acceptodds
Under review as a conference paper at ICLR 2027

TrajectoryOp: Explicit-Interface Operator Learning for Moving Shocks

Abstract

Neural operators provide rapid predictions for families of PDE solutions, but moving shocks remain difficult to represent and learn from limited supervised numerical data. We introduce , an explicit-interface operator for conservation laws with a single tracked shock. TrajectoryOp consists of an FNO encoder, two state decoders, and a trajectory network. The FNO encoder extracts features from the discrete initial field, the two state decoders represent the primitive states on the two sides of the shock, and the trajectory network predicts the shock trajectory from the initial shock location. The solution is represented by the two regional states separated by the predicted shock interface. Regional PDE residuals constrain the states inside the two predicted regions, and the Rankine–Hugoniot condition couples these states to the shock motion along the predicted interface over the full time interval, including where numerical state labels are unavailable, thereby reducing the required amount of supervised numerical data. Compared with fully-supervised operator learning methods, proposed achieves almost performance with only about 3% supervised information, demonstrating significant advantages under such sparse supervised case.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Operator learning, Hyperbolic conservation law, physics-informed learning, scientific machine learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_suaGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

TrajectoryOp learns a solution operator with an explicit shock interface. An FNO encodes the initial field, two KAN decoders predict regional primitive states, and an initial-field-conditioned trajectory network locates the interface separating them. Training combines initial and sparse state labels with regional conservation-law residuals and a Rankine-Hugoniot interface loss. The physical constraints cover the entire prescribed time interval, including times without state labels. A graph interface x = xi(y,t) extends the formulation to two dimensions.

The paper tests Burgers, shallow water, one-dimensional Euler rarefaction-contact-shock solutions, and a two-dimensional Euler shock-entropy interaction. With 3.125% of the specified dense nonzero-time state-label tuple budget, TrajectoryOp reduces error against the strongest matched sparse baseline by approximately 77%, 48%, 41%, and 45%. Loss and joint interface/constraint ablations support the design. Dense FNO nevertheless remains more accurate on three of four main benchmarks. The contribution is a useful integration of established interface physics into operator learning, supported primarily by experiments rather than new theoretical guarantees.

Critical strengths
  1. An interpretable mechanism for reducing state supervision. The interface is an explicit output whose motion is coupled to predicted traces through RH, while regional PDE losses constrain unlabeled states. This avoids forcing the tracked jump into a single continuous output representation and amortizes interface learning across initial conditions. It is a consequential, physically motivated design even though its individual ingredients are established.

  2. Consistent sparse-budget improvements with useful controls. All four tests improve over the matched sparse external baselines. The paper includes loss ablations and a clearly labeled joint interface/RH ablation. Its label accounting and validation-time restrictions are transparent, and the FNO comparator is larger than TrajectoryOp in the reported one-dimensional Euler configuration. These features make the reported empirical signal more credible than an architecture comparison alone.

Critical weaknesses
  1. The central smooth-region assumption is not reconciled with the tested wave structures. The Euler RCS family contains an untracked contact discontinuity. More directly, the stated shallow-water ranges admit two shocks: with g = 1, h_L = 1.2, h_R = 0.8, u_L = 0.6, u_R = -0.3, and x_0 = 0, the exact middle state is h_* = 1.47397, u_* = 0.36180. Both waves are compressive shocks, with speeds -0.68150 and 1.14736; both remain inside the domain through T = 0.5. No shallow-water pattern filter is stated. One hard interface leaves another discontinuity inside a branch treated as smooth, undermining the explanation for the regional strong residual. This does not prove that finite-resolution predictions are impossible, especially with diffuse targets, but it requires an explicit treatment and validation. Conservation/RH alone also does not ensure entropy admissibility, and the stated reference boundary conditions are not included in the described network constraints.

  2. The evidence does not yet isolate the physical and label-efficiency advantage over suitable alternatives. The external baselines are data-only; the internal single-decoder control uses a global strong residual across shocks. An appropriate weak/conservative physics-informed learner is missing. References use spatially first-order Rusanov discretization without convergence or exact Riemann checks, and no quantitative shock-location, conservation, or admissibility metrics are reported. All runs use one seed. The 3.125% ratio is correct for nonzero-time state tuples, but not total information, independent simulations, or computational cost. Sparse TrajectoryOp has approximately 1.57, 2.30, and 1.72 times dense FNO's error on shallow water and the two Euler tasks, so broad comparable-performance language needs qualification.

Questions for authors
  1. Does the actual shallow-water dataset exclude two-shock states? Please provide any filter and wave-pattern counts, identify the tracked front, and explain how the other shock is handled when such states are retained. Likewise, how is the Euler contact treated within the regional strong-residual rationale? The explicit within-range example above would be a useful diagnostic.

  2. How are the extrapolation and reflecting boundary conditions imposed on the learned operator, or justified as inactive over the test horizon? What evidence supports admissibility of the predicted fronts on the current benchmarks? I am asking about the existing experiments, not requiring a general entropy theorem or arbitrary new boundary families.

  3. Do the accuracy conclusions persist against exact one-dimensional Riemann solutions and a refined or higher-order two-dimensional reference? Please report interface-location and shock-local errors, conservation/jump diagnostics, and errors specifically after the last labeled time, rather than only global averages including t_0.

  4. How does the contribution differ from LeFloch and Xiang's RH-guided FNO/explicit-shock construction (arXiv:2607.23143), and does the sparse-label advantage persist against an appropriate weak-form or conservative physics-informed learner, such as an adapted UNFV? The joint ablation already supplied is helpful, but does not answer this comparison. Relevant primary sources: https://arxiv.org/abs/2607.23143 and https://arxiv.org/abs/2505.23702.

  5. How robust are the main and ablation differences across training seeds and label selections? Please also provide sufficient implementation detail to reproduce the exact two-dimensional perturbation and physical-coordinate differentiation through moving collocation points. If computational savings are intended beyond fewer state tuples, what are the numerical data-generation, training, and inference costs?

5 comments

Q1 & Weakness 1: Shallow-Water Wave Analysis and Euler Contact Discontinuity

Wave-type statistics and dominant-shock tracking in shallow water: The shallow-water dataset in Table 1 was sampled directly from the stated parameter ranges without filtering out two-shock states (Appendix A.1 & B.6). We verified the reviewer's calculation for the diagnostic example (h∗=1.47397,u∗=0.36180h_{\ast} = 1.47397, u_{\ast} = 0.36180, wave speeds −0.68150-0.68150 and 1.147361.14736). Across the 1024 test cases of this unfiltered dataset, 515 cases (50.3%) satisfy h∗≤hLh_{\ast} \le h_L and consist of a left-going smooth rarefaction wave and a right-going shock, while the remaining 509 cases (49.7%; training: 516/1024, validation: 65/128) satisfy h∗>hLh_{\ast} > h_L and contain two shocks. Because uL>0>uRu_L > 0 > u_R and hL>hRh_L > h_R, the h∗>hLh_{\ast} > h_L cases are strongly asymmetric: the right-going wave is the dominant shock (median depth ratio h∗/hR≈2.0h_{\ast}/h_R \approx 2.0), whereas the left-going shock has a small jump amplitude (median depth ratio h∗/hL≈1.12h_{\ast}/h_L \approx 1.12). Initialized at the largest initial jump ξ0(a)\xi_0(a), TrajectoryOp tracks the right-going shock across the dataset (t8t_8 position error of 1.451.45 cells), while the weak left-going wave stays inside the left region x<ξψ(t)x < \xi_\psi(t) and is approximated smoothly by the left decoder V^−\hat{V}^-.

Subgroup breakdown and single-shock (h∗≤hLh_{\ast} \le h_L) benchmark: Table 5 separates the unfiltered test set into these two subsets. Sparse TrajectoryOp is more accurate than Dense FNO on the h∗≤hLh_{\ast} \le h_L rarefaction-shock subset (0.01180.0118 vs. 0.01220.0122), whereas smoothing the weak left shock accounts for the higher error on the h∗>hLh_{\ast} > h_L subset (0.01710.0171 vs. 0.01330.0133). When we restrict the initial-state sampling to h∗≤hLh_{\ast} \le h_L so that every case contains a single shock and retrain all models (Table 6), sparse TrajectoryOp (3.125% label budget) closely matches Dense FNO in both L1L_1 error and shock position, while dense TrajectoryOp achieves the lowest error (0.00720.0072 vs. 0.01100.0110). We report this single-shock benchmark in the main text (Section 4.3, Table 2), and the unfiltered dataset remains in Table 1 with its subset analysis in Appendix B.6.

1D Euler contact discontinuity: As detailed in our reply to Reviewer nFCc (Q4), the contact discontinuity moves slower than the primary shock (u∗<sshocku_{\ast} < s_{\text{shock}}) and lies inside the left region x<ξψ(t)x < \xi_\psi(t) in 99.9% of test cases at t8t_8. Because velocity and pressure are continuous across a contact wave ([u]=[p]=0[u]=[p]=0), the left decoder V^−\hat{V}^- represents the density transition smoothly, contributing only 0.00360.0036 (∼8%\sim 8\%) to the relative L1L_1 error without disturbing the regional PDE residual (Appendix B.7).

Table 5: Subgroup breakdown on the original shallow-water test set (full space-time relative L1L_1 error against exact Riemann solutions)

Model & SupervisionOriginal Full Set (1024)h∗>hLh_{\ast} > h_L Subset (509)h∗≤hLh_{\ast} \le h_L Subset (515)
TrajectoryOp (Sparse)0.01440.01710.0118
FNO (Dense)0.01270.01330.0122

Table 6: Shallow-water evaluation with h∗≤hLh_{\ast} \le h_L (left-going smooth rarefaction and right-going shock, evaluated against exact Riemann solutions)

Model & SupervisionFull Space-Time L1L_1t5∼t8t_5 \sim t_8 Extrap. L1L_1t5∼t8t_5 \sim t_8 Shock Pos. Error
TrajectoryOp (Dense)0.00720.00820.69 cells
TrajectoryOp (Sparse)0.01150.01401.05 cells
FNO (Dense)0.01100.01261.02 cells
Shift-DeepONet (Sparse)0.03930.05938.66 cells
FNO (Sparse)0.04990.093879.51 cells

Q2: Boundary Conditions and Post-Hoc Entropy Admissibility Diagnostics

Diagnostics on the test sets confirm that the predicted solutions respect both properties (Appendix A.2 & B.7):

  • Boundary conditions: In 1D, all wave fronts over [0,T][0, T] remain strictly inside the interior of Ω\Omega (for 1D Euler on [−2,2][-2, 2], extreme wave positions lie in [−1.42,1.79][-1.42, 1.79]), and boundary-cell state variations across all three benchmarks are at most 8.7×10−68.7 \times 10^{-6}, so the extrapolation boundaries remain inactive. In 2D Euler, over t5∼t8t_5\sim t_8, the mean wall normal velocity violation ∣v∣\vert{}v\vert{} (exact reflecting condition: v=0v = 0) is 0.00060.0006 for sparse TrajectoryOp, compared with 0.00190.0019 for Dense FNO.
  • Entropy admissibility: Across all 1024 test cases and all nonzero times (t1∼t8t_1\sim t_8), the predicted shocks of Burgers, shallow water, and 1D Euler satisfy the Lax characteristic compression inequalities in 100% of cases. For 1D Euler, the thermodynamic entropy increases across the predicted shock in 84.3% of all case-time pairs.

Q3 & Weakness 2: Exact 1D Riemann Evaluation, 4×4\times-Refined 2D Convergence Check, and Late-Time/Conservation Metrics

Global exact-Riemann errors, t5∼t8t_5\sim t_8 extrapolation errors, shock-position errors, and local shock-band (±3\pm 3 cells) errors are reported in General Response Tables 1–2 (Section 4.3, Table 2 & Appendix B.1, Table 10). Table 7 below provides the t4/t8t_4/t_8 shock-position errors (comparing the maximum-gradient jump location across all models and the explicit trajectory ξψ\xi_\psi of TrajectoryOp) together with the conservation drift over t5∼t8t_5\sim t_8 (Appendix B.1, Table 11).

For the 2D benchmark, we recomputed 8 uniformly spaced test cases at 2×2\times (512×256512 \times 256) and 4×4\times (1024×5121024 \times 512) grid resolutions using the same Rusanov solver (observing an L1L_1 convergence order of 0.6∼0.70.6\sim 0.7 between 2×2\times and 4×4\times, consistent with a first-order scheme on discontinuous flows) and conservatively averaged the 4×4\times solution onto the base grid (Table 8; Appendix B.1, Table 12 & Figure 8). The original 1×1\times training labels themselves have an error of 0.00790.0079 at t8t_8 relative to the 4×4\times reference. Evaluated against the 4×4\times-refined reference, sparse TrajectoryOp achieves lower error than Dense FNO at all stored times, showing that Dense FNO's lower error on the 1×1\times grid comes from fitting the numerical dissipation of the first-order training labels. The revised manuscript reports both the 1×1\times numerical-grid metrics (Table 1) and the exact and refined-reference evaluations (Section 4.3, Table 2).

Table 7: Late-time shock position error (grid cells at t4/t8t_4 / t_8) and conservation drift (t5∼t8t_5 \sim t_8)

Diagnostic Metric & ModelBurgersShallow Water1D Euler2D Euler
Flow-Jump Pos. Error (t4/t8t_4 / t_8): TrajectoryOp (Sparse)0.61 / 1.210.67 / 1.294.76 / 8.660.91 / 2.64
Flow-Jump Pos. Error (t4/t8t_4 / t_8): FNO (Sparse)7.53 / 43.719.22 / 134.8113.16 / 62.1689.63 / 102.49
Flow-Jump Pos. Error (t4/t8t_4 / t_8): FNO (Dense)0.55 / 1.530.87 / 1.704.01 / 10.351.00 / 0.80
Explicit ξψ\xi_\psi Error (t4/t8t_4 / t_8): TrajectoryOp (Sparse)0.58 / 1.190.61 / 1.264.75 / 8.650.96 / 2.64
Conservation Drift (t5∼t8t_5 \sim t_8): TrajectoryOp (Sparse)0.00520.00740.0106--
Conservation Drift (t5∼t8t_5 \sim t_8): FNO (Sparse)0.04170.03890.0395--

Table 8: 2D Euler evaluation against 4×4\times-refined reference solutions (relative L1L_1 error across 8 test cases)

Reference / ModelFull Space-Timet1∼t4t_1 \sim t_4t5∼t8t_5 \sim t_8t8t_8
Original 1×1\times Numerical Labels0.00490.00400.00690.0079
TrajectoryOp (Sparse)0.00500.00410.00690.0081
FNO (Dense)0.00610.00540.00790.0087
FNO (Sparse)0.00890.00520.01430.0233

Q4: Comparison with LeFloch & Xiang (arXiv:2607.23143) and Weak-Form / Conservative Learners (arXiv:2505.23702)

Comparison with PITT (LeFloch & Xiang, 2026): We have added a citation and discussion in Section 2.2. While PITT also uses the Rankine-Hugoniot relation to advance shocks, it is formulated and demonstrated for one-dimensional scalar balance laws (the Schwarzschild–Burgers model): it builds the explicit steady-state and generalized Riemann solutions of this model into the network and advances shocks by an analytical RH law at each rollout step, and its extension to systems is left to a follow-up work. In contrast, TrajectoryOp learns the shock trajectory ξψ\xi_\psi as a continuous space-time operator output, applies to coupled conservation systems (shallow water, 1D Euler), and extends to 2D curved interfaces x=ξψ(y,t)x = \xi_\psi(y, t) (2D Euler shock–entropy interaction).

Comparison with weak-form and conservative finite-volume learners: Under the same 3.125% sparse budget, we added a single-decoder Weak-Form Residual baseline (integrating against smooth compactly supported test functions) and a Conservative Finite-Volume (FV) Residual baseline (penalizing one-step Rusanov flux residuals across cells, following UNFV; Section 4.4, Table 3 & Appendix B.3). As shown in General Response Table 2 (and the ablation table in Reviewer aDub Q2), TrajectoryOp achieves lower t5∼t8t_5\sim t_8 extrapolation error than both the single-decoder Weak-Form and Conservative FV baselines across all four benchmarks; moreover, pairing the same weak-form residual with our two-branch representation outperforms the single-decoder weak-form model on every benchmark, confirming the benefit of explicit interface separation beyond the choice of weak or conservative loss.

Q5: Multi-Seed Robustness, Implementation Details, and Computational Cost

We repeated the main experiments, including the single-decoder strong-form model, across 3 seeds (1234, 2345, and 3456), where each seed sets both the network initialization and the selection of sparse labeled cases and spatial points (Table 9; Appendix B.2, Table 13). Across all four benchmarks, sparse TrajectoryOp achieves lower t5∼t8t_5\sim t_8 extrapolation error than the single-decoder strong-form model, sparse FNO, sparse DeepONet, and sparse Shift-DeepONet in every seed (3/3), and dense TrajectoryOp outperforms dense FNO in every seed on Burgers, shallow water, and 2D Euler. In addition, while the shock-position error of sparse FNO fluctuates widely across seeds (27∼8027\sim 80 cells), sparse TrajectoryOp stays near 1 cell on Burgers and shallow water with small variance.

Reviewer_nFCcOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

TrajectoryOp is a physics-informed neural operator for conservation laws with a single explicitly tracked shock. It combines an FNO encoder, two regional state decoders, and a network predicting the shock trajectory. Training enforces the governing PDE within each predicted region and the Rankine–Hugoniot condition across the interface, alongside initial-condition and sparse state supervision. The intended contribution is to reduce reliance on numerical solution labels by explicitly representing and physically constraining the moving discontinuity.

The method is evaluated on Burgers, shallow-water, and one- and two-dimensional Euler problems. Using 3.125% of the dense reference budget of evolution-state labels, it outperforms the tested data-driven baselines under the same sparse budget. This percentage excludes the initial fields, which remain available to every method. Comparisons with dense FNO are mixed, including when TrajectoryOp receives the same dense label budget. The paper also provides supervision-budget studies and internal ablations of the physical losses and interface representation.

Critical strengths

1. A physically motivated representation of the moving discontinuity.

Separating the shock trajectory from the regional states is a meaningful architectural choice. It allows regional PDE constraints and interface conservation to be applied in their respective settings, while retaining an operator that generalizes across initial conditions rather than requiring a separate optimization for every instance. This is the paper’s strongest conceptual contribution.

2. Consistent improvements under the tested sparse-supervision budget, supported by internal controls.

TrajectoryOp achieves the lowest error among the compared methods in the sparse block on all four benchmarks. Importantly, the evidence is not limited to comparisons against data-only models: Tables 6 and 7 include physical-loss ablations and a physics-informed single-state decoder. These experiments provide useful support for the proposed combination of representation and physical constraints.

Critical weaknesses

1. The advantage over existing physics-informed alternatives is not sufficiently established.

The external baselines are trained without physical losses. The internal physics-informed control in Table 7 is valuable, but it removes both the explicit interface and its Rankine–Hugoniot constraint. Together with Table 6, this provides partial evidence about the components, but does not establish competitiveness against existing physics-informed operators designed to handle discontinuities. This matters because incorporating physics is itself a plausible source of improvement under sparse supervision; the additional benefit of this particular formulation should be demonstrated more directly.

2. The practical value of the label savings and the resources required for reproduction remain unclear.

The numerical labels are generated by simulation, so additional computation could instead be used to produce more training data. The paper does not provide the hardware and wall-clock measurements needed to assess that trade-off, and several implementation details remain underspecified. Consequently, I cannot determine when the reported label savings translate into a practical advantage or estimate the resources required for an independent reproduction. I am not assuming that TrajectoryOp is slower; the relevant measurements are missing.

Questions for authors
  1. Significant English Error in the Abstract The abstract contains a significant grammatical error: “TrajectoryOp achieves almost performance.” Could you explain how this was missed?
  2. Data Availability, Computational Cost, and Scope I am not sure the paper clearly distinguishes two cases:

When more data can be generated, as in these simulations, is it better to invest computation in TrajectoryOp or in generating more data for a standard operator? It is unclear which is preferable when accounting for both data generation and training. What hardware and computational budget were used, and how do runtimes compare with the baselines?

When more data cannot realistically be obtained, could you give a concrete practical example and explain the limitation?

An advantage under sparse supervision is a valid contribution; please make that scope explicit. If broader superiority is intended, experiments with substantially more data would be needed. 3. Comparison with Other Physics-Informed Methods Table 7 provides a useful internal physics-informed comparison. How would TrajectoryOp compare with existing shock-aware, physics-informed operators under the same data budget? Could a different representation with comparable physical constraints perform as well?

Since Table 7 removes both the explicit interface and the Rankine–Hugoniot constraint, could their contributions be distinguished further? 4. Additional Discontinuities As I understand it, the 1D Euler experiment includes a contact discontinuity inside a region assumed to be smooth. Is this problematic for the regional PDE residual? How is it handled? 5. Two-Dimensional Shock Deformation In Figure 4(d), the numerical reference shock appears curved, while the predicted TrajectoryOp interface remains nearly straight at the final time. How does this support the claim that the model recovers the deformed shock? Could you quantify the interface-position error? 6. Implementation and Reproducibility With the hard selection in Eq. (6), how does the trajectory network receive gradients, particularly when the Rankine–Hugoniot loss is removed?

Could you provide more complete architecture, sampling, and training details in an appendix or code? I am not sure an independent researcher or coding agent could reproduce the experiments and assess computational efficiency without guessing important implementation choices. 7. Other Grammatical Errors There are further grammatical errors, including “an explicitly shock trajectory” and “is composed by.” I would expect a submission to a conference of this level to be free of such basic errors.

10 comments

Q1 & Q7: Grammatical Errors in the Abstract and Main Text

We apologize for the proofreading oversights. We have corrected the missing word in the abstract, revised "an explicitly shock trajectory" to "an explicit shock trajectory" and "is composed by" to "is composed of" in the main text, and carefully proofread the entire manuscript to fix all remaining grammatical issues.

Q2 & Weakness 2: Data Scaling, Scope of Applicability, and Runtime Profile

Comparison with Dense FNO and scaling training data (n=64→1024n = 64 \to 1024): In the original Table 1, evaluating against first-order 1×1\times Rusanov grids favors Dense FNO because the model directly fits the solver's numerical smearing across shocks. Evaluated against exact 1D Riemann solutions and 4×4\times-refined 2D references (General Response Table 1; Section 4.3, Table 2), sparse TrajectoryOp (3.125% budget) outperforms Dense FNO on Burgers and 2D Euler and matches it on single-shock shallow water, while dense TrajectoryOp achieves the lowest error on Burgers and shallow water. Furthermore, when we increase the training trajectories of Dense FNO from n=64n=64 to 256256 and 10241024 (Table 4; Appendix B.5, Table 15), its exact-solution error stops decreasing beyond n=256n=256 on Burgers and shallow water because additional trajectories come from the same first-order solver and carry the same shock smearing. By separating the jump via an explicit interface and enforcing regional PDEs and the RH relation, TrajectoryOp avoids smoothing the discontinuity and achieves lower errors on Burgers (n=16,64n=16, 64) and shallow water (n=64n=64) than Dense FNO trained with n=1024n=1024 trajectories.

Scope of applicability and practical use cases: TrajectoryOp is designed for two practical regimes: (i) incomplete temporal or spatial observations (t≤T/2t \le T/2), such as shock-tube high-speed schlieren or flash X-ray diagnostics that capture only early snapshots (t1∼t4t_1\sim t_4) at sparse sensors before the wave leaves the viewing window—where adding more runs nn cannot supply late-time supervision (t5∼t8t_5\sim t_8) and data-driven operators lose track of the shock, whereas TrajectoryOp advances the front via the RH condition; and (ii) reducing fine-grid simulation costs, where reducing numerical shock smearing for a purely data-driven operator requires refining the solver grid in both space and time (for example, 4×4\times grid refinement in 2D increases solver time from ∼1.5 s\sim 1.5\text{ s} to ∼90 s\sim 90\text{ s} per case), whereas TrajectoryOp recovers sharp shocks and accurate shock positions directly from coarse 1×1\times labels.

Hardware and runtime profile: On a single NVIDIA RTX 2080 Ti GPU (Appendix B.5), generating 64 coarse 1D cases takes 6.6∼9.3 s6.6\sim 9.3\text{ s} and one coarse 2D case takes ∼1.5 s\sim 1.5\text{ s} (the sparse T/2T/2 setting uses 12.5% of the dense solver rollout cost). Training for 5000 epochs takes 26∼43 min26\sim 43\text{ min} for 1D TrajectoryOp and ∼0.9 h\sim 0.9\text{ h} in 2D, compared with 8∼9 min8\sim 9\text{ min} for 1D FNO (n=1024n=1024) and 18 min18\text{ min} for 2D FNO (n=128n=128). At inference time, no derivatives are required, and TrajectoryOp predicts both the state field and the shock trajectory ξψ\xi_\psi in ∼17 ms\sim 17\text{ ms} per case (∼0.9 ms\sim 0.9\text{ ms} for FNO).

Table 4: Scaling Dense FNO training trajectories (n=64→1024n = 64 \to 1024) vs. TrajectoryOp in the extrapolation interval (t5∼t8t_5\sim t_8, relative L1L_1 error / shock position error in cells; Revised Manuscript Appendix B.5, Table 15)

Model & Trajectory Count (n)Burgers (L1 / Pos.)Shallow Water (L1 / Pos.)1D Euler (L1 / Pos.)2D Euler (L1)
TrajectoryOp (Sparse, n=16n=16)0.0090 / 0.980.0140 / 1.050.0615 / 7.190.0059
TrajectoryOp (Dense, n=64n=64)0.0070 / 0.960.0082 / 0.690.0368 / 3.960.0011
Dense FNO (n=64n=64)0.0098 / 0.910.0126 / 1.020.0336 / 7.000.0016
Dense FNO (n=256n=256)0.0100 / 0.960.0111 / 0.780.0269 / 4.720.0015
Dense FNO (n=1024n=1024; 512 in 2D)0.0104 / 0.950.0120 / 0.740.0255 / 4.450.0015

Q3 & Weakness 1: Comparison with Physics-Informed Alternatives and Component Decoupling

Please refer to Table 2 in the General Response and our reply to Reviewer aDub (Q1 & Q2), where we compare TrajectoryOp against discontinuity-aware operators (Cut-DeepONet, RiemannONets) and single-decoder physics-informed baselines (global strong-form PDE, weak-form residual, and conservative finite-volume residual), and separate the contribution of the two-branch representation from that of LRH\mathcal{L}_{RH} (Revised Manuscript Section 4.4, Table 3 & Appendix B.3–B.4).

Q4: Handling the Contact Discontinuity in 1D Euler

Following the hybrid front-tracking/capturing paradigm in computational fluid dynamics (Mao, 1991; Glimm et al., 1998; Shyue, 1998), TrajectoryOp explicitly tracks the primary nonlinear shock ξψ(t)\xi_\psi(t) while representing the contact discontinuity within the left regional decoder V^−\hat{V}^-. Because the contact wave propagates at the local fluid velocity u∗u_{\ast}, which is slower than the right-going shock speed, it remains inside the left subdomain x<ξψ(t)x < \xi_\psi(t) (in 99.9% of test cases at t8t_8). Unlike a genuinely nonlinear shock where intersecting characteristics cause discontinuous jumps in all state variables, a contact discontinuity is a linearly degenerate wave across which velocity and pressure remain strictly continuous ([u]=0,[p]=0[u]=0, [p]=0) and characteristics run parallel. Consequently, the spatial derivatives ∂xu\partial_x u and ∂xp\partial_x p in the Euler equations remain smooth across the contact wave, while the left KAN branch V^−\hat{V}^- represents the advected density jump as a steep, differentiable transition layer matching the finite-volume training data. Across all 1024 test cases, the local band (±5\pm 5 cells around the exact contact location) contributes 0.00360.0036 to the relative L1L_1 error (about 8% of the total error under sparse supervision and 9% under dense supervision; Appendix B.7), confirming that it does not destabilize the regional PDE residual.

References:

  • De-Kang Mao. A treatment of discontinuities in shock-capturing finite difference methods. Journal of Computational Physics, 92(2):422–455, 1991.
  • James Glimm, John W. Grove, Xiao Lin Li, Keh-Ming Shyue, Yanni Zeng, and Qiang Zhang. Three-dimensional front tracking. SIAM Journal on Scientific Computing, 19(3):703–727, 1998.
  • Keh-Ming Shyue. An efficient shock-capturing algorithm for compressible multicomponent problems. Journal of Computational Physics, 142(1):208–242, 1998.

Q5: 2D Shock Deformation and Interface Error Quantification

In the 2D shock-entropy interaction (Figure 5(d), formerly Figure 4(d)), the upstream circular density perturbation deforms the downstream acoustic/entropy field as it passes through the shock and induces a transverse bending of the shock front spanning about 8.48.4 base-grid cells at t8=Tt_8 = T in the 4×4\times-refined reference solution. We have revised the caption in Appendix A.3 to state that TrajectoryOp recovers the streamwise shock propagation and the deformed downstream interaction field beyond the labeled interval. Evaluated against the pressure-jump front of the 4×4\times-refined reference solution (Appendix A.3 & B.1, Table 11, Figure 8), sparse TrajectoryOp—without any state labels after t4t_4—achieves a mean absolute front position error of 0.960.96 cells at t4t_4 and 2.642.64 cells at t8t_8, compared with 89.6389.63 and 102.49102.49 cells for sparse FNO.
Q6: Gradient Flow under Piecewise Selection, Meaning of the Controlled Ablation, and Reproducibility Details

Under the piecewise selection of Eq. (6), the trajectory network is trained by the RH loss. LRH\mathcal L_{RH} depends on ξψ\xi_\psi through the shock speed ∂tξψ\partial_t\xi_\psi and through the interface points (t,ξψ)(t,\xi_\psi) at which both regional decoders V^−\hat V^- and V^+\hat V^+ are evaluated, so the shock moves at the speed that the predicted states on its two sides require. The state loss and the regional PDE loss train the two regional decoders. Without the RH loss, the trajectory therefore remains at its initial estimate ξ0(a)\xi_0(a) (Appendix A.5, Table 8).

To study the interface representation without the RH loss, the controlled ablation replaces the step in Eq. (6) during training by the smooth weight wϵ=σ((x−ξψ)/ϵ)w_\epsilon=\sigma((x-\xi_\psi)/\epsilon), through which the state loss and the regional PDE loss also move ξψ\xi_\psi; Eq. (6) is kept at evaluation (General Response Table 2; Section 4.4, Table 3). With the same weak-form residual, this two-branch model is more accurate than the single decoder on all four benchmarks, and TrajectoryOp, which couples the shock motion to the two regional states through the RH relation, gives the lowest error on Burgers.

Algorithm A summarizes the collocation and automatic-differentiation procedure. Complete architectural hyperparameters, sampling counts, and the 2D perturbation formula are given in Appendix A, and an anonymous code repository is available at: https://anonymous.4open.science/r/TrajectoryOp-123456/

Algorithm A (Revised Manuscript Appendix A, Algorithm 2): regional PDE and interface RH residuals for one case aa. In 1D, the transverse coordinate yy and the terms involving F^y\hat F_y and ξψ,y\xi_{\psi,y} are omitted.

Input: latent field Hh(a)H_h(a); regional decoders V^−\hat V^- and V^+\hat V^+; trajectory network ξψ\xi_\psi; domain points (tk,xk,yk)(t_k,x_k,y_k), k=1,…,MPDEk=1,\ldots,M_{\mathrm{PDE}}; interface coordinates (tm,ym)(t_m,y_m), m=1,…,MRHm=1,\ldots,M_{\mathrm{RH}}.

Output: LPDE\mathcal L_{\mathrm{PDE}} and LRH\mathcal L_{\mathrm{RH}}.

  1. for k=1,…,MPDEk=1,\ldots,M_{\mathrm{PDE}} do
  2.   σk←−\sigma_k\leftarrow - if xk<ξψ(a;yk,tk)x_k<\xi_\psi(a;y_k,t_k), otherwise σk←+\sigma_k\leftarrow +
  3.   U^σk←C(V^σk(a;tk,xk,yk))\hat U^{\sigma_k}\leftarrow\mathcal C(\hat V^{\sigma_k}(a;t_k,x_k,y_k)), F^xσk←Fx(U^σk)\hat F_x^{\sigma_k}\leftarrow F_x(\hat U^{\sigma_k}), F^yσk←Fy(U^σk)\hat F_y^{\sigma_k}\leftarrow F_y(\hat U^{\sigma_k})
  4.   RPDEσk←∂tU^σk+∂xF^xσk+∂yF^yσkR_{\mathrm{PDE}}^{\sigma_k}\leftarrow\partial_t\hat U^{\sigma_k}+\partial_x\hat F_x^{\sigma_k}+\partial_y\hat F_y^{\sigma_k}, by automatic differentiation in (t,x,y)(t,x,y)
  5. end for
  6. LPDE←∑σ=±⟨∥RPDEσ∥22⟩\mathcal L_{\mathrm{PDE}}\leftarrow\sum_{\sigma=\pm}\langle\lVert R_{\mathrm{PDE}}^{\sigma}\rVert_2^2\rangle, where each average is taken over the points with σk=σ\sigma_k=\sigma
  7. for m=1,…,MRHm=1,\ldots,M_{\mathrm{RH}} do
  8.   xm←ξψ(a;ym,tm)x_m\leftarrow\xi_\psi(a;y_m,t_m); obtain ξψ,t\xi_{\psi,t} and ξψ,y\xi_{\psi,y} at (ym,tm)(y_m,t_m) by automatic differentiation
  9.   evaluate V^−\hat V^- and V^+\hat V^+ at (tm,xm,ym)(t_m,x_m,y_m) and form the jumps [U^][\hat U], [F^x][\hat F_x], [F^y][\hat F_y]
  10.   RRH←−ξψ,t[U^]+[F^x]−ξψ,y[F^y]R_{\mathrm{RH}}\leftarrow-\xi_{\psi,t}[\hat U]+[\hat F_x]-\xi_{\psi,y}[\hat F_y]
  11. end for
  12. LRH←⟨∥RRH∥22⟩\mathcal L_{\mathrm{RH}}\leftarrow\langle\lVert R_{\mathrm{RH}}\rVert_2^2\rangle, averaged over m=1,…,MRHm=1,\ldots,M_{\mathrm{RH}}
Reviewer_aDubOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes TrajectoryOp for conservation laws with a single shock. It uses an FNO encoder, two regional state decoders, and a shock trajectory network. Training combines sparse state labels with PDE residuals and Rankine–Hugoniot constraints. Experiments on Burgers', shallow-water, and 1D/2D Euler problems show better results than data-driven baselines under the same supervision budget. However, the experiments do not clearly separate the contributions of the explicit interface design and the additional physical constraints.

Critical strengths

1.The explicit shock representation is physically motivated. Separating the two regional states and constraining the interface through the Rankine–Hugoniot condition is a coherent approach to handling discontinuities. 2.The experiments cover four conservation-law benchmarks, and the component ablations provide initial evidence that both regional PDE residuals and interface constraints contribute to performance under sparse supervision.

Critical weaknesses

1.The novelty over existing interface-aware and physics-informed methods remains unclear. The paper does not clearly establish what is fundamentally new beyond combining explicit shock tracking with physical constraints. It also lacks direct comparisons with related methods for discontinuous solutions, such as RiemannONets, R-adaptive DeepONet, and Cut-DeepONet, making it difficult to assess its advantages over existing approaches. 2.The experiments do not clearly isolate the benefit of the explicit-interface design. The baselines lack comparable physical constraints, while the ablation removes both the interface representation and its associated jump constraint. Therefore, the results do not establish whether the gains come from the explicit interface itself or mainly from the additional physical constraints.

Questions for authors

1.How does TrajectoryOp differ fundamentally from existing interface-aware and physics-informed methods, beyond combining explicit shock tracking with physical constraints? 2.Table 7 removes both the explicit-interface representation and its associated Rankine–Hugoniot constraint. Can the authors clarify what evidence supports the independent contribution of the interface design, rather than the effect of the additional physical constraints?

3 comments
AuthorsOpenReview ↗

Response to Reviewer aDub

Q1 & Weakness 1: Fundamental Difference from Existing Methods and Comparison with Discontinuity-Aware Operators

Modeling Difference (Explicit Interface Dynamics vs. Unconstrained Parameterization)

The core contribution of TrajectoryOp lies in formulating the moving shock interface as an explicit operator output ξψ\xi_\psi separating two regional state decoders (V^−,V^+)(\hat{V}^-, \hat{V}^+), where the one-sided state jumps [U^][\hat{U}] and [F^][\hat{\mathcal{F}}] govern the shock speed ∂tξψ\partial_t \xi_\psi through the Rankine-Hugoniot (RH) condition over the entire space-time domain. As also noted by Reviewer nFCc and Reviewer suaG, separating the shock trajectory from the two regional state decoders allows smooth PDE residuals and the Rankine-Hugoniot (RH) jump condition to be imposed in their respective domains while avoiding the need to fit a sharp jump through a single continuous network:

  1. Existing discontinuity-aware operators—such as RiemannONets (learned basis functions), R-adaptive DeepONet (coordinate warping), and Cut-DeepONet (piecewise lifting via a step feature H(x−ξ)H(x-\xi))—change the spatial parameterization so that sharp gradients can be represented. However, in these methods, the discontinuity location is not constrained by the conservation law and remains decoupled from the state values on the two sides; consequently, the model must infer the shock trajectory entirely from dense evolution-state labels. Our experiments in Table 3 directly confirm this limitation: once state labels stop at t4=T/2t_4 = T/2, these methods have no mechanism to advance the interface over t5∼t8t_5\sim t_8, leading to large extrapolation errors (0.0680∼0.17170.0680\sim 0.1717 for Cut-DeepONet and 0.4823∼1.15040.4823\sim 1.1504 for RiemannONets).
  2. Evolution-step operators with embedded RH updates (PITT, LeFloch & Xiang, 2026): Among operators that learn discrete time-stepping updates, PITT is the closest to our work as it also uses the RH relation to advance shocks. However, PITT is formulated and demonstrated for one-dimensional scalar balance laws (the Schwarzschild–Burgers model): it builds the explicit steady-state and generalized Riemann solutions of this model into the network and advances shocks by an analytical RH law at each rollout step, and its extension to systems is left to a follow-up work. In contrast, TrajectoryOp learns the shock trajectory ξψ\xi_\psi as a continuous space-time operator output, applies to coupled conservation systems (shallow water, 1D Euler), and extends to 2D curved interfaces x=ξψ(y,t)x=\xi_\psi(y,t) (2D Euler shock–entropy interaction).

Advantages of This Formulation & Quantitative Comparison (Table 3): We evaluated Cut-DeepONet and RiemannONets under the exact same data splits and label budgets in both sparse (n=16,s=64,t1∼t4n=16, s=64, t_1\sim t_4, 3.125% budget) and dense (n=64,s=256,t1∼t8n=64, s=256, t_1\sim t_8, 100% budget) settings (Revised Manuscript Appendix B.4, Table 14). (R-adaptive DeepONet is not included because no reference code is publicly available.) Modeling the interface explicitly as an RH-coupled output brings three concrete benefits:

  1. Higher accuracy with far fewer state labels: Under the 3.125% sparse budget, TrajectoryOp reduces the t5∼t8t_5\sim t_8 error of Cut-DeepONet by 64%–87%. Even when Cut-DeepONet and RiemannONets are trained with 100% dense labels across the full time horizon, sparse TrajectoryOp still achieves 32%–79% lower error than dense Cut-DeepONet and 80%–96% lower error than dense RiemannONets.
  2. Accurate shock localization in unseen time intervals: Over t5∼t8t_5\sim t_8, sparse TrajectoryOp keeps the shock position error at about one grid cell on Burgers (0.98 cells) and shallow water (1.05 cells), compared with 27–80 cells for sparse FNO (General Response Table 1). Revised Manuscript Figure 3 plots the shock position error at each time level: the error of sparse FNO grows rapidly once state labels stop at t4t_4, whereas that of sparse TrajectoryOp stays at the level of dense FNO; Figure 7 compares the predicted shock trajectories with the exact ones on representative test cases.
  3. Overcoming the numerical dissipation of coarse training labels: Because the jump is handled by the interface selection rather than smoothed by a continuous trunk, TrajectoryOp avoids fitting the numerical smearing of the first-order training solver, achieving lower exact-solution errors on Burgers and shallow water than Dense FNO trained with 16×16\times more trajectories (n=1024n=1024, Reviewer nFCc Table 4).

Table 3: Comparison with discontinuity-aware operators in the unlabeled extrapolation interval (t5∼t8t_5\sim t_8, relative L1L_1 error against exact Riemann solutions; Revised Manuscript Appendix B.4, Table 14)

Method & Supervision BudgetBurgersShallow Water1D Euler
TrajectoryOp (Sparse, 3.125%)0.00900.01400.0615
Cut-DeepONet (Sparse, 3.125%)0.06800.07230.1717
RiemannONets (Sparse, 3.125%)0.62730.48231.1504
TrajectoryOp (Dense, 100%)0.00700.00820.0368
Cut-DeepONet (Dense, 100%)0.03320.04040.0905
RiemannONets (Dense, 100%)0.20720.12080.3173

Q2 & Weakness 2: Separating the Explicit Interface from the Physical Constraints

The original Table 7 compares TrajectoryOp with a single decoder trained with a global strong-form residual, so the representation and the loss change together. TrajectoryOp differs from a single-decoder model in two respects. Its solution is represented by two regional decoders V^±\hat V^\pm separated by the predicted shock location ξψ\xi_\psi (Eq. 6), which is what we mean by the explicit-interface design. This representation also allows two physical losses: a PDE residual LPDE\mathcal L_{\mathrm{PDE}} inside each region, and an RH residual LRH\mathcal L_{RH} that ties the shock speed ∂tξψ\partial_t\xi_\psi to the jumps across the shock. In the revision we change these one at a time under the same 3.125% sparse budget. The results are collected in the table below and reported in Section 4.4 (Table 3) and Appendix B.3.

We first keep the loss fixed and change only the representation. The single decoder is now trained with three physical losses of its own, so that the baselines have physical constraints comparable to ours: the global strong-form residual (Row 3), a weak-form residual (Row 2), and a conservative finite-volume residual (General Response Table 2). For the two-branch model we remove LRH\mathcal L_{RH}, and during training we replace the step in Eq. (6) by a smooth weight wϵ=σ((x−ξψ)/ϵ)w_\epsilon=\sigma((x-\xi_\psi)/\epsilon), with ϵ\epsilon decreasing from 32 cells to 1 cell, so that the state and PDE losses can move the interface. Eq. (6) is kept at evaluation. With the same weak-form residual, the two-branch model is more accurate than the single decoder on all four benchmarks (Row 5 vs. Row 2), and without any physical loss it is more accurate on the three 1D benchmarks (Row 4 vs. Row 1). Each regional decoder only has to represent a smooth state, whereas a single decoder has to fit the jump.

We then keep the two-branch representation fixed and change the physical losses. The regional PDE loss alone (Row 6) does not determine how the shock moves after the last labeled time, and on 1D Euler the trajectory drifts away from the shock. The RH loss alone (Row 7) constrains the shock speed but not the states inside each region. With both losses (Row 8) the error is lower than with either one on all four benchmarks, and the piecewise form of Eq. (6) (Row 9, TrajectoryOp) keeps the shock sharp, lowering the Burgers error from 0.0154 to 0.0090. None of the single-decoder models with physical losses is as accurate as TrajectoryOp on any benchmark.

The explicit interface therefore improves accuracy when the loss is held fixed, and the physical losses improve accuracy when the representation is held fixed. The RH loss in particular can only be imposed because the interface is an explicit output, which is the main difference from the methods discussed in Q1.

Summary of Controlled Ablations (t5∼t8t_5\sim t_8 relative L1L_1 error under 3.125% sparse supervision; excerpted from General Response Table 2 / Revised Manuscript Table 3)

RowSolution RepresentationPhysical ConstraintBurgersShallow Water1D Euler2D Euler
1Single DecoderNone (Lstate\mathcal{L}_{\text{state}} only)0.42970.23580.20590.0129
2Single DecoderWeak-Form Residual0.02190.02870.06770.0104
3Single DecoderGlobal Strong PDE (Orig. Table 7)0.05140.03640.06760.0069
4Two-Branch, Smooth (wϵw_\epsilon)None (Lstate\mathcal{L}_{\text{state}} only)0.07430.11430.13740.0180
5Two-Branch, Smooth (wϵw_\epsilon)Weak-Form Residual0.01270.01900.05960.0051
6Two-Branch, Smooth (wϵw_\epsilon)Regional LPDE\mathcal{L}_{\text{PDE}} only0.03820.03640.38590.0454
7Two-Branch, Smooth (wϵw_\epsilon)LRH\mathcal{L}_{RH} only0.08070.08970.21140.0103
8Two-Branch, Smooth (wϵw_\epsilon)Regional LPDE+LRH\mathcal{L}_{\text{PDE}} + \mathcal{L}_{RH}0.01540.01290.05630.0066
9Two-Branch, Piecewise (Eq. 6, TrajectoryOp)Regional LPDE+LRH\mathcal{L}_{\text{PDE}} + \mathcal{L}_{RH}0.00900.01400.06150.0059

Comments on the paper

2 comments
AuthorsOpenReview ↗

Common Response to Area Chair and All Reviewers

We thank the Area Chair and the reviewers for their comments. In TrajectoryOp the shock location is an output of the operator, and its motion is determined by the Rankine–Hugoniot relation between the predicted states on the two sides. Existing operators for discontinuous solutions either learn the shock location from state labels or apply physical losses to a representation without an interface. The three questions raised by more than one reviewer concern this difference, and we answer them below.

  • Comparison with existing operators for discontinuous solutions (Reviewer aDub W1, Q1; Reviewer nFCc W1, Q3; Reviewer suaG W2, Q4): Existing operators either learn where the shock is from state labels, or enforce conservation on a field that has no interface. Cut-DeepONet and RiemannONets belong to the first group. Under the same splits and budgets, sparse TrajectoryOp is more accurate on the 1D benchmarks than both of them, even when they are trained with 100% dense labels (Reviewer aDub Table 3). For the second group, we implemented a conservative finite-volume learner following UNFV. It is less accurate than TrajectoryOp on all four benchmarks under the same sparse budget (Table 2). PITT also advances shocks with the RH relation, but it is formulated for 1D scalar balance laws. TrajectoryOp learns the shock trajectory as an operator output for coupled systems and 2D curved interfaces. These comparisons are added to Section 2.2, Section 4.4 (Table 3) and Appendix B.4 (Table 14).
  • Separating the explicit interface from the physical constraints (Reviewer aDub W2, Q2; Reviewer nFCc W1, Q6; Reviewer suaG W2): The original Table 7 changed the representation and the loss together, so we now change one at a time. When the same weak-form residual is applied to both a single decoder and the two-branch representation, the two-branch representation is more accurate on all four benchmarks. When the two-branch representation is fixed, neither the regional PDE loss nor the RH loss is sufficient alone, and using both gives the lowest error of the three (Table 2). The two parts therefore both contribute. The RH loss can be imposed only because the interface is an explicit output, and a single decoder does not have one. These ablations are reported in Section 4.4 (Table 3), Appendix A.5–A.6 and Appendix B.3, and the residual evaluation is given in Appendix A, Algorithm 2.
  • The shock stays in place where no labels exist (Reviewer suaG W2, Q3; Reviewer nFCc W2, Q2, Q5): The training labels come from a first-order Rusanov solver, which spreads each shock over several cells. We therefore evaluate against exact 1D Riemann solutions (Table 1) and a 4×4\times-refined 2D reference (Reviewer suaG Table 8). Against these references, sparse TrajectoryOp uses 3.125% of the evolution-state labels. It is more accurate than Dense FNO on Burgers and 2D Euler and matches it on shallow water. The labels end at t4t_4, yet over t5∼t8t_5\sim t_8 the shock position error on Burgers and shallow water remains about one grid cell, whereas sparse FNO is off by tens of cells (Table 1; Figure 3 of the revised manuscript). This is the practical consequence of constraining the shock motion: the RH relation keeps advancing the interface after the state labels stop. These results are reported in Section 4.3 (Table 2, Figure 3) and Appendix B.1 (Tables 10–12, Figures 7–8). Results over three training seeds, Dense FNO trained on up to 1024 trajectories, and runtimes are given in the responses to Reviewer suaG (Q5) and Reviewer nFCc (Q2).

Table 1: 1D evaluation against exact Riemann solutions (full space-time relative L1L_1 error, shock position error in grid cells over unlabeled t5∼t8t_5 \sim t_8, and local shock-band ±3\pm 3 cells L1L_1 error; Revised Manuscript Section 4.3, Table 2 & Appendix B.1, Table 10)

Evaluation Metric & ModelBurgersShallow Water1D Euler
Full Space-Time L1L_1: TrajectoryOp (Sparse)0.00810.01150.0469
Full Space-Time L1L_1: TrajectoryOp (Dense)0.00620.00720.0310
Full Space-Time L1L_1: FNO (Dense)0.00880.01100.0267
Shock Position Error (cells, t5∼t8t_5 \sim t_8): TrajectoryOp (Sparse)0.981.057.19
Shock Position Error (cells, t5∼t8t_5 \sim t_8): TrajectoryOp (Dense)0.960.693.96
Shock Position Error (cells, t5∼t8t_5 \sim t_8): FNO (Sparse)27.2879.5138.78
Shock Position Error (cells, t5∼t8t_5 \sim t_8): FNO (Dense)0.911.027.00
Local Shock-Band (±3\pm 3 cells) L1L_1: TrajectoryOp (Sparse)0.00180.00130.0064
Local Shock-Band (±3\pm 3 cells) L1L_1: FNO (Sparse)0.00710.00460.0065
Local Shock-Band (±3\pm 3 cells) L1L_1: FNO (Dense)0.00340.00300.0049

*Note: Mean over 3 training seeds (standard deviations in Reviewer suaG Q5, Table 9). Shallow water denotes the strict single-shock benchmark (h∗≤hLh_\ast \le h_L, detailed in Reviewer suaG Q1). The 1st-order Rusanov training labels themselves have intrinsic L1L_1 discretization errors of 0.00240.0024, 0.00500.0050, and 0.01410.0141 against exact Riemann solutions on the three 1D benchmarks.

Table 2:Controlled ablation separating the two-branch solution representation from physical constraints, and comparison with physics-informed and discontinuity-aware baselines (t5∼t8t_5 \sim t_8 relative L1L_1 error under 3.125% sparse supervision, n=16,t1∼t4n=16, t_1 \sim t_4; Revised Manuscript Section 4.4, Table 3 & Appendix B.3–B.4)

Representation & Physics ConfigurationBurgersShallow Water1D Euler2D Euler
TrajectoryOp (Hard Piecewise + Regional PDE + RH)0.00900.01400.06150.0059
Two-Branch Smooth (Regional PDE + RH)0.01540.01290.05630.0066
Two-Branch Smooth (Weak-Form Residual)0.01270.01900.05960.0051
Two-Branch Smooth (Regional PDE Only)0.03820.03640.38590.0454
Two-Branch Smooth (RH Only)0.08070.08970.21140.0103
Two-Branch Smooth (No Physics)0.07430.11430.13740.0180
Single Decoder (Weak-Form Residual)0.02190.02870.06770.0104
Single Decoder (Global Strong PDE, Orig. Table 7)0.05140.03640.06760.0069
Single Decoder (Conservative FV Residual)0.04820.03920.07520.0091
Single Decoder (No Physics)0.42970.23580.20590.0129
FNO (No Physics)0.07180.09380.10920.0100
Cut-DeepONet (No Physics)0.06800.07230.1717--
RiemannONets (No Physics)0.62730.48231.1504--

Note: 1D benchmarks are evaluated against exact Riemann solutions; 2D Euler is evaluated against the numerical reference (4x-refined 2D evaluation is provided in Reviewer suaG Q3). Cut-DeepONet and RiemannONets are evaluated on the 1D benchmarks, following the 1D Riemann-problem settings for which they were proposed.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.