Pith. sign in

REVIEW 4 major objections 5 minor 50 references

Topology-aware Neural Flux Prediction Guided by Physics

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read GNNs are direction-blind because message passing low-pass filters nodal signals; PhyNFP substitutes discretized difference operators and PDE constraints so high-frequency, direction-bearing components survive.

desk verdict A plausible new architecture with a confounded direction-sensitivity metric; worth refereeing but needs major experimental revision. read the letter →

arxiv 2506.05676 v1 pith:SPSZAAVO submitted 2025-06-06 cs.LG

Haoyang Jiang , Jindong Wang , Xingquan Zhu , Yi He This is my paper
classification cs.LG
keywords graphneuralnetworksfluxpredictiondirectedgraphsdiscretizeddifferencematricesphysics-informedlearningSaint-VenantequationsAw-Rasclehigh-frequencycomponents
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper sets out to explain and fix a specific blindness of GNNs on directed flow networks: message passing acts as a low-pass filter that erases the high-frequency, direction-localized variations in flux signals, which is why ordinary GNNs predict about equally well on a graph whether its edges point forward or backward. Its proposed fix, PhyNFP, replaces the adjacency matrix with discretized difference operators and writes the message-passing layer as a learned discretization of the governing conservation law—Saint-Venant for river networks, Aw-Rascle for traffic networks—so that high-frequency information is preserved instead of smoothed away. If the paper is right, direction awareness no longer has to be teased out of data by symmetric aggregation; it is built into the propagation operator itself, and a model can be expected to respect which way flow points on any graph whose dynamics obey a conservation law. The paper reports that this design reduces forward-setting flux prediction error by 31.6% on a Danube river network and 4.9% on a traffic network relative to the average of eight baselines, while producing the positive forward-versus-reverse sensitivity gap that standard GNNs lack.

What carries the argument

The carrying object is the discretized difference matrix $\hat D$: $\hat D_{ii}=1$ and $\hat D_{ij}=-1$ when $v_j$ is an upstream neighbor of $v_i$, so multiplying a nodal signal by $\hat D$ yields directed spatial gradients. Two scaled versions are used in the message-passing layer: $D_1=\hat D/\Delta x$ (spatial distance, propagation rate) and $D_2=(\Delta z/\Delta x)\hat D$ (elevation difference, gravitational forcing), with $\Delta x=\phi_1(e_{ij})$ and $\Delta z=\phi_2(e_{ij})$ learned from edge features. The layer update is then a discretized conservation law—Saint-Venant momentum for rivers, Aw-Rascle mass conservation for traffic—with learnable $\Delta t$ and $\hat g$ acting as stability-balancing coefficients. In the frequency domain the operator $I+\alpha\hat D$ has magnitude $|1+\alpha-\alpha e^{-j\omega}|$, so it leaves the low-frequency component at $\omega=0$ unchanged and amplifies high frequencies (at $\omega=\pi$ the magnitude is $|1+2\alpha|$); that frequency response is what replaces the low-pass behavior of standard message passing and lets direction survive.

What would settle it

A decisive check would be to train PhyNFP on the same graphs with all edge directions randomly shuffled, or replaced by an undirected topology of equal degree distribution, while keeping the reverse-graph control; if $DS$ remains strongly positive when direction carries no physical information, the claimed directional encoding is not what produces the gap. A second check is to measure the DTFT magnitude of the learned layer operator directly: if it does not amplify high frequencies near $\omega=\pi$ relative to a GCN's symmetrically normalized operator, the high-frequency preservation mechanism is not actually present in the trained model.

Watch

Extended reading notes

Core claim

The paper's claim is that the low-pass filtering behavior of GNN message passing is the mechanism behind their directional blindness: with $h^{l+1}_i = U^l(h^l_i, \sum_{j\in N_{in}[v_i]} M^l(h^l_i,h^l_j,e_{ji}))$, repeated aggregation drives neighboring embeddings together and suppresses the high-frequency nodal variations that carry flow direction. PhyNFP's contribution is to replace the adjacency matrix with discretized difference operators $\hat D$ (with entries $1$ on the diagonal and $-1$ for upstream neighbors) and to write each message-passing layer as a learned discretized PDE update—for rivers, $h^{l+1}=h^l-\Delta t\left(h^l\odot(D_1 h^l W_1)+\hat g\cdot(D_2 h^l W_2)\right)$ with $D_1=\hat D/\Delta x$ and $D_2=\Delta z\,\hat D/\Delta x$; for traffic, the analogous Aw-Rascle update using density and velocity embeddings. Under this construction the composite operator $I+\alpha\hat D$ has DTFT magnitude $|1+\alpha-\alpha e^{-j\omega}|$, which amplifies high frequencies near $\omega=\pi$ while leaving low frequencies unchanged, so the layer preserves exactly the directional signal ordinary message passing averages out. The empirical upshot claimed by the paper is that PhyNFP gives forward-setting MSE of 0.0801 on the river network and 0.0696 on the traffic network, beating the baseline averages of 0.1170 and 0.0732, and positive direction sensitivity $DS=\mathrm{MSE}_{\text{reverse}}-\mathrm{MSE}_{\text{forward}}$ of +0.0105 and +0.0028, where most baselines sit near zero or negative.

Load-bearing premise

The load-bearing premise is that $DS=\mathrm{MSE}_{\text{reverse}}-\mathrm{MSE}_{\text{forward}}$ is a clean measure of directional awareness—that reversing every edge makes the prediction task no harder and no easier for any model except through the physical meaning of direction, so the gap in errors cannot be attributed to optimization noise, model capacity, or incidental differences in the reversed graph.

Editorial extensions

If this is right

  • A PhyNFP-trained model should systematically outperform its same-capacity GNN counterpart whenever flow direction matters, and the margin should widen as the graph becomes more tree-like; the paper reports 31.6% lower forward MSE (0.0801 vs 0.1170 on the river network) and 4.9% lower forward MSE (0.0696 vs 0.0732 on the traffic network).
  • Injecting a local perturbation upstream should propagate coherently downstream in the forward setting, whereas in the reverse setting the same perturbation should travel upstream as an instability signature of solving the PDE backward; the paper reports exactly this pattern for PhyNFP and not for GCN or ResGCN.
  • The learned time step $\Delta t$ should adapt to problem conditioning, settling at a smaller value in the reverse, ill-posed setting; the paper observes a converged value about 21.58% smaller when edges are reversed, which it reads as evidence that the layer behaves like a discretized PDE solver rather than a generic smoother.
  • The framework should transfer to other conservation-law flow networks by swapping the governing PDE and the definitions of $D_1$ and $D_2$, since the paper frames the difference-matrix construction as generalizable and tests it on two different physical systems with two different governing equations.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Our inference: the direction-sensitivity metric $DS$ could serve as a cheap diagnostic for any spatiotemporal GNN, independent of physics training; a near-zero or negative $DS$ would flag a model that silently ignores edge orientation, and a positive $DS$ would indicate that reversal actually changes the learned mapping.
  • Our inference: the paper's contrast between the tree-like river network and the cyclic traffic network implies a testable structural claim—direction sensitivity should decrease monotonically as cycles are added to a synthetic directed graph even when the same PDE constraint is used, because cycles create message-passing routes that blur forward and reverse topologies.
  • Our inference: the learned $\Delta t$ could be interpreted as an inferred Courant number; plotting converged $\Delta t$ values against physical flow speeds would test whether the network is recovering a meaningful stability limit rather than merely tuning the loss, a check the paper does not run.
  • Our inference: preserving high-frequency components instead of smoothing them may change how these models extrapolate under distribution shift, such as extreme flood events or sudden congestion, because sharp local signals are no longer attenuated; this is a direct consequence of the proposed mechanism that the paper does not evaluate.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes PhyNFP, a graph neural network framework for flux prediction on directed graphs. The method replaces the adjacency matrix used in standard message passing with discretized upwind difference matrices, and augments the message-passing update with PDE-inspired terms: Saint-Venant momentum terms for river networks and Aw-Rascle mass-conservation terms for traffic networks. Experiments on a Danube river dataset (LamaH-CE2) and a traffic dataset (PEMS-04) compare PhyNFP against eight baselines in forward and reverse edge-direction settings. The paper reports lower forward MSE for PhyNFP (0.0801 vs. 0.1170 averaged baselines on river; 0.0696 vs. 0.0732 on traffic) and a larger direction-sensitivity metric DS = MSE(reverse) - MSE(forward), which it interprets as evidence that the model internalizes edge direction and preserves high-frequency components. The central claim is that explicit difference matrices plus physics-guided implicit constraints make GNNs direction-sensitive and improve flux prediction.

Significance. If established, the core idea is a worthwhile contribution: replacing symmetric adjacency with direction-sensitive difference operators is a principled and reasonably general way to inject directional inductive bias, and the PDE-inspired updates in Eqs. (8), (11), (12), and (13) are coherent discretizations. The authors provide public code, evaluate on two real-world directed networks, include an ablation isolating the effect of the PDE constraints, and report horizon sensitivity. The forward-MSE improvement on the river dataset is substantial. However, the distinctive claims about directional and high-frequency sensitivity rest on the DS metric, which is confounded with task difficulty, and on qualitative perturbation plots without uncertainty quantification. As presented, the evidence does not yet support the headline sensitivity uplifts, and the traffic-dataset gains are small enough that repeated-seed statistics are essential. With additional controls and statistical rigor, the paper could become publishable.

major comments (4)
  1. [Section 4 (Direction Sensitivity; Table 1; Appendix C)] The definition DS(M) = L_M(Reverse) - L_M(Forward) is treated as a measure of directional awareness, but reversing every edge changes the boundary structure, node roles, and conditioning of the prediction problem. The manuscript itself argues in Appendix C that the reverse task is an ill-posed inverse problem with amplified noise, so a model that has fit the forward graph well may incur larger reverse error simply because the reverse task is harder, independent of any representation of direction. Table 1 contains baselines with negative DS (GraphSAGE -0.0075, MP-PDE Solver -0.0044, GCN -0.0008 on river), showing that the reversed graph can be easier for some architectures, so DS is not a monotone proxy for direction sensitivity. This confound invalidates the inference that the 96.5% and 79.9% RDS uplifts reflect directional encoding. Please add controls that hold task difficulty fixed, e.g., random edge-direction permutations, degree- and boundary-matched reversed graphs, or synthetic graphs with known directional ground truth, and report DS relative to a permutation baseline.
  2. [Section 4 (RDS definition; Table 1)] The relative direction sensitivity RDS(M1, M2) = (DS(M2) - DS(M1)) / DS(M1) divides by a near-zero and sometimes negative denominator. The average baseline DS on the river dataset is 0.0004, with individual values ranging from -0.0075 to +0.0032; ratios to such denominators are numerically unstable, so the claimed 26x, 96.5%, and 79.9% improvements are not meaningful summary statistics. In the traffic dataset the absolute DS difference between PhyNFP (0.0028) and the baseline average (0.0006) is 0.0022, which is close to the scale of the noise floor given that Table 1 reports only point estimates with no repeated seeds or confidence intervals. Please report per-model DS distributions over random seeds, bootstrap intervals, and a stable effect-size measure such as absolute DS difference or a permutation test, rather than percentage changes relative to near-zero baselines.
  3. [Section 3.3 and RQ4 (Eq. (12); Figure 2; Appendix B)] The central architectural claim is that PhyNFP 'preserves high-frequency components' and 'models rapid spatial and directional variations,' but no experiment measures the frequency content of the model's predictions or internal representations. Appendix B analyzes the frequency response of the difference operator D and the composite I + alpha D, not of a trained PhyNFP model; that analysis shows only that the operator is high-pass, not that the GNN retains high-frequency information in its outputs. Figure 2 is qualitative, reports no error bars on the perturbation-response curves, and does not quantify the difference between GCN, ResGCN, and PhyNFP. Please add a direct test, for example the DTFT or a high-pass-filtered error of test-set predictions under PhyNFP versus baselines, or a quantitative perturbation-propagation amplitude with confidence intervals.
  4. [RQ5 and Appendix C (Figure 3; Figure 4)] The evidence for physics extraction in the reverse setting is post-hoc interpretation rather than a controlled test. The learned time step Delta-t converging to a value 21.58% smaller in the reverse setting (Figure 3) is presented as evidence of PDE-consistent stability behavior, but no comparison is made to a non-physics model trained with the same initialization, and no formal stability analysis links the learned scalar to the discretization scheme. Similarly, Figure 4 reports that PhyNFP propagates a perturbation upstream in the reversed graph, and this is interpreted as a signature of solving the PDE backward; but there is no quantitative metric, no comparison with a model that is direction-aware without PDE constraints, and no ablation showing that this behavior is necessary for the reported performance. Please add a controlled experiment, e.g., freezing Delta-t across forward and reverse settings, or comparing with PhyNFPDM under the same reverse perturbation, and report a quantitative propagation measure.
minor comments (5)
  1. [Section 3.1] There is a typo: 'we leverage the the upwind scheme' should read 'we leverage the upwind scheme.'
  2. [Section 2] The text refers to 'Figure 2 (right)' when discussing reversed and undirected message-passing loss trends, but Figure 1 contains the right panel with MSE trends; the figure reference appears to be a typo.
  3. [RQ4 and Figure 2] The y-axis description is inconsistent: the main text and figure caption describe the y-axis as 'the difference between perturbed and unperturbed predictions,' while another caption line calls it 'responsive prediction errors (in MSE)'; please use one consistent definition and state units.
  4. [Appendix A and Figure 4] The appendix caption says the perturbation propagates to upstream nodes v2 through v6, while the main-text description of Figure 2 lists nodes v2, v3, and v4; the node sets and graph drawings should be reconciled.
  5. [References] Some references are incomplete or malformed (e.g., the Kesting and Treiber entry and the RS & Williams entry); please check that all entries have complete titles and venues.

Circularity Check

2 steps flagged · score 2.0 of 10

Central MSE benchmarks are independent, but two RQ5 'physics extraction' validations are self-referential rather than external predictions.

  1. self definitional [Section 4, RQ5 (reverse-setting perturbation analysis); Eq. (12) in Section 3.3]
    "For PhyNFP, the perturbation incorrectly propagates upstream (to v2, v3, . . .). While physically incorrect for forward flow, this behavior is the expected signature of solving the PDE backward from downstream data. The response of PhyNFP in this setting, which directly reflects the ill-posed and potentially unstable nature of this inverse problem, thus demonstrates its capture of the PDE-encoded dynamics."

    Eq. (12) is not an independent probe: it is constructed from the PDE-discretized update Eq. (8) with explicit upwind difference matrices D1 and D2 and the term h^l ⊙ (D1 h^l W1). Applying that layer to a reversed graph therefore propagates a perturbation along whichever direction D points. Observing upstream propagation is entailed by the layer's definition, so citing that behavior as 'demonstrating' capture of PDE-encoded dynamics is self-definitional: the evidence is the architecture itself, not an external physical quantity. No PDE residual or independent physical measurement is reported.

  2. fitted input called prediction [Section 4, RQ5 (learned time-step evidence); Figure 3]
    "Further evidence comes from the learned time step parameter ∆t. As shown in Figure 3, ∆t stabilizes at a higher value in the forward setting, whereas in the reverse setting it converges to a value approximately 21.58% smaller. This reflects the need for tighter step sizes to ensure stability when solving ill-posed inverse problems (Baumeister, 1987)."

    ∆t is a learnable scalar trained by backpropagation on the prediction MSE; nothing in the objective ties it to CFL stability or to the PDE. The smaller learned value in the reverse setting is a fitted outcome that could equally arise from a shallower effective gradient, optimization noise, or a changed loss surface. Using that fitted parameter as 'further evidence' that the model captures inverse-problem stability is a fitted input relabeled as a physics prediction, since Eq. (12) treats ∆t as a free balancing scalar rather than a quantity derived from the governing equations.

full rationale

The central performance claim is not circular: Table 1 reports held-out forward and reversed MSE, and the learnable W1, W2, ∆t, and g parameters are ordinary trained parameters rather than calibrated surrogates for the reported metrics. The DS metric is a design assumption about task difficulty, not a quantity fitted to itself, so concerns about DS conflating reversal with difficulty are methodological rather than derivationally circular. However, two passages in RQ5 present as validation what is actually a restatement of the architecture: the reverse-perturbation response follows from the upwind difference matrices hard-coded into Eq. (12), and the learned ∆t is a fitted scalar whose smaller reverse value is post-hoc interpreted as numerical stability. Neither passage undermines the independent benchmark results, and no load-bearing self-citation or imported uniqueness theorem appears, so the paper is only mildly self-referential rather than derivationally circular.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

No new physical entities are posited. The method introduces learnable scalars and difference operators, not new forces, particles, dimensions, or conserved quantities.

free parameters (4)
  • learnable time step Delta t = initialized to 0.7, converges lower in reverse settings
    Introduced in Eq. (12)/(13) as a learnable scalar modulating update magnitude; its converged value is used as evidence of physics behavior, so the central claim depends on it.
  • learnable gravity scale \hat g = not reported
    Learnable scalar in Eq. (12) scaling the elevation-driven term; fitted during training and not independently measured.
  • learnable mappings phi_1, phi_2 for Delta x and Delta z = MLPs, values not reported
    Eq. (4) defines Delta x = phi_1(e_ij) and Delta z = phi_2(e_ij); these determine D1 and D2, which are the central mechanism for directional sensitivity.
  • layer weights W1 and W2 = learned, not reported
    Standard learnable projection matrices in Eq. (12)/(13); they refine embeddings but are ordinary neural parameters.
assumptions (4)
  • domain assumption Flow in the river network is governed by the simplified Saint-Venant momentum equation, and traffic flow is governed by the Aw-Rascle mass conservation equation.
    Section 3.2 uses these PDEs to derive the update rules; if the simplified PDEs misrepresent the data generation, the physical guidance could inject systematic bias.
  • domain assumption Message-passing aggregation acts as a low-pass filter that suppresses high-frequency nodal signals.
    Assumed from cited literature and Figure 1; it motivates the whole approach and the direction-sensitivity hypothesis.
  • domain assumption Reversing every edge is equivalent to solving an ill-posed inverse hyperbolic problem, so a directional model should perform worse in the reverse setting.
    Appendix C argues this by analogy with characteristic tracing; the DS metric depends on this premise.
  • standard math Upwind finite-difference discretization is numerically stable and adequate at these graph scales.
    Section 3.1 invokes upwind schemes; the only stability evidence is the high-pass DTFT analysis in Appendix B, not a full CFL or stability analysis.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Topology-aware Neural Flux Prediction Guided by Physics." pith.science (2026). https://pith.science/paper/SPSZAAVO

@misc{pith2026250605676,
  author       = {Pith},
  title        = {Pith review of: Topology-aware Neural Flux Prediction Guided by Physics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SPSZAAVO}},
  note         = {Machine review of arXiv:2506.05676}
}
read the original abstract

Graph Neural Networks (GNNs) often struggle in preserving high-frequency components of nodal signals when dealing with directed graphs. Such components are crucial for modeling flow dynamics, without which a traditional GNN tends to treat a graph with forward and reverse topologies equal.To make GNNs sensitive to those high-frequency components thereby being capable to capture detailed topological differences, this paper proposes a novel framework that combines 1) explicit difference matrices that model directional gradients and 2) implicit physical constraints that enforce messages passing within GNNs to be consistent with natural laws. Evaluations on two real-world directed graph data, namely, water flux network and urban traffic flow network, demonstrate the effectiveness of our proposal.

Figures

Figures reproduced from arXiv: 2506.05676 by the authors.

Figure 1
Figure 1. Left: Trends of temporal gradients w.r.t. the increasing number of message-passing layers. Right: MSE Trends of GCN in the original (Forward), inverse (Reverse), and undirected network settings w.r.t. the increasing number of message-passing layers. Technical Challenges GNNs leverage neighborhood aggregation to yield node embeddings that harmonize information from both nodal features and graph topology. Denoted by h… view at source ↗
Figure 2
Figure 2. Trends of prediction results in response to a local and rapid flux change. (a) The change occurs in node v1 and propagates to the downstream nodes v2, v3, and v4. The responsive prediction errors (in MSE) across the four nodes from (b) GCN, (c) ResGCN, and (d) our PhyNFP framework. Traffic network compared to the River network, stemming from two main factors. First, the governing physics dif￾fer: river flow modeling… view at source ↗
Figure 3
Figure 3. Evolution of the learned time-step parameter ∆t over training epochs for the forward and reverse settings in the river network. The model starts with an initial ∆t = 0.7. value in the forward setting, whereas in the reverse setting it converges to a value approximately 21.58% smaller. This re￾flects the need for tighter step sizes to ensure stability when solving ill-posed inverse problems (Baumeister, 1987). These … view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Trends of prediction results in response to a local and rapid flux change. (a) The change occurs in node v1 and propagates to the upstream nodes v2 through v6. The responsive prediction errors (in MSE) across the six nodes are shown for (b) GCN, (c) ResGCN, and (d) our…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

50 extracted references · 49 canonical work pages

  1. [1]

    and Rascle, M

    Aw, A. and Rascle, M. Resurrection of second-order models of traffic flow. SIAM journal on applied mathematics, 60 0 (3): 0 916--938, 2000

    work page 2000

  2. [2]

    Stable solution of inverse problems

    Baumeister, J. Stable solution of inverse problems. Springer, 1987

    work page 1987

  3. [3]

    and Vazquez, M

    Bermudez, A. and Vazquez, M. E. Upwind methods for hyperbolic conservation laws with source terms. Computers & Fluids, 23 0 (8): 0 1049--1071, 1994

    work page 1994

  4. [4]

    Beyond low-frequency information in graph convolutional networks

    Bo, D., Wang, X., Shi, C., and Shen, H. Beyond low-frequency information in graph convolutional networks. In AAAI, volume 35, pp.\ 3950--3957, 2021

    work page 2021

  5. [5]

    E., and Welling, M

    Brandstetter, J., Worrall, D. E., and Welling, M. Message passing neural PDE solvers. In ICLR, 2022

    work page 2022

  6. [6]

    Brunner, G. W. Hec-ras (river analysis system). In North American water and environment congress & destructive water, pp.\ 3782--3787, 2002

    work page 2002

  7. [7]

    N., Cho, J., and Yi, H

    Bui, K.-H. N., Cho, J., and Yi, H. Spatial-temporal graph neural network for traffic forecasting: An overview and open research issues. Applied Intelligence, 52 0 (3): 0 2763--2774, 2022

    work page 2022

  8. [8]

    Spectral Methods: Evolution to Complex Geometries and Applications to Fluid Dynamics

    Canuto, C. Spectral Methods: Evolution to Complex Geometries and Applications to Fluid Dynamics. Springer-Verlag, 2007

    work page 2007

Show all 50 references
  1. [9]

    I., Bronstein, M., Webb, S., and Rossi, E

    Chamberlain, B., Rowbottom, J., Gorinova, M. I., Bronstein, M., Webb, S., and Rossi, E. Grand: Graph neural diffusion. In ICML, pp.\ 1407--1418, 2021

    2021

  2. [10]

    Pde-gcn: Novel architectures for graph neural networks motivated by partial differential equations

    Eliasof, M., Haber, E., and Treister, E. Pde-gcn: Novel architectures for graph neural networks motivated by partial differential equations. In NeurIPS, volume 34, pp.\ 3836--3849, 2021

    2021

  3. [11]

    A spatial--temporal graph deep learning model for urban flood nowcasting leveraging heterogeneous community features

    Farahmand, H., Xu, Y., and Mostafavi, A. A spatial--temporal graph deep learning model for urban flood nowcasting leveraging heterogeneous community features. Scientific Reports, 13 0 (1): 0 6768, 2023

    2023

  4. [12]

    Rainfall-runoff modeling in a flashy tropical watershed using the distributed HL-RDHM model

    Fares, A., Awal, R., Michaud, J., Chu, P.-S., Fares, S., Kodama, K., and Rosener, M. Rainfall-runoff modeling in a flashy tropical watershed using the distributed HL-RDHM model. Journal of Hydrology, 519: 0 3436--3447, 2014

    2014

  5. [13]

    Ferrari, A., D'Oria, M., Vacondio, R., Dal Pal \`u , A., Mignosa, P., and Tanda, M. G. Discharge hydrograph estimation at upstream-ungauged sections by coupling a bayesian methodology and a 2-d GPU shallow water model. Hydrology and Earth System Sciences, 22 0 (10): 0 5299--5316, 2018

    2018

  6. [14]

    K., Pan, M., and Wood, E

    Fisher, C. K., Pan, M., and Wood, E. F. Spatiotemporal assimilation--interpolation of discharge records through inverse streamflow routing. Hydrology and Earth System Sciences, 24 0 (1): 0 293--305, 2020

    2020

  7. [15]

    S., Riley, P

    Gilmer, J., Schoenholz, S. S., Riley, P. F., Vinyals, O., and Dahl, G. E. Message passing neural networks. Machine learning meets quantum physics, pp.\ 199--214, 2020

    2020

  8. [16]

    Grady, L. J. and Polimeni, J. R. Discrete calculus: Applied analysis on graphs for computational science, volume 3. Springer, 2010

    2010

  9. [17]

    Attention based spatial-temporal graph convolutional networks for traffic flow forecasting

    Guo, S., Lin, Y., Feng, N., Song, C., and Wan, H. Attention based spatial-temporal graph convolutional networks for traffic flow forecasting. In AAAI, volume 33, pp.\ 922--929, 2019

    2019

  10. [18]

    Inductive representation learning on large graphs

    Hamilton, W., Ying, Z., and Leskovec, J. Inductive representation learning on large graphs. In NeurIPS, volume 30, 2017

    2017

  11. [19]

    Revisiting graph neural networks: Graph filtering perspective

    Hoang, N., Maehara, T., and Murata, T. Revisiting graph neural networks: Graph filtering perspective. In ICPR, pp.\ 8376--8383, 2021

    2021

  12. [20]

    Spatio-temporal graph neural networks for predictive learning in urban computing: A survey

    Jin, G., Liang, Y., Fang, Y., Shao, Z., Huang, J., Zhang, J., and Zheng, Y. Spatio-temporal graph neural networks for predictive learning in urban computing: A survey. IEEE Transactions on Knowledge and Data Engineering, 2023

    2023

  13. [21]

    and Zhang, H

    Jin, W. and Zhang, H. M. The formation and structure of vehicle clusters in the payne--whitham traffic flow model. Transportation Research Part B: Methodological, 37 0 (3): 0 207--223, 2003

    2003

  14. [22]

    Floodgnn-gru: A spatio-temporal graph neural network for flood prediction

    Kazadi, A., Doss-Gollin, J., Sebastian, A., and Silva, A. Floodgnn-gru: A spatio-temporal graph neural network for flood prediction. Environmental Data Science, 2023

    2023

  15. [23]

    and Treiber, M

    Kesting, A. and Treiber, M. Traffic flow dynamics: data, models and simulation. no. Book, Whole)(Springer Berlin Heidelberg, Berlin, Heidelberg, 2013), 2013

    2013

  16. [24]

    and Sun, Y

    Kirschstein, N. and Sun, Y. The merit of river network topology for neural flood forecasting. In ICML, 2024

    2024

  17. [25]

    Lama H - CE : Large-sample data for hydrology and environmental sciences for central europe

    Klingler, C., Schulz, K., and Herrnegger, M. Lama H - CE : Large-sample data for hydrology and environmental sciences for central europe. Earth System Science Data Discussions, 2021: 0 1--46, 2021

    2021

  18. [26]

    Large-scale river network modeling using graph neural networks

    Kratzert, F., Klotz, D., Gauch, M., Klingler, C., Nearing, G., and Hochreiter, S. Large-scale river network modeling using graph neural networks. In EGU General Assembly Conference Abstracts, pp.\ EGU21--13375, 2021

    2021

  19. [27]

    lighthill--whitham--richards

    Leclercq, L. Hybrid approaches to the solutions of the “lighthill--whitham--richards” model. Transportation Research Part B: Methodological, 41 0 (7): 0 701--709, 2007

    2007

  20. [28]

    B., and M \"u ller, K.-R

    LeCun, Y., Bottou, L., Orr, G. B., and M \"u ller, K.-R. Efficient backprop. In Neural networks: Tricks of the trade, pp.\ 9--50. Springer, 2002

    2002

  21. [29]

    LeVeque, R. J. Finite volume methods for hyperbolic problems, volume 31. Cambridge university press, 2002

    2002

  22. [30]

    Multipole graph neural operator for parametric partial differential equations

    Li, Z., Kovachki, N., Azizzadenesheli, K., Liu, B., Stuart, A., Bhattacharya, K., and Anandkumar, A. Multipole graph neural operator for parametric partial differential equations. In NeurIPS, volume 33, pp.\ 6755--6766, 2020

    2020

  23. [31]

    Multi-resolution partial differential equations preserved learning framework for spatiotemporal dynamics

    Liu, X.-Y., Zhu, M., Lu, L., Sun, H., and Wang, J.-X. Multi-resolution partial differential equations preserved learning framework for spatiotemporal dynamics. Communications Physics, 7 0 (1): 0 31, 2024

    2024

  24. [32]

    P DE - N et: Learning PDE s from data

    Long, Z., Lu, Y., Ma, X., and Dong, B. P DE - N et: Learning PDE s from data. In ICML, pp.\ 3208--3216, 2018

    2018

  25. [33]

    Parameterization of distributed hydrological models: learning from the experiences of lumped modeling

    Moreda, F., Koren, V., Zhang, Z., Reed, S., and Smith, M. Parameterization of distributed hydrological models: learning from the experiences of lumped modeling. Journal of Hydrology, 320 0 (1-2): 0 218--237, 2006

    2006

  26. [34]

    Graph neural ordinary differential equations

    Poli, M., Massaroli, S., Park, J., Yamashita, A., Asama, H., and Park, J. Graph neural ordinary differential equations. In AAAIW, 2019

    2019

  27. [35]

    S., Punekar, S

    Roudbari, N. S., Punekar, S. R., Patterson, Z., Eicker, U., and Poullis, C. From data to action in flood forecasting leveraging graph neural networks and digital twin visualization. Scientific reports, 14 0 (1): 0 18571, 2024

    2024

  28. [36]

    RS, A. J. S. R. M. and Williams, J. Large area hydrologic modeling and assessment part i: model development. Journal of American Water Resources Association, 34 0 (1): 0 73, 1998

    1998

  29. [37]

    K., Chamberlain, B., Rowbottom, J., Mishra, S., and Bronstein, M

    Rusch, T. K., Chamberlain, B., Rowbottom, J., Mishra, S., and Bronstein, M. Graph-coupled oscillator networks. In ICML, pp.\ 18888--18909, 2022

    2022

  30. [38]

    K., Chamberlain, B

    Rusch, T. K., Chamberlain, B. P., Mahoney, M. W., Bronstein, M. M., and Mishra, S. Gradient gating for deep multi-rate learning on graphs. In ICLR, 2023

    2023

  31. [39]

    Large eddy simulation for incompressible flows: an introduction

    Sagaut, P. Large eddy simulation for incompressible flows: an introduction. Springer Science & Business Media, 2005

    2005

  32. [40]

    Improving your graph neural networks: a high-frequency booster

    Sun, J., Zhang, L., Zhao, S., and Yang, Y. Improving your graph neural networks: a high-frequency booster. In ICDMW, pp.\ 748--756, 2022

    2022

  33. [41]

    Graph attention networks

    Veli c kovi \'c , P., Cucurull, G., Casanova, A., Romero, A., Li \`o , P., and Bengio, Y. Graph attention networks. In ICLR, 2018

    2018

  34. [42]

    Vreugdenhil, C. B. Numerical methods for shallow-water flow, volume 13. Springer Science & Business Media, 2013

    2013

  35. [43]

    Simplifying graph convolutional networks

    Wu, F., Souza, A., Zhang, T., Fifty, C., Yu, T., and Weinberger, K. Simplifying graph convolutional networks. In ICML, pp.\ 6861--6871, 2019

    2019

  36. [44]

    Computational river dynamics

    Wu, W. Computational river dynamics. Crc Press, 2007

    2007

  37. [45]

    Graph wavelet neural network

    Xu, B., Shen, H., Cao, Q., Qiu, Y., and Cheng, X. Graph wavelet neural network. In ICLR, 2019

    2019

  38. [46]

    Spatio-temporal graph convolutional networks: a deep learning framework for traffic forecasting

    Yu, B., Yin, H., and Zhu, Z. Spatio-temporal graph convolutional networks: a deep learning framework for traffic forecasting. In IJCAI, pp.\ 3634--3640, 2018

    2018

  39. [47]

    and Krstic, M

    Yu, H. and Krstic, M. Traffic congestion control for aw--rascle--zhang model. Automatica, 100: 0 38--51, 2019

    2019

  40. [48]

    Topology learning-based fuzzy random neural networks for streaming data regression

    Yu, H., Lu, J., and Zhang, G. Topology learning-based fuzzy random neural networks for streaming data regression. IEEE Transactions on Fuzzy Systems, 30 0 (2): 0 412--425, 2020

    2020

  41. [49]

    Real-time prediction system of train carriage load based on multi-stream fuzzy learning

    Yu, H., Lu, J., Liu, A., Wang, B., Li, R., and Zhang, G. Real-time prediction system of train carriage load based on multi-stream fuzzy learning. IEEE Transactions on Intelligent Transportation Systems, 23 0 (9): 0 15155--15165, 2022

    2022

  42. [50]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.