/5 seeds)?
  4. Section 4.1 mentions “two cells slightly favouring ablations” in Table 9, but I cannot find which cells it means. I find only the ungated model on the wing (inside the unstable interval) as an ablation favoured. In addition, the plain and jointly trained free-operator rows are slightly lower on the tube (inside, Table 10), but those are not ablations.
  5. In App. A.3: "Records are stored every 0.4 convective times, while the model step is Δt=0.399\Delta t = 0.399." why does it have to be like this?

Minor comments.

  • Appendix F. 8 says the error “grows with the distance”, but the wing rows decrease (2.72→0.612.72 \to 0.61).
  • Typos: “52.21and82.55” (line 962), Table 8 caption “held out the same statistic...” is ungrammatical (maybe missed "is" after "held out".)
  • Missing explanation and citation for “SINDy-CP” (App. B.4 line 1079).
2 comments
AuthorsOpenReview ↗

Response to Reviewer C3tE - Part 1/2

Thank you for the careful reading and constructive questions. The submission is scoped to Hopf bifurcations, as stated in Section 5. In all three applications, each stability transition involves a single complex-conjugate eigenvalue pair. We will clarify in that section that real-eigenvalue crossings, a second unstable pair and quasi-periodic motion were outside the experimental scope and were not tested. Your questions required new fits, training runs and rollouts, which we conducted and report below.

Q1. Noise sensitivity and fitting the initial state

Identification from noisy measurements is outside the stated scope of the submitted paper. Section 5 explicitly restricts the experiments to noise-free simulations with a supplied equilibrium and states that noise robustness remains untested. We added this study in response to your question to examine this additional setting.

We added Gaussian noise with standard deviation 1% of each coordinate’s standard deviation to the wake records and refitted stage one on five independent draws. We score against the clean reference. Crossing error is the distance from the reference to the nearest predicted crossing, in percent of the critical Reynolds number. Every draw has at least one crossing, so all five enter each statistic.

The last two columns count draws with exactly one crossing and with correct signs at every condition, respectively.

DegreeInitial statesMedian error (%)Worst error (%)One crossingAll signs correct
4 (reported recipe)Recorded0.1628.53/53/5
2Recorded0.090.213/53/5
4Fitted0.059.15/54/5

On clean records, the first two configurations give 0.15 percent (the Table 5 control fit) and 0.09 percent, each with one crossing.

Under this additional noisy-data test, the reported recipe produces incorrect crossing counts in two of the five draws. One of these has no crossing near the reference, with the nearest at Re=56.7\mathrm{Re}=56.7. Jointly fitting initial states and the operator, as you suggested, returns a single crossing in all five draws and lowers the median error. One draw places the crossing 9 percent above the reference. On clean records, fitting the initial states improves the crossing estimate. Subcritical rollout error increases from 0.12 to 0.26 when rollouts are initialised at the measured initial states. Degree 2 has the most consistent locations, all within 0.21 percent, although two draws have spurious crossings.

An exploratory single-draw test with equal absolute noise across records, giving small records low signal-to-noise ratios, yields one crossing and errors at most 0.14 percent for all three configurations. Stability labels and the supplied equilibrium were fixed, isolating noise in the trajectory term. Equilibrium uncertainty remains untested. We will report these results, include fitted initial states as an option in stage one, and document the observed sensitivity to noise in Section 5.

AuthorsOpenReview ↗

Response to Reviewer C3tE - Part 2/2

Q2. Integration and crossing computation

Crossings are independent of the rollout integrator. They are sign changes of the largest real part of the eigenvalues of A(p)A(p), located by linear interpolation on a dense parameter scan separate from the record grid (Appendix C). Refining the roots by bisection shifts each by about 0.001 percent of its normaliser or less, changing no reported digit.

Without retraining, we re-rolled the reported models with finer RK4 steps and an exponential scheme treating A(p)A(p) exactly. Values are means over five seeds. Wing values cover the 14 angles inside the unstable interval in Appendix C, pooling training, development and test conditions as in Table 9.

SchemeTube test nRMSEWake test nRMSEWing nRMSE
RK4, record step (reported)0.1080.1690.524
RK4, 2 sub-steps0.1080.1750.522
RK4, 4 sub-steps0.1080.1760.522
Exponential, A(p)A(p) exact0.1080.1720.522

The wake has 14–19 steps per limit-cycle period, as you noted, versus 71–90 on the tube and about 34 on the wing. The largest aggregate error change is 0.007 on the wake, about a tenth of the gap to the next architecture in Table 3. The exponential scheme changes aggregate error by approximately 0.003 or less. Finer RK4 steps do not improve the wake’s aggregate error for these fixed models, trained through the original RK4 map (Appendix B.3). The largest condition-level differences occur just above the wake crossing, where slow growth makes saturation time sensitive to small growth-rate changes (Appendix E.2). At individual test conditions, nRMSE against the reference changes by up to 0.25 in either direction. The wing test median in Table 3 is unchanged under every scheme, and all rollouts remain bounded.

As Appendix A.3 states, Rec=44.11\mathrm{Re}_{c}=44.11 is the critical value of this discretisation, not the continuum. It is the appropriate reference because the surrogate learns from that simulator’s records. We did not perform a resolution study and will state this in the revision.

Q3. Degree 1 on the wake

We selected the degree using the operator’s development trajectory error. The continuation started at degree 2 and proceeded to higher degrees. Degree 4 gave the lowest error (Appendix B.1). Degree 1 was below that starting point and was fitted afterwards as a check (Appendix F.5). Including it would not change the choice, as its first-stage development error is 0.145 versus 0.122 for degree 4.

After stage two, degree 1 offers similar accuracy with 128 operator coefficients versus 320. We retain the reported model to avoid selection after seeing test results and will clarify these points in the revision.

Q4. Section 4.1 and Table 9

The sentence in Section 4.1 will be corrected to match the tables. The only Table 9 cell favouring an ablation is the ungated wing model inside the unstable interval, at 0.52 versus 0.54. In Table 10, the plain and jointly trained free-operator models are marginally better on tube test conditions inside the interval, at 0.12 versus 0.13, but these are architectures, not ablations.

Our claim of lowest test error among neural architectures refers to the median over all test conditions (Table 3). On the tube, the advantage comes from the stable regime, where those two models score 0.69 and 0.41 versus 0.10. We will correct Section 4.1.

Q5. Wake sampling interval and model step

There is no mismatch. The model step equals the stored interval, while 0.4 is nominal. Snapshots are stored on whole lattice steps. With 20 nodes per diameter and inlet speed 0.07 in lattice units, one convective time is 285.7 lattice steps. Thus 0.4 convective times would require 114.29 steps. Storing every 114 gives 0.399. Appendix A.3 will state the actual interval, 0.399 convective times.

Minor comments

  • Appendix F.8. The wing rows show error decreasing from 2.72 to 0.61 percent as distance increases from 2.78 to 8.33 percent (Table 16). Both errors lie within the 2.78% reference resolution, which limits conclusions about accuracy but does not change their order. The statement that error grows with distance holds only for the tube and wake. We will correct it.
  • Both typos will be corrected.
  • Appendix B.4 will define SINDy-CP as Nicolaou et al.’s (2023) sparse-regression formulation with control parameters, using products of state and parameter monomials, and add the citation.
Reviewer_LFCJOpenReview ↗
Rating
3(+2)
Confidence
3
Informativeness
2(+1)
Presentation
2
Soundness
3(+2)
Read review
Paper summary

The paper learns and freezes a parameter-dependent linear model, then trains a gated neural correction to capture nonlinear dynamics. The gate preserves the identified equilibrium Jacobian exactly, preventing nonlinear training from shifting its stability boundaries.

Critical strengths

Extension of neural ODEs to systems with Hopf bifurcations.

Critical weaknesses

I understand, and I accept the arguments. I have raised the score to reflect the soundness of the argument.

Questions for authors

None

2 comments
AuthorsOpenReview ↗

Response to Reviewer LFCJ

Thank you for the review. We agree that the equilibrium Jacobian alone does not determine Hopf criticality and that local stability does not imply global attraction. However, Proposition 1 claims neither determination of Hopf criticality from the Jacobian nor global attraction. We clarify below why these observations do not invalidate the proposed gate.

1. Supercritical versus subcritical Hopf bifurcations

Proposition 1 guarantees preservation of the identified equilibrium Jacobian during nonlinear training. It does not determine the bifurcation type from that Jacobian. The submitted main text states this distinction on page 4, lines 184–185.

"It does not determine whether a Hopf bifurcation is supercritical or subcritical, which depends on nonlinear terms."

Appendix B.7 further explains that these nonlinear terms are supplied by the trainable residual.

The architecture permits either Hopf criticality. To demonstrate this directly, consider a two-dimensional instance with state scales chosen so that r=∥z∥2r=\lVert z\rVert_2, and take

A(μ)=(μ−ωωμ),B=bI,ω>0,A(\mu)=\begin{pmatrix}\mu & -\omega \cr \omega & \mu\end{pmatrix},\qquad B=bI,\qquad \omega > 0,

where BB is the coefficient matrix of equation (1). Equations (1)–(2) then give

r˙=μr+br3r2+r02,φ˙=ω.\dot r=\mu r+\frac{b r^3}{r^2+r_0^2},\qquad \dot\varphi=\omega.

Near the equilibrium,

r˙=μr+br02r3+O(r5).\dot r=\mu r+\frac{b}{r_0^2}r^3+O(r^5).

Choosing b<0b < 0 or b>0b > 0 produces supercritical or subcritical Hopf behavior, respectively, while leaving the equilibrium Jacobian unchanged. Thus, freezing the linear operator does not fix the sign of the nonlinear coefficient that distinguishes the two cases.

This establishes that the architecture can represent both types. As with the nonlinear learning baselines, whether the fitted model recovers the physical system's criticality depends on identifying the relevant nonlinear terms accurately. Preserving the equilibrium Jacobian does not remove that representational freedom. We do not claim guaranteed identification of criticality. The paper reports the accuracy of the learned limit cycles separately, with amplitude errors of 1.5, 3.4 and 17.6 percent on the three systems (Table 11).

2. Local stability versus stability under large disturbances

We agree that preserving a locally stable Jacobian does not guarantee global attraction. The purpose of the gate is to prevent residual training from changing the equilibrium spectrum. It imposes no global-attraction constraint.

Indeed, the example in the review,

x˙=−x+2x31+x2,\dot x=-x+\frac{2x^3}{1+x^2},

is exactly an instance of our architecture with

A=−1,B=2,r=∣x∣,r0=1.A=-1,\qquad B=2,\qquad r=\lvert x\rvert,\qquad r_0=1.

Its equilibrium Jacobian is −1-1, exactly as Proposition 1 requires, while the origin has the finite basin of attraction (−1,1)(-1,1). The example therefore satisfies the claimed preservation property. Its finite basin does not demonstrate a failure of the gate.

For r≥0r\geq 0, the same expression is the radial equation of the planar family in point 1 with b=2b=2 and r0=1r_0=1, evaluated at μ=−1\mu=-1. This family undergoes a subcritical Hopf bifurcation at μ=0\mu=0. At μ=−1\mu=-1, the stable origin has a basin bounded by an unstable periodic orbit of radius one. The example therefore shows that the architecture represents the subcritical behavior raised in the first point.

With the first-stage operator fixed, the gate preserves its crossings throughout residual training. Consistent with this guarantee, Table 17 reports zero variation across second-stage training seeds. Removing the gate gives first-crossing standard deviations of 0.102 percent of the unstable-interval width on the tube and 0.212 in Reynolds number on the wake.

The paper evaluates nonlinear trajectory accuracy, limit-cycle geometry, and finite-horizon rollout boundedness separately from equilibrium spectral preservation. These empirical results are not presented as a global stability guarantee.

In the revision, we will make explicit that the stability boundaries considered here concern linear stability of the supplied equilibrium. Hopf criticality and basins of attraction depend on the full nonlinear vector field, including the learned residual, and are not guaranteed by equilibrium-Jacobian preservation.

Neither observation contradicts the guarantee claimed in Proposition 1. The architecture permits both Hopf criticalities, and the reviewer's scalar example satisfies the Jacobian-preservation property exactly. We respectfully ask the reviewer to reconsider the soundness assessment and the overall recommendation.

AuthorsOpenReview ↗

Response to Updated Review by Reviewer LFCJ

Thank you for considering our feedback and for updating the review.

The scope of Proposition 1 is stated in the submitted paper. On page 4 it says that the proposition fixes the crossing locations and count but does not determine whether a Hopf bifurcation is supercritical or subcritical (lines 184 and 185, and Appendix B.7), and the paper makes no claim about basins of attraction. The revision will add that the stability boundaries concern linear stability of the supplied equilibrium, and will repeat this scope in Section 5, where the paper lists its limitations. It will also define each auxiliary loss term in Appendix B.2 and state in each table caption which conditions the reported values cover.

To further improve the presentation, we would be glad to clarify any sections or passages that you felt required the most effort. We will revise them in the updated version.

Reviewer_9v8mOpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes a two-stage parametric neural ODE for systems undergoing Hopf bifurcations. It first identifies and freezes a shared polynomial linear operator from early-time trajectories, with stability-label regularization. An amplitude-gated nonlinear residual then learns nonlinear dynamics while preserving the identified equilibrium Jacobian exactly. Experiments on a Rijke tube, cylinder wake, and flexible wing demonstrate recovery of stability crossings and prediction of transient trajectories and limit-cycle geometry at unseen parameter values. Ablations examine the roles of trajectory fitting, operator freezing, and residual gating.

Critical strengths

I find the construction simple and well matched to the problem. Identifying weak growth rates separately, then preventing nonlinear training from modifying the equilibrium Jacobian, directly addresses a meaningful failure mode of parametric surrogates. The cubic onset of the residual is also a natural inductive bias near a Hopf bifurcation. The physical demonstrations are substantial, particularly the coupled flexible-wing simulation. The stable-record-only boundary prediction in Appendix F.8 is especially compelling, as it tests recovery beyond the observed stability regimes rather than merely reproducing supplied labels.

Critical weaknesses

The nonlinear prediction results rely on substantial system-specific loss design. Appendix B.2 combines envelope, phase, frequency, amplitude and derivative terms differently across systems. The full-recipe neural ODE control usefully shows that these losses alone do not explain the improvement. However, it does not establish how well the proposed architecture works with a simple trajectory objective. It remains unclear how much manual loss design and calibration would be required for a new physical system, and whether the reported gains are robust to the relative weighting of the auxiliary terms.

Questions for authors
  1. With the identified operator fixed, how does trajectory-only training compare with the full auxiliary objective in trajectory error and limit-cycle amplitude/frequency accuracy?
  2. How were the relative auxiliary-loss weights chosen, and how sensitive are the results to their relative balance? Tuning a shared multiplier does not resolve this question.
2 comments
AuthorsOpenReview ↗

Response to Reviewer 9v8m - Part 1/2

Thank you for the review. The second-stage objective is system specific, as described in Appendix B.2. Both questions required new training runs on all three systems, which we performed using the reported operator, gate, curriculum and seeds. The frozen operator is unchanged.

Q1. Trajectory-only training with the operator fixed

All auxiliary weights are set to zero. Values are means over three seeds.

SystemObjectivenRMSEAmp. error (%)Freq. error (%)
WakeReported0.1684.00.19
Trajectory only0.1142.60.15
TubeReported0.1021.70.055
Trajectory only0.11312.00.047
WingReported0.55115.40.09
Trajectory only0.87763.90.57

Tube and wake scores use the test conditions of Table 3. Both tube rows use the final epoch, which matches the reported model's checkpoint-selected trajectory and amplitude errors. Under the submitted checkpoint rule, the tube trajectory-only errors are 0.150 nRMSE, 10.4% amplitude and 0.084% frequency. The wing's Table 3 test median falls at a stable angle and is 0.147 in every run. We therefore report medians over the 14 angles inside the unstable interval (Appendix C), pooling training, development and test angles as in Table 9, which gives 0.54 versus 0.55 on test angles alone in Table 10.

The different effects may reflect each system's nonlinear response, although these experiments do not isolate the cause. The wake undergoes a supercritical Hopf bifurcation and follows the Stuart-Landau law (Appendix A.3), and trajectory error alone suffices there. On the tube, weak growth throughout the unstable interval (Appendix A.2) means much of each record is slow growth already captured by the frozen operator. This may explain why trajectory error changes less than amplitude error, which rises from 1.7 to 12.0%. The wing has a steep onset and strongly asymmetric branch (Appendix A.1), and the auxiliary terms improve its nonlinear predictions in these experiments.

With trajectory-only training, the proposed model substantially outperforms the corresponding plain neural ODE, whose test errors are 0.441 on the tube and 1.001 on the wake (Table 3, the variant labelled trajectory loss and curriculum in Figure 20). This supports an architectural contribution beyond the auxiliary objective (Appendix E.4).

AuthorsOpenReview ↗

Response to Reviewer 9v8m - Part 2/2

Q2. Choice of weights and sensitivity to their balance

We chose the relative auxiliary-loss weights manually for each system, using the training and development conditions. For the tube and wake, we then fixed those relative weights and used a grid search on the development conditions to select a shared multiplier. The shared multiplier changes the overall strength of the terms it scales while preserving their relative weights. We did not perform a systematic joint search over the individual weights. The new experiments assess that balance directly by varying selected weights and removing terms individually and in groups.

These weights are part of the shared comparison recipe, so changing them requires retraining each affected row (Appendix B.3). A second-stage fit takes about 4.3 hours on the tube, 2.2 hours on the wake and 1.2–6.8 hours on the wing (Appendix B.6).

We first varied one auxiliary weight at a time, using one seed per system. Tube and wake trajectory errors use Table 3's test conditions. Wake amplitude and frequency errors use supercritical test conditions, while the tube uses all conditions with a limit cycle. Tube rows use the final epoch. Wing values use the same 14 angles.

SystemTerm variedValuenRMSEAmp. error (%)Freq. error (%)
WakePhase, the only active term00.1051.30.17
Reported0.1420.70.11
3.3 times0.1521.30.17
10 times0.1815.80.23
TubeTransient envelopeHalved0.1001.70.03
Reported0.1041.60.09
Doubled0.0991.50.03
WingSettled amplitudeHalved0.5702.50.17
Reported0.5528.60.13
Doubled0.56920.10.09

Halving or doubling the tested weight changes nRMSE by at most 0.005 on the tube and 0.02 on the wing, where the reported weight gives the lowest nRMSE of the three. Wake error rises by 0.04 over a tenfold weight range and is lowest at zero, as in Q1. All frequency errors remain below 0.25%. Under these weight changes, wing amplitude error ranges from 2.5 to 20.1%, compared with 8.6–27.5% across the reported model's seeds. We will clarify that Appendix B.2 gives the wake phase weight before the 0.3 multiplier, with an effective value of 0.03.

We then removed each term in turn and selected groups of terms, using one seed. These runs set the corresponding weights to zero. Tube values use Table 3's test conditions at the final epoch. Wing values use the same 14 angles.

Tube, terms removednRMSEAmp. error (%)Cycle error (%)
None (reported)0.1042.04.2
Any one of nine terms, range0.095 to 0.1161.1 to 3.23.6 to 4.6
Four envelope terms together0.1091.02.3
Two frequency terms together0.1021.64.3
Two settled terms together0.0983.24.1
Vector-field anchor0.1261.66.4
All ten (trajectory only)0.12111.212.5
Wing, terms removednRMSEAmp. error (%)Freq. error (%)
None (reported)0.5528.60.13
Any one of five terms, range0.461 to 0.5694.1 to 25.80.02 to 0.10
Carrier phase1.09236.30.37
All six (trajectory only)0.84260.80.32

Of the individual tube removals, only the vector-field anchor changes trajectory error beyond the reported seed spread of 0.017. The other nine stay within it. Removing the four envelope or two frequency terms together also makes little difference. Wing carrier phase is the most consequential term. Over three seeds, mean nRMSE is 1.139 without it and 0.716 with it alone, versus 0.551 for the full objective, so the other terms still matter together.

Auxiliary terms are not uniformly needed for accurate trajectories. Removing them improves wake trajectory error and modestly worsens tube trajectories, but substantially degrades tube amplitude accuracy and wing predictions. Most other perturbations have smaller trajectory effects, while the settled-amplitude weight changes wing amplitude accuracy. Most sensitivity tests use one seed. The frozen operator and crossings remain unchanged in every run (Table 4, Proposition 1).

For a new system, the auxiliary objective requires calibration. We will state this requirement explicitly in Section 5. Several external baselines also require system-specific libraries, thresholds, latent dimensions or regularisation weights (Appendix B.4). On the tube, adding the auxiliary terms changes the unstructured model's test error from 0.441 to 1.265 (Table 3), a larger change than 0.113 to 0.102 for the proposed model. We will add these tables to the appendix and define each auxiliary term in Appendix B.2.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.