acceptodds
Under review as a conference paper at ICLR 2027

Tracking the Discontinuity: Explicit Front Tracking for Neural Euler Solvers

Abstract

Physics-informed neural networks smear shocks because of their objective, not their optimiser: across a sharp shock whose two states satisfy the Rankine-Hugoniot conditions, the pointwise strong-form residual scales as the inverse of the width, so its squared norm is minimised by spreading the jump. We develop explicit front tracking for the compressible Euler equations, in which smooth networks are separated by fronts that move by their jump conditions and the residual is never evaluated at a discontinuity. Four components make this work for a system with several interacting wave families. The number, characteristic family and admissibility of the fronts are derived from the exact star state of the Riemann data, with a contact kept only if it carries strength. Two-dimensional fronts are level sets anchored to the initial discontinuity, with the jump conditions imposed along the learned normal at points carried onto the front by Newton projection. The plateaus beside a shock, which lie in the null space of the residual, are anchored through the jump conditions to the exact far-field states rather than to neighbouring learned ones. And the best of several runs is selected by an uncalibrated physics score, without ground truth. On Riemann problems with discontinuities the method is 8 to 14 times more accurate than strong-form and relaxation networks at the same budget on Sod and 386 to 879 times on a problem of two shocks and a contact, anchoring improves the latter a further sixfold, and selection finds the most accurate of three seeds in 11 of 13 settings, against one in three by chance. The method reproduces a planar oblique shock in two dimensions and, on a nonconservative system, realises the jump condition of each prescribed path, where a standard path-conservative scheme responds with the wrong sign. It does not yet handle a curved front, and a classical solver remains three to four orders of magnitude cheaper whenever a rarefaction is present.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Physics-Informed Neural Networks, Scientific Machine Learning, Compressible Euler Equations, Front Tracking, Level-Set Method

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_ZArSOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies neural solution of compressible Euler equations with discontinuities. Its central observation is that the failure of strong-form PINNs near shocks is intrinsic to the objective: as a neural network sharpens a Rankine-Hugoniot-consistent shock, the pointwise residual grows rather than decreases. The proposed solution is therefore not to improve optimization, but to change the representation itself. The solution is represented by separate smooth neural branches divided by explicitly tracked discontinuities; PDE residuals are evaluated only in smooth regions, while the fronts are governed by Rankine-Hugoniot and Lax admissibility conditions.

Critical strengths

The paper identifies and directly attacks the correct failure mode. Rather than adding increasingly sophisticated optimization machinery to a fundamentally problematic strong-form objective, the paper changes the representation so that the residual is never evaluated at the discontinuity.

Critical weaknesses

The current construction relies on unusually strong prior knowledge and is not yet a general Euler solver. The most important issue is that the method requires the exact Riemann star state to determine which fronts exist and their characteristic families. For Riemann data, however, that star state essentially already determines the exact solution. The authors themselves explicitly acknowledge that “a method that needs Riemann data to set up its ansatz is not a general solver.” Furthermore, the 1D parameterization prevents front interactions, shock formation is not addressed, and even a single genuinely curved 2D front currently fails.

Questions for authors

Could you please explain or address the weaknesses raised above?

1 comment

Thank you for the accurate summary. We agree with each limitation you list (as they are explicitly stated on out paper, and the sentence you quote is our own assessment of the same limits. We try to address each of the points raised:

What the star state is used for. It decides only which fronts exist and their families; its values never enter the loss (Section 3.2, Figure 6). You are right that for Riemann data it already determines the exact solution. We chose Riemann problems for that reason: an exact reference lets us attribute the error by region (Appendix E). As Section 5.5 states, the comparison measures how well the ansatz represents a known solution, it's not stated as a use case.

Why we did not learn the structure: We did try to and the evidence is presented in the paper too. When the network chose the family itself, one seed in three chose the wrong one on Sod and placed the front 0.256 from the shock (Appendix B). The prior knowledge is doing real work, and we do not claim otherwise.

What does not depend on the star state.

The diagnosis. The scaling argument and its numerical check in Section 1 hold for any shock. The two-dimensional oblique test. It is given the initial front, the family and boundary data, with no Riemann solve (Section 5.4). The selection score. It uses no reference solution, although our argument for it is one-dimensional (Section 3.5). The prescribed-path study. Appendix K uses a nonconservative system where the jump condition is supplied.

Interaction, formation and curved fronts. All three are open.

Interaction. The one-dimensional fronts are ordered by construction, which excludes it. Shock formation. Section 6 names two obstacles: the fixed Lax margin would bias weak shocks, and the exclusion band has a fixed width. Curved fronts. The cylindrical blast is a negative result against a floor declared before running, and we have not isolated its cause (Appendix D).

We present the paper as a diagnosis and a set of measured components on a test bed with exact answers, evaluated as rigorously as we could, not as a general Euler solver.

Reviewer_PHuNOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper addresses an important failure mode of PINNs when applied to nonlinear hyperbolic PDEs, which is the tendency to smear shocks. (This is a problem faced by most solvers, not just PINNs.) The paper proposes an explicit front-tracking ansatz, in which smooth neural network branches are separated by explicitly tracked discontinuities. The method is evaluated on several one-dimensional Riemann problems and a two-dimensional oblique shock, showing accuracy gains over strong-form and relaxation PINNs.

Critical strengths
  1. The paper addresses an important problem in computational science. The authors consider, in addition to the classical sod tube, nonconservative and 2-dimensional systems, which is unusual in the PINN literature.

  2. The authors report substantial accuracy gains over two baseline PINNs.

Critical weaknesses
  1. The biggest issue the paper has is that it seems to present a small advancement over the work of Lorin and Novruzi (2024, 2026), who proposed the idea of computing smooth local solutions in subdomains bounded by discontinuity lines, the latter defined from the Rankine-Hugoniot jump conditions.

  2. The paper seems to still be in a development stage. For example, it cannot deal with curved interfaces in 2D and does not deal well at all with rarefaction waves. The paper is dense and difficult to follow. I am familiar with the theory of nonlinear hyperbolic PDEs and it was difficult to follow for me, let alone someone from a general audience who does not have this background. The text seems to have been largely generated by AI and was not polished enough to make the presentation flow.

Questions for authors

This paper is promising but still seems to be in a development stage. There are a few questions that need to be addressed before it is ready, in my opinion. I list some of the questions below, not in order of importance, which could improve the paper and make it a substantial advance over the current literature on PINNs for nonlinear hyperbolic PDEs.

  1. Computational cost. The proposed method trains for ~1600 seconds (three seeds serially on one GPU) to reach an accuracy that a classical finite-volume solver reaches in a fraction of a second for Sod. Can the authors improve the efficiency of the method to make it a practical choice?

  2. There is a complex loss function with five terms, and the ablation shows that it is very sensitive to weighting choice. Can the authors develop a principled way to find the weights that does not require fine-tuning, which is needed in a new problem for which a solution is not known beforehand?

  3. The method applies to piecewise constant initial conditions and requires the structure of the solution to be known, exactly as traditional Riemann solvers. Can the method be extended to smooth initial conditions that develop shocks? This would offer a distinct advantage over traditional solvers.

  4. On curved interfaces, the level-set formulation is in principle capable of addressing them. What specifically prevents the method from handling curved fronts, is it the Newton projection, the anchoring strategy, or the training dynamics? What would a concrete roadmap to addressing this look like?

  5. A methodology to address rarefaction waves efficiently is needed, as most practical problems feature rarefaction waves.

2 comments

Thank you for the detailed review and the concrete questions. We try to answer each one in kind.

Firstly about soundness: The concerns raised are about novelty, maturity and presentation. We could not identify a claim that is said to be unsupported by its evidence. If there is one, we would be grateful to have it named.

Relation to Lorin and Novruzi (W1). We agree that neural shock fitting and its multi-dimensional extension are their work, and say so at the end of Section 1. Our additions are narrower and each is measured for it's strengths and weaknesses in this paper:

Wave structure derived from the data. With the family hard-coded for Sod, the mirrored problem is 96× worse than Sod; with it derived, 2.6× (Section 3.2).

Level-set fronts that fix no axis, where the 2026 paper uses a graph over a chosen axis.

Anchoring plateaus to the far-field data, a 6.09× median improvement on the colliding problem (Section 5.2).

Run selection without ground truth, which finds the best of three seeds in 11 of 13 settings (Section 5.3).

The prescribed-path study on a nonconservative system (Appendix K).

Maturity (W2, Q4, Q5). We agree on both limits and state them in the abstract.

Curved fronts. We have not isolated the cause. The projection and the planar jump defect were verified only on planar fronts (Appendix D), and the blast test leaves the contact and fan to one smooth branch. Any of the projection, the anchoring, the unrepresented waves or the optimisation could be responsible, and we cannot yet say which.

Rarefactions. With no front, the method is an ordinary PINN (Section 5.1), and relaxation networks are 1.8× more accurate on the two-rarefaction problem. The two are composable, but we have not tried it (Section 6).

Presentation. We accept that the paper is too dense, we tried to condense as much information in the page limit provided, results were moved to appendix where necessary. In the final revision we will move derivations out of the results and state each result before its caveats.

Cost (Q1). The gap on Sod is larger than the review states: about one second against 3546 s for three seeds (Table 3). The 1614 s figure is the colliding problem. We evaluated one training budget and did not trace a cost-accuracy curve (Appendix J), so we cannot claim the method is practical on cost.

Loss weights (Q2). The three groups are normalized automatically by their magnitude at initialization, and the same group weights are used on every Euler problem. No run was tuned separately, and the supplementary code reproduces this. The sensitivity is real: Appendix C shows per-term normalization is 11× worse on Sod, and Table 13 shows a tenfold Lax weight is also worse. For a problem with no known solution, the selection score of Section 3.5 is our partial answer. We have no principled rule beyond this.

Smooth initial data (Q3). Not addressed. Section 6 names two obstacles: the fixed Lax margin would bias weak shocks, and the exclusion band has a fixed width. Lorin and Novruzi (2026) already handle shock formation for scalar laws with graph fronts.

Thank you for your responses. My overall impression of the paper has not changed. I still think it needs more work, to make the method much more efficient, to be able to deal with rarefaction waves and curved fronts, to make it less sensitive to weighting of the loss components, and to be able to deal with smooth initial conditions.

Reviewer_gJHJOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The authors introduce an explicit front-tracking method to train a neural network as solution for Eulers equations to prevent the shock smearing typical of standard PINNs. They demonstrate accuracy gains on Riemann problems and successfully extends the approach to two-dimensional planar shocks. The evaluation compares against a classical solver, which is a refreshing change from many other PINN papers - but also demonstrates that the approach does not help to solve the problem compared to classical methods.

Critical strengths
  • The paper provides a scaling analysis explaining why the L2 residual objective in standard PINNs inherently leads to shock smearing. It is not a thorough analysis but gives insights into the issues.
  • The accuracy improvements over traditional strong-form and relaxation networks on Riemann problems (especially the colliding shock case) are present; the fact that a comparison was done against classical methods at all makes this paper more scientific than many other PINN papers.
  • The use of level sets and Newton projection provides a potential path for extending neural front tracking to higher dimensions.
Critical weaknesses

-The requirement of the exact star state to determine the wave structure means the method is currently a specialized tool for Riemann problems rather than a general PDE solver.

  • The method fails to track curved fronts, which indicates a limitation in the current level-set implementation or training.
  • The method is significantly slower than classical schemes except in very specific cases (highly discontinuous solutions), especially (as discussed by the authors) when considering that the classical solver is running on CPU, not GPU.
  • Very limited evaluation (not a lot of PDEs), and almost no theoretical analysis beyond the L2 residual.
Questions for authors
  • How can the method be adapted to determine the wave structure (number and type of fronts) for non-Riemann initial data (i.e. one that develops shocks over time)?
  • What specific modifications to the level-set anchoring or the loss function are needed to support curved fronts?
  • Since the "colliding shock" advantage over classical solvers depends on using a non-optimized NumPy implementation, can you provide a comparison against a compiled C++/Fortran solver? What benefits does the method have in comparison to using that one?
  • Could the authors elaborate on how the method would handle wave interactions (e.g., shocks merging), given that the current 1D implementation explicitly excludes this?
1 comment

Thank you for the review, and for noting the classical comparison (for some it strikes at the very heart of relevance of PINN based approaches). We respond in kind to each weaknesses pointed out and questions raised.

Firstly - Soundness. The weaknesses listed concern scope: Riemann data, curved fronts, cost and breadth. Each is stated in our abstract or in Section 6. We could not identify a claim in the review that is said to be unsupported by its evidence. If there is one, we would be grateful to have it named so that we can address it.

Star state (W1, Q1). The star state decides only which fronts exist and their families. Its values never enter the loss, and no reference solution enters training or the selection among seeds (Section 3.2, Figure 6). The choice among plateau variants did use the reference, as Section 5.3 states. We agree this is not a general solver. For shocks forming from smooth data, Section 6 names two obstacles we have not solved: the fixed Lax margin would bias weak shocks, and the exclusion band has a fixed width.

Curved fronts (W2, Q2). The cylindrical blast is reported as a negative result against a floor declared before running. We have not isolated whether the unrepresented contact and fan, the anchoring or the optimisation is responsible (Appendix D), so we cannot yet say which modification is needed.

Cost and a compiled baseline (W3, Q3). We agree. Section 5.5 states that the colliding advantage would not survive an optimised classical baseline, and Appendix J shows that removing our own solver’s strip overhead already reduces it from 5.0× to between 2.0× and 3.0×. The paper’s accuracy claim is relative to neural baselines at the same number of iterations: 8 to 14× on Sod and 386 to 879× on the colliding problem. We do not claim a cost advantage over any classical solver. Where we see a possible use is in imposing a prescribed jump condition. On a nonconservative model system (Appendix K), the tracked front follows the prescribed path with slope 0.998 to 1.004, where a Lax-Friedrichs-type path-conservative scheme gives −0.26 at 6400 cells. The path is given to the construction, and we do not compare with in-cell reconstruction schemes, which also capture it.

Wave interaction (Q4). The one-dimensional fronts are ordered by construction, which excludes interaction. Level sets are the natural route to merging, but we have not demonstrated it.

Breadth and theory (W4). The evaluation covers six one-dimensional Euler problems, six plateau variants on two of them, four neural baselines with three seeds on all three Riemann problems, eleven further formulations on Sod, oblique shocks at three angles, the cylindrical blast and the nonconservative system. We agree the theory is limited to the scaling argument, which we check numerically in Section 1.

Comments on the paper

1 comment

We thank the three reviewers. We summarize here what is common to the reviews, the individual responses give further details.

Points of agreement. All three reviews identify the same limitations: the construction needs Riemann data, does not handle wave interaction or shock formation, fails on a curved front, and costs far more than a classical solver whenever a rarefaction is present. We agree with every one, and do so state them in our paper from the very first version. Each one pointed out is stated in the abstract or in Section 6, and the cost comparison is our own Section 5.5.

Scope of the claims. The paper does not present a general Euler solver or a competitor to classical schemes. It claims a diagnosis of why strong-form PINNs smear shocks, and a set of components whose effect is measured on problems with exact solutions, against neural baselines at the same number of iterations.

Soundness. Two reviews rate soundness at 1. We could not find in either, a claim that is said to be unsupported by its evidence. The concerns listed are about scope, novelty and presentation. We have asked both reviewers to name any such claim, and we would address it directly.

Presentation. We accept that the paper is dense and there is informational overload in certain sections, we tried our best to preserve readability while adhering to the page limits imposed.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.