acceptodds
Under review as a conference paper at ICLR 2027

FLINT: A Simultaneous Emulator for Discontinuous and Smooth Structures with PINNs

Abstract

Physics-informed neural networks (PINNs) have made substantial progress on hyperbolic conservation laws containing sharp discontinuities. However, accurately capturing discontinuities while preserving transported fine-scale structure remains a challenge. We show that this failure worsens as the spatial frequency of initial data increases, exposing a core limitation of global space–time PINNs. We introduce FLINT (FLux balance for INitial-data Transport), which combines several recent advances in the PINN literature with our novel contributions: shock-selective residual weighting and variable-height control volumes anchored directly to initial conditions. The PDE residual is down-weighted at points where sharp gradients suggest either a shock or contact discontinuity, while the divergence theorem transforms local conservation into global conservation laws over control volumes, which propagate initial data into later times. We evaluate FLINT on 1D benchmark problems for compressible Euler equations, including the newly proposed Sod- diagnostic (a shock tube with tunable initial density oscillations), the Shu–Osher problem, and a 1D cylindrical blast wave. Across these benchmarks, FLINT captures wave amplitude and phase while maintaining shock and contact locations and global conservation, outperforming baselines in every metric. On Shu–Osher, FLINT reduces wave-region error by and wave amplitude error by relative to the stronger baseline, while accurately handling geometric source terms and coordinate singularities in the blast wave. By addressing these 1D transport challenges, FLINT establishes a necessary foundation for scaling global space–time PINNs to complex multi-dimensional compressible flows.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Physics-Informed Neural Networks, Scientific Machine Learning, Hyperbolic Conservation Laws, Neural PDE Solvers

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_4aJqOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes the "FLux balance for INitial-data Transport" (FLINT) concept to address the failure of global space-time PINNs in preserving high-frequency initial data during transport. Key contributions are a weighting mechanism that suppresses the strong-form PDE residual specifically at compressive shocks while maintaining it elsewhere, and evaluation/validation showing superior performance in wave amplitude, phase, and global conservation compared to other neural network approaches.

Critical strengths
  • The use of anchored control volumes provides a direct non-local link between initial conditions and later predictions, effectively overcoming the "smoothing" effect often seen in PINNs.
  • The introduction of the benchmark and more problems provide a good test of transport and wave-capturing capabilities.
  • The inclusion of proposition (B.1) combined with the linear-advection experiments provide a clear theoretical explanation for why small PDE residuals do not necessarily guarantee accurate wave recovery.
Critical weaknesses
  • The contribution seems to be incremental; a loss term designed for very specific shock scenarios. The experiments do not help alleviate this assessment, because they are done on very simple toy problems with no comparison to the state of the art (traditional solvers for these csaes). Claims about potential higher dimensionality are not backed up.
  • No theoretical analysis or substantial contribution.
  • The authors acknowledge that the computational expense remains high compared to traditional finite-volume solvers, although they argue for the potential of PINNs in higher dimensions. It is not clear which problems they mean in particular, as high-dimensional PDEs with shocks are not particularly common.
  • It is not clear how much the presented approach depends on the network architecture and how much the results are attributable to the choice of loss funciton (the new contribution). Capturing shocks with a continuous (even smooth) network is not ideal, and integrating a discontinuity as an output of the network may help more - akin to traditional Riemann solvers.
  • Some hyperparameters are hand-tuned per-problem, which makes it hard to assess the immediate generalizability of the framework in practice.
Questions for authors
  • Could the authors discuss if the anchored control volume approach can be extended to problems where the initial state is not known exactly (e.g., inverse problems or noisy data)?
  • Regarding the hand-tuned hyperparameters, do the authors have suggestions for a more automated way to determine parameters (like gate threshold) across different physics problems?
  • How does the method compare against classical solvers for these problems, especially classical shock tracking algorithms?
3 comments
AuthorsOpenReview ↗

Author Response to Reviewer 4aJq Part 1

We thank 4aJq for their thoughtful review. We will address each of your comments and questions here.

W1. Incremental, shock-specific contribution, tested only on toy problems without classical solver baselines, and with unsupported higher-dimensional claims. First we thank the reviewer for suggesting we pursue a higher-dimensional problem, as the extension of FLINT to higher spatial dimensions is straightforward, we now provide an extension to a 2D problem in Section 4.4 lines [475-506] with additional information including in Appendix A.5 lines [1176-1225]. The key result of extending FLINT to two dimensions is that it outperforms the baselines in 14 of the 16 reported metrics, with only marginal differences in the remaining two.

Regarding novelty of contribution, we direct the reviewer to our response to W1 from reviewer iU9n.

Regarding the toy nature of the test problems and comparison to state of the art: we note that we do compare to industry standard high-resolution shock capturing methods including a finite volumes scheme with both a limited linear and WENO5 reconstruction. In hypersonic flows, unresolved shocks, often interacting with complex wave structure, are nearly ubiquitous. It is for this reason that most traditional solvers for hypersonic systems include the Shu-Osher test in their verification suites. So while we have not modeled the nearly unlimited set of flow structures, we believe we have chosen a relevant representative subset.

W2. Additional theoretical analysis. We would like to thank the reviewer for acknowledging the inclusion of Prop B.1 with stating that proposition (B.1) combined with the linear-advection experiments provide a clear theoretical explanation for why small PDE residuals do not necessarily guarantee accurate wave recovery.

FLINT is a neural network based optimization method for PDE problems. We believe that our method is evaluated thoroughly and our claims are supported empirically.

In regards to additional theoretical analysis, we would appreciate guidance from the reviewer as to which parts (and/or components) lacks theoretical support.

W3. FLINT is costlier than finite-volume solvers, and the higher-dimensional benefit names no concrete problem with shocks. We thank the reviewer for suggesting a more careful analysis of cost. We now include a full study and ablation in the revised manuscript, Appendix D lines [1449-1457,1486-1499], Figure 18 lines [1535-1559] and Table 9 lines [1512-1533]. To convey matched computational budgets we increased the number of iterations to meet FLINT where necessary and we also matched FLINTs run times. The key result is the improvements persist under these matched budgets, since FLINT is both the fastest and the most accurate method on Sod-k, and on Shu Osher the baselines do not reach FLINT's accuracy even when they are given more training time, and they still lose the entropy wave that FLINT resolves. The trade-off for Shu-Osher is accuracy over a longer training time and this is something we will continue to investigate in future work.

Regarding the presence of higher-dimensional PDEs with or without shocks: we primarily mean three-dimensional fluid flows coupled to kinetic/transport systems, which are descried by a six-dimensional Boltzmann equation. For transport of, e.g., chemical particulates or ion species, these flows are particularly important for engineering applications such as semiconductor manufacturing and hypersonic vehicles (Anderson, 2006; Candler, 2019). If the transported field is photons, its coupling to the gas produces a rich phenomenology of radiation shocks (Mihalas and Weibel-Mihalas, 1984; Lowrie and Edwards, 2008), with applications that include laser fusion and reentry vehicles. We have revised the outlook of the manuscript to include these examples.

W4. The effects of the architecture and the loss are not separated, and a network that outputs the discontinuity directly may capture shocks better. We thank the reviewer for this helpful question and agree that the role of the network should be separated from that of the loss. We have added a network size ablation in Appendix E.2 lines [1598-1610] and Table 10 lines [1620-1645] of the revised manuscript. The key result is that the network capacity does have an impact on the overall optimization process. Lowering the network capacity with all other settings fixed, is unable to capture the wave structure in both Sod-kk and Shu-Osher. Furthermore, increasing the capacity of the network does not necessarily lead to better performance.

AuthorsOpenReview ↗

Author Response to Reviewer 4aJq Part 2

Q1. Noisy data.

Thank you for this thoughtful question. We now include an extended robustness experiment by adding noise to the initial conditions of Shu-Osher. This extended experiment is in the revised manuscript under Appendix E.3 lines [1612-1619] and Table 11 lines [1647-1662]. The key result is that this experiment suggests that FLINT is robust to noise. To test this, we trained FLINT with the reported configuration for Shu-Osher, with no changes to the method, but initial data corrupted by random noise in all three variables, and scored it against the WENO5 reference. The noisy initial data are

\tilde{u}(x) = u_0(x) + \sigma\, c_0(x)\, n_u(x), \quad \tilde{p}(x) = p_0(x)\,\bigl(1 + \sigma\, n_p(x)\bigr),$$ where $\rho_0$, $u_0$ and $p_0$ are the exact initial density, velocity and pressure, $\sigma$ is the noise level, and $c_0 = \sqrt{\gamma p_0 / \rho_0}$ is the local sound speed with $\gamma = 1.4$, which sets the size of the velocity noise. The noise $n_\rho$, $n_u$ and $n_p$ is drawn from a standard normal distribution, $\mathcal{N}(0, 1)$, independently for each variable. **W5 + Q2. Hyper-parameter selection.** Thank you for this important question. To our knowledge, the baselines we compare against do not give a procedure for choosing their settings on a new problem. PINNs-WE reports its settings for each problem without saying how they were chosen. WE-PINN fixes its loss weights to 1 and chooses its network and quadrature from a sweep scored against the exact Burgers solution. At submission we used the same approach and chose our settings for each problem against the WENO5 reference. We acknowledge that this is not practical for a new problem, where no reference exists. **Per the reviewer's suggestion, we have explored a more practical approach. We have now developed a set of rules, as advice for future practitioners that do not depend on the reference solution, which we include in the revised Appendix H lines [1818-1889] Our original prescription and experiments remain unchanged. Additionally, we found a single configuration that works across all of the problems presented. We found that taking the Shu-Osher configuration and using the rules described in the revised Appendix H provides results very close to the results reported in the paper.** **Q3. Comparison with classical solvers and shock tracking.** We thank the reviewer for this important question and agree that classical solvers are the right standard for these problems. We used **WENO5 as the classical reference** for Sod-$k$ and Shu-Oshe. On these one dimensional problems classical solvers are much faster than PINNs, and WENO5 with 800 cells reaches errors comparable to those of FLINT in about one second on one CPU core. We agree that shock tracking and shock fitting schemes (Glimm et al., 1981; Chern et al., 1986; Moretti, 1987; Salas, 2010) are an appealing way to treat these flows. **To address this question we ran the shock tracking algorithm of Glimm et al., 1981 and Chern et al., 1986. and the results are in the revised manuscript in Appendix I lines [1893-1991].** **The key result is that the shock-tracking solutions converge toward the reference solutions used throughout this paper. This provides additional confidence in our evaluation of FLINT and the baselines, suggesting that our conclusions are not strongly dependent on the choice of reference solver.** **While FLINT does not outperform traditional solvers in one dimension, it does outperform the SOTA in the PINN literature, and is a step towards the practical application of PINNs in this space.** We find that our loss straightforwardly extends to two dimensions, where it has the lowest error among the PINN methods we tested on the quadrant Riemann problem **see reviewer iU9n W3**, and we hope it helps future PINNs handle these scenarios.
AuthorsOpenReview ↗

Follow-Up to Author Responses and Revisions

Dear reviewer 4aJq: Thank you again for your time and constructive feedback. We have provided point-by-point responses and additional experimental results, and hope these address your concerns.

As the discussion deadline approaches, we would appreciate it if you could review our replies when your schedule permits. Please let us know if anything remains unclear or requires further clarification. We are more than delighted to have further discussions and improve our manuscript. If our response has addressed your concerns, we would be grateful if you could re-evaluate our work.

Reviewer_iU9nOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces FLINT, a global space–time physics-informed neural network framework for solving compressible Euler equations while capturing both discontinuities and fine-scale smooth waves. Its main contributions are variable-height control volumes anchored to the prescribed initial data, which directly constrain later predictions through integral flux balance, and a residual-weighting scheme combining solution steepness with local compression to selectively reduce strong-form enforcement near shocks. These components are combined with interior control-volume constraints, positivity-preserving outputs, and SOAP optimization.

Validation covers a proposed Sod-k diagnostic with tunable initial density oscillations, the Shu–Osher shock–entropy-wave interaction, and a cylindrical blast wave solved in one-dimensional radial coordinates. Comparisons with PINNs-WE and WE-PINN use high-resolution numerical references and report solution errors, wave-amplitude errors, and conservation diagnostics across multiple seeds. On Shu–Osher, FLINT reports approximately 4–6-fold reductions in wave-region errors and 7–31-fold reductions in wave-amplitude errors relative to the stronger baseline. Additional studies examine component contributions and optimizer effects. The blast-wave experiment demonstrates handling of geometric source terms, but has no baseline comparison; validation remains restricted to one spatial dimension.

Critical strengths
  1. The paper addresses an important failure mode beyond shock localization. Capturing discontinuities while preserving nearby fine-scale waves is a meaningful challenge for neural conservation-law solvers. The Sod-k diagnostic and Shu–Osher benchmark directly test this capability, making the evaluation more informative than shock-tube tests dominated by piecewise smooth states.

  2. The reported improvements in wave preservation are substantial. Across three seeds on Shu–Osher, FLINT reduces wave-region errors by approximately 4–6 times and wave-amplitude errors by 7–31 times relative to the stronger evaluated baseline, alongside improved conservation diagnostics. These results provide substantive evidence of practical benefit on the tested problems, even though the methodological novelty is incremental.

Critical weaknesses
  1. The conceptual advance over existing control-volume PINNs is limited. The principal contribution anchors randomly sampled, variable-height control volumes to the prescribed initial data, while the residual weighting combines existing gradient and compression indicators. These are reasonable extensions, but they do not introduce a fundamentally new conservation formulation. The empirical gains are useful, yet the paper does not sufficiently establish a broader methodological insight beyond this particular training design, limiting the significance of the contribution.

  2. The validation is too narrow to establish broader applicability. All experiments involve one spatial dimension, including the cylindrical blast wave, which is solved in radial coordinates and has no baseline comparison. Quantitative comparisons are restricted to two neural baselines, and the method requires problem-specific tuning and substantial quadrature sampling. Consequently, the experiments demonstrate improved performance on selected benchmarks but leave uncertain whether the gains persist across more varied flows and under comparable computational budgets. This limits the evidence supporting the proposed framework’s general utility.

Questions for authors
  1. Clarification on the contribution of initial-data anchoring. With the architecture, SOAP optimizer, residual weighting, and total quadrature budget held fixed, how does FLINT compare with a version that replaces the anchored volumes with additional interior control volumes? Does anchoring provide gains beyond those attributable to increased sampling or computation?

  2. Clarification on computational fairness. What are the total training time, peak memory, and network-evaluation budgets for FLINT and the baselines? Do the improvements persist under matched computational budgets, including the SOAP-trained baseline comparisons?

  3. Clarification on quadrature sensitivity. How sensitive are the wave-preservation results to the number of quadrature nodes and the widths of the anchored volumes? In particular, do the reported gains persist as the initial oscillation frequency increases when the quadrature budget is fixed?

3 comments
AuthorsOpenReview ↗

Author Response to Reviewer iU9n Part 1

We thank reviewer iU9n for their thoughtful review and questions towards clarification of our method and results.

W1. Do not introduce a fundamentally new conservation formulation and he paper does not sufficiently establish a broader methodological insight

We agree with the reviewer that some components of FLINT may have been introduced in previous methods. However, the particular formulation of FLINT builds on (and advances) several insights in the PINN literature. In the paper, our primary novel intellectual contributions are:

  • We identify a weakness of current state-of-the-art PINNS, which is that they do not capture fine wave structure and shocks simultaneously.

  • Inspired by classical solvers, we address this weakness in FLINT by injecting initial conditions more directly into the spacetime domain through control-volumes anchored at t=0t=0, adhering to conservation laws. The weighted residuals preserve the shock.

  • In addition to existing test problems with shocks (stressful Shu-Osher test and the cylindrical blast wave), we evalute FLINT and SOTA baslines against our Sod-kk initial conditions which further stress a methods's ability to capture fine wave structure in a controlled way. FLINT consistently outperforms other baselines.

W2a. All experiments involve one spatial dimension While we followed the majority of PINN literature, which focuses on primarily 1D problems, we thank the reviewer for suggesting the inclusion of a test problem with two spatial dimensions, as the extension of FLINT to this domain is straightforward. We now include an additional 2D benchmark, discussed below and included in the revised manuscript in Section 4.4 lines [475-506] with additional information including in Appendix A.5 lines [1176-1225]. The key result of extending FLINT to two dimensions is that it outperforms the baselines in 14 of the 16 reported metrics, with only marginal differences in the remaining two.

W2b. Two neural baselines. The two neural baselines used were the only ones available for these problems in the literature that had an end-to-end solution. We would be grateful for any additional baselines the reviewer recommends. For the cylindrical blast wave we could not compare with neural baselines because it would require extensive modifications to our baselines code, but we do compare to the classical baseline MUSCL–HLLC (minmod-limited MUSCL reconstruction (van Leer, 1979; Sweby, 1984) with the HLLC approximate Riemann solver (Toro et al., 1994; Toro, 2009) and SSP-RK3 time stepping (Gottlieb et al., 2001).

W2c. Hyper-parameter Tuning and Quadrature
Thank you for this important question. To our knowledge, the baselines we compare against do not give a procedure for choosing their settings on a new problem. PINNs-WE reports its settings for each problem without saying how they were chosen. WE-PINN fixes its loss weights to 1 and chooses its network and quadrature from a sweep scored against the exact Burgers solution. At submission we used the same approach and chose our settings for each problem against the WENO5 reference. We acknowledge that this is not practical for a new problem, where no reference exists. Per the reviewer's suggestion, we have explored a more practical approach. We have now developed a set of rules, as advice for future practitioners that do not depend on the reference solution, which we include in the revised Appendix H lines [1818-1889] Our original prescription and experiments remain unchanged. Additionally, we found a single configuration that works across all of the problems presented. We found that taking the Shu-Osher configuration and using the rules described in the revised Appendix H provides results very close to the results reported in the paper.

For the question around quadrature we answer this in Part 2: Q3 below.

W2d. Varied flows and matched computational budgets. We thank the reviewer for these important questions. The question on matched computational budgets is answered below in Part 2: Q2.

In regards to varied flows, this work is focused on the class of PDEs with hyperbolic conservation laws. While the space of all possible flows within this class is uncountably large, we have shown that FLINT can work on a substantial set of "primitive" flows that compose to more complex phenomena. This includes shocks from Mach 1.71.7 to Mach 33, entropy waves at four frequencies from k=5k = 5 to k=100k = 100, a shock interacting with a wave train, a cylindrical blast wave with geometric source terms in radial coordinates, the plain Sod tube and the Burgers rarefaction, and in this rebuttal a 2D quadrant piecewise initial condition evolving into slip lines and a vortical roll up at the center presented in Section 4.4 lines [475-506] with additional information including in Appendix A.5 lines [1176-1225].

AuthorsOpenReview ↗

Author Response to iU9n Part 2

Q1. Ablation on initial-data anchoring. We thank the reviewer for this suggestion. We provide the ablation as an update in Appendix C.1, Table 8, lines [1404-1431]. The key result is that the gain is anchoring itself. With everything fixed, replacing the anchored volumes by twice as many interior volumes leaves the error unchanged on Sod-kk and Shu-Osher, whereas anchoring lowers it six- and eight-times (to 0.015 and 0.036) and recovers the wave behind the shock.

Q2. Comparisons of computational cost. We agree with the reviewer that this metric should be reported and we thank them for the suggestion. We now include a full study and ablation in the revised manuscript, Appendix D lines [1449-1457,1486-1499], Figure 18 lines [1535-1559] and Table 9 lines [1512-1533]. To convey matched computational budgets we increased the number of iterations to meet FLINT where necessary and we also matched FLINTs run times. The key result is the improvements persist under these matched budgets, since FLINT is both the fastest and the most accurate method on Sod-k, and on Shu Osher the baselines do not reach FLINT's accuracy even when they are given more training time, and they still lose the entropy wave that FLINT resolves. The trade-off for Shu-Osher is accuracy over a longer training time and this is something we will continue to investigate in future work.

Q3. Clarification on quadrature sensitivity. We thank the reviewer for this important question. The results do not depend on the number of quadrature nodes once a face has enough of them. The number of nodes NN on a face introduces an error in the evaluation of the residual that is proportional to 1N2\frac{1}{N^2}, so NN can be chosen from the error we want to achieve, while making sure this error does not dominate the error in the interior and anchored control-volume residual. We note that the spacing between quadrature nodes must also be sufficiently fine to capture the spatial variation in, e.g., the entropy wave. i.e., L/N<L/kL/N \lt L/k for domain width LL.

The experiments agree with this. The key result is that FLINT is not overly sensitive to the number of nodes NN, for N≥121.N\geq 121. The experiment suggest that we could successfully lower the number of quadrature nodes and is something we will continue to explore in future work. The experiment is Appendix E.1 lines [1503-1511,1562-1565], Figure 19 lines [1566-1579] and Figure 20 lines [1585-1596] in the revised manuscript.

AuthorsOpenReview ↗

Follow-Up to Author Responses and Revisions

Dear reviewer iU9n: Thank you again for your time and constructive feedback. We have provided point-by-point responses and additional experimental results, and hope these address your concerns.

As the discussion deadline approaches, we would appreciate it if you could review our replies when your schedule permits. Please let us know if anything remains unclear or requires further clarification. We are more than delighted to have further discussions and improve our manuscript. If our response has addressed your concerns, we would be grateful if you could re-evaluate our work.

Reviewer_8DYvOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

FLINT is a global space–time PINN for hyperbolic conservation laws that aims to preserve transported fine-scale structure alongside shocks and contacts. It combines compression-gated residual weighting with interior flux-balance constraints and variable-height control volumes whose lower faces use the prescribed initial state directly. Experiments cover frequency-controlled Sod-k, Shu–Osher, Burgers, and a radial cylindrical blast wave. Initial-face ablations and SOAP-based baseline comparisons support the benefits of the approach, particularly for oscillatory wave reconstruction.

Critical strengths

The strongest contribution is the direct coupling of known initial data to later predictions through integral conservation. The experiments provide targeted evidence that this helps preserve transported wave structure.

  1. The initial-face ablation holds control-volume geometry fixed while replacing the prescribed initial state with the network prediction. Its results support the benefit of injecting exact initial data into conservation constraints, beyond merely placing boxes at t = 0 (Appendix C.1; Table 7).
  2. On Shu–Osher, FLINT substantially improves both wave-region error and amplitude fidelity, with reported reductions of roughly 4–6× and 7–31× relative to the stronger default baseline. These metrics directly address the paper's stated problem of retaining oscillations near discontinuities (Table 2).
  3. The evaluation includes a controlled frequency sweep, multiple training seeds, and baseline retraining with SOAP for the same number of optimization steps. The latter comparison shows that optimizer choice alone does not explain the observed accuracy gains (Appendices A.1 and C.2).
Critical weaknesses

The empirical results are encouraging, but the contribution of the residual gate, the practical tuning requirements, and the scope of physical reliability remain incompletely established.

  1. The compression gate is a central contribution, but its independent effect is not isolated by an end-to-end ablation within the same FLINT architecture and training setup. Figure 2 visualizes gated and ungated weights on a trained solution; this does not establish the accuracy gain from training with the gate rather than gradient-only weighting.
  2. Appendix F selects quadrature nodes against a reference, box counts using solution performance, and anchored-loss weights separately for each problem. Robustness without reference-dependent tuning is therefore unclear. Matching optimization steps also does not match computation when FLINT uses hundreds of anchored boxes with 385 quadrature nodes per face; wall-clock and network-evaluation costs are not reported.
  3. Physical reliability is demonstrated empirically rather than guaranteed: the Burgers rarefaction test is useful, but does not establish entropy admissibility for the Euler cases. Claims of superiority in every metric also need qualification. For Sod-k, the energy conservation residual is 3.21 × 10^-3 for FLINT versus 1.80 × 10^-3 for WE-PINN (Table 1).
Questions for authors
  1. Can you add a controlled ablation of compression-gated versus gradient-only weighting with the same network, optimizer, and sampling budget? Ideally, test it both with and without initial-data anchoring on Sod-k and Shu–Osher to separate the contributions of the two proposed components.

  2. What reference information was used to select the anchored-box count and loss weight, and how well does a common configuration transfer without per-problem reference-based selection? Please report wall-clock training cost or total network evaluations alongside the equal-step comparison to clarify the accuracy–cost trade-off.

2 comments
AuthorsOpenReview ↗

Author Response to Reviewer 8DYn

We thank reviewer 8DYv for their insightful comments, questions, and patience. We will address the comments and questions below:

W3. Entropy Admissibility. Thank you for this very helpful comment. We now test entropy admissibility on the Euler cases directly, by comparing the entropy each solution produces with the exact amount the shock must generate under the Rankine–Hugoniot conditions, and we will include this test in the revised manuscript. The key result is that FLINT produces entropy on Sod-kk and Shu-Osher. This evidence suggests that entropy admissibility is met in these cases even though it is not enforced explicitly. Like traditional finite-volume solvers, FLINT works with the global conservation of the conserved quantities and with the weak form over control volumes, and we believe this is why entropy is captured implicitly. This experiment is Appendix A.2 lines [1025-1079].

Thank you as well for pointing out the metric claims. The referee is correct, FLINT is best or tied for best on 75 of the 79 metric entries that have against a baseline, and we have updated the abstract in the revised manuscript to reflect this.

W1 + Q1. Controlled ablation of gate vs. anchored boxes. Yes, we provide the ablation study below. The key result is that both components are necessary and complementary. Anchoring to the initial data lowers the error six- to eight-times over every un-anchored variant, and the compression gate then keeps the wave behind the shock with 4$\times$ the accuracy at the same network, optimizer and sampling budget.

This experiment is Appendix C.2 lines [1433-1446] and Figure 17 lines [1458-1484] in the revised manuscript.

W2 + Q2. Hyper-parameter selection and generalization. Thank you for this important question. To our knowledge, the baselines we compare against do not give a procedure for choosing their settings on a new problem. PINNs-WE reports its settings for each problem without saying how they were chosen. WE-PINN fixes its loss weights to 1 and chooses its network and quadrature from a sweep scored against the exact Burgers solution. At submission we used the same approach and chose our settings for each problem against the WENO5 reference. We acknowledge that this is not practical for a new problem, where no reference exists. Per the reviewer's suggestion, we have explored a more practical approach. We have now developed a set of rules, as advice for future practitioners that do not depend on the reference solution, which we include in the revised Appendix H lines [1818-1889] Our original prescription and experiments remain unchanged. Additionally, we found a single configuration that works across all of the problems presented. We found that taking the Shu-Osher configuration and using the rules described in the revised Appendix H provides results very close to the results reported in the paper.

AuthorsOpenReview ↗

Follow-Up to Author Responses and Revisions

Dear reviewer 8DYv: Thank you again for your time and constructive feedback. We have provided point-by-point responses and additional experimental results, and hope these address your concerns.

As the discussion deadline approaches, we would appreciate it if you could review our replies when your schedule permits. Please let us know if anything remains unclear or requires further clarification. We are more than delighted to have further discussions and improve our manuscript. If our response has addressed your concerns, we would be grateful if you could re-evaluate our work.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Contributions and Additional Experimental Results

Dear Area Chair and reviewers,

Thank you for your time and effort in handling our manuscript. The main contribution of FLINT is identifying a failure mode in the learned representation of PINNs applied to hyperbolic conservation laws and providing a solution. We show that PINNs can achieve small PDE residuals while failing to preserve fine-scale wave structures from initial conditions. FLINT helps this issue through initial-data-anchored conservation constraints and discontinuity selective residual weighting. Our goal is to improve the effectiveness of PINNs, and note that while FLINT was developed with flows like Shu-Osher in mind, it also improved performance on the cylindrical blast wave, 2D Quadrant Riemann, and the more fundamental problems like the Sod shock tube and Burgers inviscid equation. In response to the reviewers, we have worked through several new experiments, with five key results summarized below.

  • Component Ablation. We ran an ablation that isolates the effects of the discontinuity-selective residual weight and the initial-data-anchored control volumes, the key result is both components are necessary for FLINT's reported performance (Appendices B and C).

  • General Guidelines for Hyper-parameters. We provide practical hyperparameter guidelines and also discuss a shared configuration that works across the tested problems (Appendix H).

  • Computational Fairness and Sensitivity. FLINT maintains better accuracy under matched optimization steps, wall-clock budgets, and SOAP optimization, although Shu--Osher requires more computational resources. We also investigate the impact of quadrature resolution and network capacity (Appendices D and E), the key result is FLINT is not sensitive to quadrature up to a certain NN and the network capacity also has an impact on the ability to learn wave-structure..

  • Extension to 2D. We extend FLINT to 2D. FLINT captures the four slip lines and central roll-up of the 2D quadrant Riemann problem. It achieves the lowest errors in 14 of 16 metrics, with marginal differences in the remaining two (Section 4.4 and Appendix A.5).

  • Miscellaneous Experiments FLINT produces positive net entropy, agrees closely with classical shock-tracking references, and maintains similar accuracy with 5% noise in the initial conditions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.