acceptodds
Under review as a conference paper at ICLR 2027

Learning Nonlinear Transient Dynamics in Fluid Flow Networks with the Neural Basis Method

Abstract

Fluid flow networks are ubiquitous across engineered and natural systems, including pipeline infrastructure and biological vasculature. In many practical regimes, they are governed by nonlinear edgewise hyperbolic laws coupled through continuity and conservation at junctions. Learning their dynamics is important in many-query scenarios requiring rapid evaluation across unseen parametric settings. Conventional methods re-solve the system for each new query, resulting in prohibitive cost. Existing physics-informed neural approaches typically impose governing equations through penalty losses over generic neural approximators, rather than embedding the physical laws in their neural formulation. Such losses act only as optimization surrogates during training, with no guarantee that their reduction reduces solution error. Moreover, the resulting optimization problem is prone to ill-conditioning, particularly for hyperbolic systems. Building on the “representation-then-projection” principle of the neural basis method, we introduce a new approach for modeling fluid flow networks. Fixed neural bases define the representation space on each edge, while projection of the discretized equations determines the expansion coefficients, embedding edgewise laws into a conservative formulation. Edge subdomains are coupled through non-overlapping domain-decomposition. A neural operator then learns, directly in this reduced space, the parametric map to the corresponding expansion coefficients. On two representative network topologies, looped and branching, we reproduce references with relative errors below , whereas the physics-informed neural network baseline fails. Across unseen boundary conditions in both interpolation and extrapolation, we maintain mean pressure errors below and mean flux errors below , while the physics-informed DeepONet baseline reaches – pressure errors and flux errors of –, failing to capture the underlying dynamics. The results demonstrate the accuracy and efficiency of the proposed framework for learning nonlinear transient dynamics in fluid flow networks, highlighting its potential for real-time monitoring, design exploration, and uncertainty quantification.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
fluid flow network, nonlinear transient dynamics, operator learning, physics-informed machine learning, neural basis method

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_QdiCOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

Traditional methods and standard physics-informed neural networks for modeling fluid flow networks are often prohibitively expensive or suffer from ill-conditioning due to their reliance on penalty losses. To address this, the authors propose a novel framework based on the "representation-then-projection" principle, which structurally embeds physical laws directly into the neural formulation using fixed neural bases and non-overlapping domain-decomposition.

Critical strengths
  • The paper addresses the uniquely challenging and novel problem of rapidly and accurately evaluating nonlinear edgewise hyperbolic laws coupled across large-scale fluid flow networks, which has direct applications to critical systems like biological vasculature and pipeline infrastructure.

  • The proposed framework demonstrates remarkable accuracy.

Critical weaknesses
  • The paper's central motivation regarding the optimization problem being "prone to ill-conditioning, particularly for hyperbolic systems" is not discussed or justified anywhere in the text.

  • Dataset visualization is missing.

  • It is unclear how the datasets were generated (by the authors or from existing works), how many trajectories and frames are included, and how the datasets were split.

  • The authors only compare their method with two very old baselines. This comparison is insufficient to determine how the proposed model performs compared to recent state-of-the-art approaches.

  • The highlighted strong performance in unseen boundary conditions for both interpolation and extrapolation is not justified by any experiment, and there are no boundary errors or boundary residuals provided.

  • Experiments and experimental settings are insufficient, as evaluations are only conducted on two datasets with the same geometry. Model performance remains unclear regarding long rollout stability, scalability to large domains, and generalization to different geometries.

  • Many variables are undefined, making the paper difficult to read. For example, in equation 4.2, Φ\Phi is not defined, and the construction of AA and bb is unexplained.

  • The proposed method shows an orders-of-magnitude improvement over the baselines, requiring an analysis of whether the baselines are simply too weak or poorly trained, as the pressure and momentum evolution in Figure 3 shows the baselines completely diverging from the ground truth.

  • The authors' claim that "Such losses act only as optimization surrogates during training, with no guarantee that their reduction reduces solution error" disagrees with many findings from existing PINNs papers and requires justification.

Questions for authors

See weakness.

1 comment
AuthorsOpenReview ↗

Response to Reviewer (Revised PDF & Additional Experiments)

We thank the reviewer for the feedback. Most of the concerns raised are already addressed in the manuscript, and we have uploaded a revised PDF with additional experiments (highlighted) to resolve the remaining points:

  • Already covered in the paper: We evaluate two distinct topologies (a looped network and a branching tree), directly test unseen Dirichlet inlet pressures (interpolation and extrapolation) in the held-out tables, and use non-autoregressive inference (avoiding rollout drift). Dataset generation (staggered finite-volume solver) and splits are detailed in the experimental setup.
  • New POD-NN baseline: In the new Additional Experiments appendix, we added a matched supervised POD-NN baseline (same trajectories, POD basis, architecture, and 2000-epoch budget). Both reduced-order methods outperform PI-DeepONet by 1–2 orders of magnitude, and NBM-OL beats POD-NN by 1.8×1.8\times–3.0×3.0\times in interpolated pressure while matching it in flux.
  • Baseline convergence & ill-conditioning: Our new convergence analysis confirms the trained PINN escapes the trivial basin (8×8\times lower loss) and plateaus by 20k epochs, yet still incurs 92%92\% flux error—empirically illustrating PINN ill-conditioning in hyperbolic regimes and why low surrogate loss does not guarantee accurate solution recovery.
  • Residuals, visualizations, and notation: The new appendix reports explicit boundary (1.7×10−141.7 \times 10^{-14}) and junction (2.4×10−82.4 \times 10^{-8}) residuals, and adds network heatmaps and spatial/temporal profiles at CFL=0.7\mathrm{CFL} = 0.7. Regarding Eq. 4.2, Φ\Phi is defined in the preceding paragraph (including its block structure), WW in the sentence immediately following it, and An,bnA^n, b^n are introduced with the finite-volume discretization and assembled row-by-row in Appendices A–B (please let us know if other symbols were intended).

We hope these clarifications and additional experiments address your concerns.

Reviewer_EPwqOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes a surrogate for transient gas flow in pipeline networks. A fixed, untrained random network supplies a basis on each pipe, and a finite-volume scheme is projected onto it and solved by least squares at each time step. The resulting coefficient trajectories are compressed with POD, and an MLP is trained on the projected FV residual to map time and inlet pressure to the latent state. Results on two small networks show close agreement with the reference solver and large gains over PINN and PI-DeepONet baselines.

Critical strengths
  1. Within the tested setting, the learned model reproduces the full transient (wave reflections, junction behavior, steady state) accurately and holds up under mild extrapolation.

  2. The formulation follows an established FV scheme closely, with junction conditions built into the solve rather than added as penalties. The fixed basis and POD give a substantial reduction in problem size. Since the system is solved by least squares, though, conservation holds only approximately.

Critical weaknesses
  1. NBM-OL is trained on solver trajectories and evaluated against the same solver, while PI-DeepONet and the PINN do not seem to get any data and must solve the full continuous problem. The baselines also appear to sit in the trivial-flow failure mode the authors describe in the appendix. There are no baselines trained on the same data and no ablation of the residual loss against a simple supervised loss, so the contribution of the method itself is unclear.

  2. Eq. 4.6 bounds the distance to the best one-step solution in the POD subspace, not the error against FV or the true solution. It follows from linear least squares with reference states, so the contrast with PINN losses isn't fair. It also can't be evaluated at unseen inlet pressures, and its constant is never reported.

  3. There is no justification or ablation for the random basis or its size. The reference scheme runs at a large time step and is strongly dissipative, so close agreement may reflect a smooth target. There is no convergence study and no test with sharper fronts.

  4. The method is tested on one fixed network with a single scalar parameter, a dense training grid and small extrapolation. Despite the efficiency motivation, no computational cost is reported, and alternatives such as adjoint-based methods aren't discussed. The "any queried time" claim is only tested on the training time grid.

Questions for authors
  1. How does accuracy depend on the basis size, and how is it chosen?

  2. How does the method behave at small CFL, where fronts are resolved?

  3. Can errors be reported against FV or a converged solution rather than NBM?

  4. What are the wall-clock costs (FV, NBM, training, inference), and when does the method pay off? How does it compare with an adjoint-based approach?

  5. How does NBM-OL compare with supervised baselines on the same data, and with its own pipeline trained on a supervised loss?

  6. Is the model accurate between training time steps?

3 comments

We thank the reviewer for the thorough and constructive feedback. We have uploaded a revised PDF with a new Additional Experiments appendix (highlighted) that directly addresses your questions and concerns:

  • Basis size ablation (Q1): We added a sweep over k∈{10,20,40,80}k \in \{10, 20, 40, 80\} on the tree network (measured directly against the FV reference). Error decreases monotonically with kk, and at k=40k = 40 the representation error (9.9×10−69.9 \times 10^{-6} in pressure, 1.3×10−41.3 \times 10^{-4} in flux) is already two orders of magnitude below the learned map's error, whereas k=80k = 80 increases solve time by 2.7×2.7\times (207207,s vs. 550550,s). Thus k=40k = 40 provides a practical balance of accuracy and speed.
  • Sharper fronts at small CFL (Q2, Q3): We repeated the NBM--FV verification at Δt=10−3\Delta t = 10^{-3},s (aΔt/Δx=0.7a\Delta t/\Delta x = 0.7) with the same fixed basis Φ\Phi. Relative L2L^2 error against FV remains low (2.8×10−42.8 \times 10^{-4} in pressure, 3.5×10−33.5 \times 10^{-3} in flux)—still well below the learned map's error—and the spatial/temporal profiles are visually indistinguishable from FV.
  • Certificate constant & conservation (W2): We now report the certificate constant (σmin⁡(M)=0.570\sigma_{\min}(M) = 0.570 at k=40k = 40, varying by <1%< 1\% across the horizon) and explicit constraint residuals (worst-case Kirchhoff imbalance of 2.4×10−82.4 \times 10^{-8} of peak flux; relative boundary residual of 1.7×10−141.7 \times 10^{-14}; pressure continuity exact by construction). We also explicitly clarify that Eq. 4.6 certifies the one-step projection at a training pair rather than at unseen inlet pressures.
  • Matched supervised baseline & PINN convergence (W1, Q5): We added a matched POD-NN baseline (latent and full-space supervised MSE) using the exact same trajectories, POD basis (r=35r = 35), architecture, and 2000-epoch budget. Both reduced-order methods outperform PI-DeepONet by 1–2 orders of magnitude, and NBM-OL beats POD-NN by 1.8×1.8\times–3.0×3.0\times in interpolated pressure while matching it in flux. We also verified that the trained PINN escapes the trivial basin (8×8\times lower loss, 0.030140.03014 vs. 0.243750.24375) and plateaus over the final 1300 epochs, yet still incurs 92%92\% flux error.
  • Scope, runtime, and continuous-time queries (W4, Q4, Q6): We report NBM solve times in the basis sweep and note that NBM-OL inference requires only a single forward pass at any (t,Pin)(t, P_{\mathrm{in}}). Extending to richer time-varying or multi-inlet boundary regimes (discussed in Conclusions and Future Work) remain directions for future work.

We hope these new experiments and clarifications address your concerns.

Reviewer_EPwqOpenReview ↗

Thanks for addressing my primary concern

My main concern was "NBM-OL is trained on solver trajectories and evaluated against the same solver, while PI-DeepONet and the PINN do not seem to get any data and must solve the full continuous problem. The baselines also appear to sit in the trivial-flow failure mode the authors describe in the appendix. There are no baselines trained on the same data and no ablation of the residual loss against a simple supervised loss, so the contribution of the method itself is unclear."

The authors have now addressed this by adding a matched POD-NN baseline. So, I have increased my score.

AuthorsOpenReview ↗

Thank you for updating your score & a brief clarification

Thank you for taking the time to re-evaluate our submission and for updating your score. We really appreciate your constructive feedback.

Just a brief clarification on one point from your quote: the trivial-flow collapse (q≡0q \equiv 0, P≡constP \equiv \text{const}) noted in Appendix E.3 refers specifically to standard, raw-scale PINN training without non-dimensionalization. With non-dimensionalization applied, our trained baselines escape that trivial basin and converge (verified quantitatively in Appendix F.5, Table 17).

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

The paper extends the Neural Basis Method to nonlinear transient gas flow on pipeline networks governed by the isothermal Euler equations. Fixed, edgewise neural bases represent pressure and mass flux, while a monolithic weighted least-squares projection of a staggered finite-volume discretization incorporates boundary and junction conditions. For parametric prediction, NBM-OL compresses reference coefficient trajectories using POD and learns a map from time and constant inlet pressure to reduced coefficients, enabling direct state evaluation without time stepping. Training uses a projected discrete residual assembled from reference trajectories. The paper also derives a residual-based bound on the distance to the best representable one-step solution.

Validation uses a seven-edge looped network and a twelve-edge branching network. The NBM solver reproduces the corresponding finite-volume solutions with relative errors of approximately 10−510^{-5} in pressure and 10−410^{-4} in flux. On held-out inlet pressures, NBM-OL reports mean pressure errors below 1% and mean flux errors below 3.3%, substantially outperforming the evaluated PINN and physics-informed DeepONet baselines. Tests cover interpolation and limited extrapolation beyond the training pressure range; topology, pipe properties, and outlet pressures remain fixed.

Critical strengths
  1. Accurate extension of the Neural Basis Method to coupled network dynamics. The edgewise representation and monolithic projection incorporate pipe equations and junction conditions within a unified discrete formulation. On both looped and branching networks, the method reproduces the finite-volume reference with relative errors of approximately 10−510^{-5} in pressure and 10−410^{-4} in flux. This provides convincing evidence that the proposed network construction accurately represents the targeted discrete dynamics.

  2. Effective parametric prediction within the tested setting. NBM-OL directly predicts the transient state at a queried time without time stepping, achieving mean pressure errors below 1% and mean flux errors below 3.3% across the reported interpolation and limited extrapolation tests. These results demonstrate practical promise for repeated queries on fixed networks, even though broader generalization remains untested.

Critical weaknesses
  1. The comparisons do not establish the added value over closely related reduced-order learning methods. The framework combines fixed neural bases, discrete least-squares projection, POD compression, and coefficient prediction, but the reported baselines are PINN and PI-DeepONet models that largely fail to recover the flux. A direct POD-based coefficient predictor trained on the same reference trajectories, together with ablations of the neural basis and residual objective, would better isolate the contribution. Since NBM-OL uses reference trajectories for both POD construction and residual assembly, comparisons against physics-only baselines do not establish whether the gains arise from the proposed formulation or access to reference information. This leaves the methodological significance insufficiently substantiated.

  2. The validation supports only a narrow form of parametric generalization. Each model is evaluated on a fixed topology with identical pipe properties, fixed outlet pressures, and a single scalar parameter controlling constant inlet pressure. Extrapolation extends only up to 10% of the training interval’s width. The experiments therefore do not establish adaptability to time-varying boundary conditions, heterogeneous pipes, or unseen network structures. Given the incremental nature of the methodological extension, this limited scope weakens the evidence for a broadly useful advance in learning network dynamics.

Questions for authors
  1. Clarification on the value beyond standard reduced-order learning. How does NBM-OL compare with a POD-based neural coefficient predictor using the same reference trajectories, latent dimension, and comparable training budget? A matched comparison would help establish the benefit of the fixed neural basis and projected-residual objective beyond conventional trajectory compression and regression.

  2. Clarification on conservation and the error certificate. Which boundary and junction conditions are satisfied exactly by the representation, and which are only minimized through the residual? Please quantify conservation violations at held-out pressures and clarify whether the one-step error bound assembled from reference states provides any guarantee for direct predictions at unseen pressures.

  3. Clarification on computational benefit. What are the complete offline costs, including reference generation, basis construction, POD, operator assembly, and network training, and the inference cost relative to the finite-volume solver? How many queries are needed to amortize these costs?

1 comment
AuthorsOpenReview ↗

Response to Reviewer (POD-NN Baseline, Conservation & Clarifications)

We thank the reviewer for the thoughtful and constructive feedback. We have uploaded a revised PDF with a new Additional Experiments appendix (highlighted) that directly addresses your questions and concerns:

  • Matched POD-NN baseline & basis ablation (W1, Q1): We added a matched supervised POD-NN baseline (both latent and full-space MSE variants) using the exact same reference trajectories, POD basis (r=35r = 35), architecture, and 2000-epoch budget. Both reduced-order methods outperform PI-DeepONet by 1–2 orders of magnitude, confirming that the NBM + POD representation drives the main accuracy gain. Within that representation, NBM-OL outperforms both supervised variants by 1.8×1.8\times–3.0×3.0\times in interpolated pressure (1.2×10−31.2 \times 10^{-3} vs. 2.1×10−32.1 \times 10^{-3} and 3.6×10−33.6 \times 10^{-3}), matches them in flux, and trains directly against the discrete governing equations rather than pure regression. We also added a basis-size ablation (k∈{10,20,40,80}k \in \{10, 20, 40, 80\}) showing monotonic convergence to the FV reference.
  • Conservation & error certificate (Q2): Junction pressure continuity is satisfied exactly by construction (via shared junction variables), while Dirichlet boundary conditions and Kirchhoff flux balance are enforced through the projected least-squares residual. Our new constraint-satisfaction analysis reports these residuals explicitly: at k=40k = 40, the relative boundary residual is 1.7×10−141.7 \times 10^{-14} and the worst-case Kirchhoff imbalance is 2.4×10−82.4 \times 10^{-8} of peak flux, with certificate constant σmin⁡(M)=0.570\sigma_{\min}(M) = 0.570. We also explicitly clarify that the one-step error bound applies at training pairs (where reference states assemble the system); assembling the residual from the model's own predictions for an online self-check at unseen pressures is discussed in Conclusions and Future Work.
  • Computational cost (Q3): We report NBM solve times in the basis sweep (207207,s at k=40k = 40 vs. 550550,s at k=80k = 80). Once trained, NBM-OL evaluates the full network state at any (t,Pin)(t, P_{\mathrm{in}}) in a single forward pass, so the offline cost (reference solves, POD, and 2000-epoch training) amortizes in many-query settings such as parameter sweeps or real-time monitoring.
  • Parametric scope (W2): We agree that extending validation to time-varying or multi-inlet boundary conditions, heterogeneous pipes, and unseen topologies is an important next step, and we explicitly acknowledge these boundaries in Conclusions and Future Work. Because the basis is constructed locally per edge and coupled only through junction rows, heterogeneous pipe parameters or larger networks fit directly into the same formulation without structural changes.

We hope these new experiments and clarifications address your concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.