Pith. sign in

REVIEW 3 major objections 5 minor 33 references

FluxNet claims that learning how much mass moves between neighboring cells, instead of predicting the next state, makes discrete conservation exact and physical bounds structural.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 05:29 UTC pith:PSDC2FDH

load-bearing objection Solid architecture paper with clean conservation and one-sided bound guarantees; the dual-bound D-head is an empirical regularizer, not a structural guarantee, and the abstract oversells it. the 3 major comments →

arxiv 2602.01941 v2 pith:PSDC2FDH submitted 2026-02-02 cond-mat.mtrl-sci cs.CEcs.LGphysics.comp-ph

FluxNet: Learning Capacity-Constrained Local Transport Operators for Conservative and Bounded PDE Surrogates

Zishuo Lan , Junjie Li , Lei Wang , Jincheng Wang This is my paper
classification cond-mat.mtrl-sci cs.CEcs.LGphysics.comp-ph MSC 68T0765M06
keywords conservation lawsneural PDE surrogatestransport operatorsbound preservationautoregressive rolloutlattice BoltzmannCahn-Hilliardshallow water equations
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper argues that the instability of long autoregressive PDE rollouts comes from predicting states directly, so it predicts a transport plan instead: for each grid cell and each neighbor direction, the network outputs how much of the conserved quantity moves in one surrogate step. Updating each cell by tallying outflows and inflows preserves the global sum exactly, at machine precision, regardless of the learned weights. A family of capacity-constrained heads turns physical bounds into architecture: outflow limited by the amount above a lower bound guarantees the lower bound; inflow limited by remaining capacity guarantees the upper bound; a dual head combining both with a consistency loss keeps two-sided violations near zero in practice. Because the model moves material over a configurable neighborhood rather than integrating flux rates, it can take much larger time steps, up to a 17.3x speedup on spinodal decomposition, while preserving both pointwise accuracy and statistical microstructure.

Core claim

The central claim is that conservation and boundedness are properties of the parameterization, not the training. Under periodic boundaries with a symmetric stencil, the update u^{t+1}_i = u^t_i - sum_j F_{i->j} + sum_j F_{j->i} cancels outflows against inflows in the global sum identically, so Proposition 1 holds for any flux values. The L-head, which sends at most a sigmoid fraction of the available amount above a lower bound, gives a hard guarantee that the bound is never crossed; the U-head does the same for upper bounds by limiting what a cell can receive. The D-head averages an outflow branch and an inflow branch and uses a dual-consistency loss to make them agree; the paper states plai

What carries the argument

The central object is a directed local transport plan F_{i->j} on a symmetric stencil, a fixed set of neighbor offsets with configurable radius. Instead of regressing u^{t+1}, a backbone encoder extracts features and a transport head turns them into per-direction transfer amounts; the conservative update sums outflows and inflows. The L-head parameterizes outflow as a_i * alpha_i * pi_{i->j}, where a_i = u_i - ell is the amount available above the lower bound, alpha_i in (0,1) comes from a sigmoid, and pi is a softmax distribution over neighbors, making the lower bound a hard structural guarantee. The U-head is its mirror image, filling a fraction of the remaining capacity b_i = u_max - u_i.

Load-bearing premise

The dual-bound claim rests on the empirical premise that the Dual Consistency Loss makes the outflow and inflow branches agree closely enough that their averaged update respects both lower and upper bounds; the paper explicitly states in Section 3.4 and Appendix A that this averaged update is not theoretically guaranteed to satisfy both bounds unless the branches agree exactly, and measured violation rates reach 7.09% in the traffic-flow FNO variant.

What would settle it

Run a dual-bounded benchmark with the D-head but turn the Dual Consistency Loss off while keeping everything else identical, and measure the pointwise violation rate as a function of |Delta-u_out - Delta-u_in|: the paper's mechanism predicts violations only where the branches disagree, so a case with large disagreement and zero violations, or large agreement and many violations, would falsify the dual-bound story.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the transport-plan parameterization is correct, conservation drift in autoregressive rollouts is eliminated by construction, to machine precision, with no post-hoc mass renormalization.
  • One-sided physical bounds such as nonnegativity and capacity limits become hard guarantees that hold for any learned parameters, so violation rates drop to zero for L-head and U-head problems.
  • Because the surrogate moves material over an enlarged neighborhood instead of integrating flux rates, it escapes CFL-style step-size limits, enabling 1000-dt steps and a 17.3x speedup on coarsening microstructures.
  • Dual-bounded fields such as traffic density and phase-field concentration can be handled with near-zero violations when the two branches are aligned by DCL, with violation magnitudes small enough not to destabilize rollouts.
  • The shallow-water LAP instantiation suggests coupled conserved fields can be transported coherently by linking momentum flux to mass flux with pressure gating, improving multi-field stability.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper: the conservation guarantee is stated for source-free periodic domains; adding source terms or open boundaries would require splitting the update into transport plus source or handling boundary fluxes explicitly, which the ghost-cell sketch does not fully settle.
  • A testable extension: the dual-bound guarantee would become strict if the two branches were combined by a learned convex combination that respects both capacities, or by an in-network projection onto the feasible set, a direction the paper itself hints at.
  • Connection: the transport-plan view is essentially a learned finite-volume or lattice-Boltzmann update; a natural stress test is to train FluxNet on a scalar conservation law with an exact solution and compare learned stencil weights against the analytic flux.
  • One could measure branch disagreement directly and check whether violation rate tracks it; the paper's own explanation predicts violations only where the branches disagree.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes FluxNet, a neural surrogate architecture for conservative PDEs on regular grids. Instead of predicting the next state, the network outputs directed transport amounts on a fixed stencil; the update rule in Eq. (2) then guarantees discrete conservation by construction (Prop. 1). Three transport heads are introduced: L-head enforces a lower bound via capacity-limited outflow (Prop. 2), U-head enforces an upper bound via capacity-limited inflow (Prop. 3), and D-head averages one L-type and one U-type branch with a dual-consistency loss (DCL, Eq. (4)) to approximate two-sided bound satisfaction. Experiments cover 1D convection-diffusion, 2D shallow water, 1D traffic flow, and 2D Cahn-Hilliard spinodal decomposition, reporting rollout accuracy at 2x temporal extrapolation, conservation drift, bound-violation rates, ablations, ERF analysis, and wall-clock speedups. The paper claims exact conservation, hard one-sided bound guarantees, and near-zero dual-bound violations via architectural design, with code made public.

Significance. The central structural ideas are clean and useful: parameterizing the learned object as cumulative transport rather than a next-state map is a simple but effective way to make discrete conservation exact and independent of learned parameters, and the L/U capacity parameterizations give genuine hard one-sided bound guarantees. The benchmark suite is diverse, includes 2x temporal extrapolation, ± error bars, ablations, and a statistical-fidelity analysis for spinodal decomposition, and the code is released. These are real strengths. However, the headline claim that the D-head provides 'near-zero dual-bound violations through architectural design' is not supported: the D-head is an empirical soft-constrained average of two one-sided branches, as the paper itself concedes in §3.4 and Appendix A. Because the dual-bounded case is the most demanding and the abstract/Table 1 overstate the guarantee, the significance of the work depends on recalibrating those claims and providing the missing violation statistics for the spinodal benchmark.

major comments (3)
  1. [Abstract, §3.4, Table 4] The abstract states that the D-head enforces 'near-zero dual-bound violations through architectural design,' and Table 1 labels it 'Dual capacity (out/in)' with the implication of a structural guarantee. This is not justified. Eq. (3) averages the two branch updates, so the averaged flux can exceed the sender's available amount and/or the receiver's remaining capacity; the paper itself notes in §3.4 and Appendix A that no theoretical dual-bound guarantee holds unless the branches agree exactly. Empirically, Table 4 reports Vlb=7.09% and Mub=3.55e-3 for the FNO backbone, and Vub=0.84% for the ResNet backbone, which are not 'near-zero' in a way that supports the abstract's phrasing. The abstract, intro, and Table 1 should be revised to present the D-head as an empirical soft-constrained mechanism with transparently reported violations, not as an architectural guarantee. This is load-bearin
  2. [Table 5 (Spinodal Decomposition)] The Cahn-Hilliard phase field is dual-bounded, and the models are FluxNet-D, yet Table 5 reports only MAE, ERF, and speedup: no bound-violation rates or conditional violation magnitudes are given. Without Vlb, Vub, Mlb, and Mub for the 10Δt, 100Δt, and 1000Δt models, the claim of effective empirical dual-bound enforcement on this benchmark is unsubstantiated. These statistics should be reported, since the paper elsewhere correctly treats violation rates as a first-class metric for dual-bounded transport.
  3. [Appendix A vs Table 4] Appendix A states that 'DCL training drives violation rates below 3% with conditional violation magnitudes of O(10^{-3}).' This directly contradicts Table 4, where the FNO-backbone FluxNet-D has Vlb=7.09% and Mub=3.55e-3, and the ResNet variant has Vub=0.84% and Mub=0.838e-3. The text should be corrected to quote the actual observed ranges, and any aggregation or conditioning should be explained. As written, the inconsistency undermines confidence in the reported empirical claims.
minor comments (5)
  1. [§3.6 vs Appendix B] §3.6 says 'All models are trained by minimizing prediction loss ... employing pushforward training,' but Appendix B states that the convection-diffusion benchmark uses one-step training. Please clarify which models use pushforward and which do not, and why.
  2. [Figure references] Figure numbering is inconsistent: the main-text ERF figure is called Figure 4 in §4.4, while Appendix E refers to it as 'Figure 5 of the main text'; the convection-diffusion figure is Appendix Figure 5 but appears in the main text as 'Figure 5 in the appendix' with a different numbering. Renumber figures consistently.
  3. [Eq. (4)] The normalization |Ω| in the DCL term is not defined in the main text; it should be stated that Ω is the set of grid cells.
  4. [Prop. 2 / Prop. 3 statements] The main text states the L-head guarantee as 'ut+1_i ≥ ℓ' while the Appendix proof gives the stricter 'ut+1_i > ℓ'; similarly for U-head. Please unify the inequality conventions.
  5. [Table 4 header] The header 'Mlb Mub' could be confused with 'mean lower/upper violation'; consider labeling them as conditional violation magnitudes (Mlb, Mub) explicitly in the caption.

Circularity Check

0 steps flagged

No circularity: conservation and one-sided bounds are structural theorems; D-head dual-bound behavior is explicitly empirical.

full rationale

FluxNet's derivation is self-contained. Proposition 1 (discrete conservation) is a structural identity of the symmetric-stencil update (Eq. 2), proven by relabeling flux sums in Appendix A; it holds for any flux field and does not depend on learned parameters or fitted values. Propositions 2 and 3 are direct consequences of parameterizing outflow as a fraction of available amount a_i = u_i - ell and inflow as a fraction of remaining capacity b_i = u_max - u_i; they are theorems about the head definitions themselves, not predictions smuggled in from data. The D-head's averaged update (Eq. 3) is explicitly conceded in §3.4 and Appendix A: "the averaged update does not theoretically guarantee satisfaction of both bounds unless the two branches agree exactly." Violation rates are reported as transparent empirical metrics (Tables 4 and 7), rather than as fitted parameters renamed as predictions. There are no load-bearing self-citations: the references to FINN, FNO, LBM, positivity-preserving schemes, and pushforward training are contextual or baseline-related, and no uniqueness theorem from the authors is invoked. The main limitation—dual-bound behavior depends on DCL-induced branch agreement and can degrade out of distribution (e.g., 7.09% upper-bound violation rate for the FNO-backbone traffic-flow variant)—is a robustness/correctness concern about empirical validation, not a circularity of the derivation. The absence of explicit violation statistics for spinodal decomposition is a reporting gap, not an input-output circularity. Overall, the central guarantees are proven structural properties of the update rule and parameterization, and the paper is evaluated against external benchmarks, so the circularity burden is not met.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The method introduces no new physical entities. The free parameters are mostly training/config hyperparameters and the stencil radius. The most consequential item is the D-head's empirical agreement premise, which is acknowledged by the authors as a limitation.

free parameters (5)
  • transport neighborhood radius r = 3×3, 5×5, 9×9 stencils for spinodal 10/100/1000Δt; 11-point stencil for traffic; 3×3 for shallow water
    Chosen by hand to cover the physical domain of dependence at the target timestep; this is the practical replacement for a CFL condition and is tuned per experiment.
  • DCL weight γ = 0.1 (traffic), 1.0 (spinodal)
    Controls how strongly the two D-head branches must agree; violation rates and accuracy depend on it.
  • momentum smoothing ε = small positive constant, exact value not specified
    Added in the shallow-water advection term F^m,adv = F^h · m/(h+ε) to avoid division by zero in dry cells; changes the near-dry momentum transport.
  • pushforward unroll length K = 5 for shallow water; unspecified for other benchmarks
    Training hyperparameter; ablations show it is critical for traffic and less important for shallow water.
  • architecture capacities = base channels 32/64; residual blocks 4/6; kernel sizes 3/5/7
    Per-benchmark hand choices for the ResNet backbone; not derived from the conservation or bound constraints.
axioms (5)
  • domain assumption The PDE dynamics are source-free and conservative on a periodic grid
    §3.1–3.2 restrict the surrogate to source-free conservation laws; §5 lists source terms as future work.
  • domain assumption A finite local stencil can represent the integrated transport over the full surrogate timestep
    The update only exchanges mass within N(i). If the domain of dependence exceeds the stencil, the surrogate is inaccurate or unstable; the paper tunes r per timestep rather than proving a bound.
  • standard math Sigmoid outputs lie in (0,1) and softmax weights are nonnegative and sum to 1
    Used in the L/U bound proofs; in finite precision sigmoid can saturate, so the strict inequalities are only approximate in implementation.
  • domain assumption Numerical solver reference data are accurate enough to serve as ground truth
    All metrics compare against spectral, finite-volume, or explicit finite-difference solvers; no error quantification of the reference data is provided.
  • ad hoc to paper DCL training will make the two D-head branches agree closely enough that the averaged update respects both bounds acceptably
    §3.4 and Appendix A explicitly state the averaged update has no theoretical dual-bound guarantee; the central dual-bound results depend on this empirical premise.

pith-pipeline@v1.3.0-alltime-deepseek · 20106 in / 13527 out tokens · 130226 ms · 2026-08-03T05:29:56.488584+00:00 · methodology

0 comments
read the original abstract

Autoregressive learning of time-stepping operators provides an effective approach to data-driven partial differential equation (PDE) simulation, yet for conservation laws, they face a fundamental challenge: learned updates may violate global conservation over long rollouts. For the important subclass of mass-conservation-type equations, the problem is compounded by inherent physical bounds (e.g., nonnegativity or concentrations in [0,1]) whose violation further destabilizes predictions. We introduce FluxNet, which learns cumulative transport amounts representing the total conserved quantity redistributed between each cell and a configurable neighborhood over the full surrogate interval. A conservative update guarantees exact discrete conservation by construction; modular capacity-constrained transport heads (L, U, and D) enforce lower bounds, upper bounds, or near-zero dual-bound violations through architectural design. Unlike flux-rate surrogates that require temporal integration and thus inherit CFL constraints, FluxNet involves no such integration; configurable transport neighborhoods enable large-timestep prediction at full spatial resolution. Ghost cells extend the framework to non-periodic boundaries. Experiments on four benchmarks (1D convection--diffusion, 2D shallow water, 1D traffic flow, 2D Cahn--Hilliard) demonstrate exact conservation, structural bound preservation, architecture modularity, and superior stability over flux-rate surrogates at large temporal strides. The code is publicly available at: https://github.com/Lan-zs/FluxNet.

Figures

Figures reproduced from arXiv: 2602.01941 by Jincheng Wang, Junjie Li, Lei Wang, Zishuo Lan.

Figure 1
Figure 1. Figure 1: Rollout MAE over time for shallow water equations. Mean Absolute Error as a function of rollout time step for FluxNet￾LAP, FNO, and CNN on the shallow water equation test set. The key physical constraint is nonnegative depth (h ≥ 0), and stable rollouts are challenging due to dry regions and strong mass-momentum coupling. Qualitative comparisons at the final rollout time (T = 2, corresponding to 2× tem￾por… view at source ↗
Figure 2
Figure 2. Figure 2: Traffic flow density profiles at selected time points. Comparison of density ρ(x, t) predictions between ground truth, FluxNet-D, FNO, and CNN at three time points (t = 0, t = 25, t = 50) [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Effective receptive field analysis for FluxNet-D (10∆t). Visualization of the ERF for each output channel. The localized and anisotropic patterns confirm that the learned flux operators depend only on local neighborhood information. ral networks. Advances in Neural Information Processing Systems, 28, 2015. Brandstetter, J., Worrall, D., and Welling, M. Message pass￾ing neural PDE solvers. arXiv preprint ar… view at source ↗
Figure 5
Figure 5. Figure 5: Concentration profiles at five time points comparing FluxNet-N (unconstrained), FluxNet-P (positive flux), and FluxNet-L (lower-bounded) variants on the 1D convection-diffusion equation. All three variants maintain good agreement with ground truth in smooth regions. The L-head produces the most accurate tracking near concentration minima approaching zero. accuracy with virtually no deviation, the P-head sh… view at source ↗
Figure 6
Figure 6. Figure 6: Time evolution of the water depth field h(x, y, t) comparing ground truth, FluxNet-LAP, FNO, and CNN at six time steps spanning the rollout horizon. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Time evolution of the x-momentum field mx(x, y, t) comparing ground truth, FluxNet-LAP, FNO, and CNN at six time steps. FluxNet-LAP maintains accurate predictions throughout, while CNN develops high-frequency artifacts. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Time evolution of the y-momentum field my(x, y, t) comparing ground truth, FluxNet-LAP, FNO, and CNN at six time steps. Similar patterns to mx are observed, with FluxNet-LAP providing the most accurate predictions. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Phase field ϕ(x, y, t) evolution comparing ground truth and FluxNet-D models trained with different time step sizes (10∆t, 100∆t, 1000∆t) at four time points (0T, 1T, 1.5T, 2T). The rightmost column shows absolute prediction error at the final time 2T. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Radial two-point correlation function S¯2(r) comparison between ground truth (phase-field simulation) and FluxNet-D models at four time points: 0T (initial), 1T (training horizon), 1.5T, and 2T (2× extrapolation). All models closely match the reference correlation functions throughout the extrapolation regime. 0 25 50 75 100 Time (×10 3Δt) 46 48 50 52 54 Phase Volume Fraction (%) FluxNet-D (10Δt) 0 25 50 … view at source ↗
Figure 11
Figure 11. Figure 11 [PITH_FULL_IMAGE:figures/full_fig_p021_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Effective receptive field (ERF) analysis for FluxNet-D trained with 100∆t time steps. Left panel: outflow branch channels arranged according to spatial offset indices within the 5×5 neighborhood. Right panel: inflow branch channels. The larger ERF compared to the 10∆t model ( [PITH_FULL_IMAGE:figures/full_fig_p022_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Effective receptive field (ERF) analysis for FluxNet-D trained with 1000∆t time steps. Left panel: outflow branch channels within the 9 × 9 neighborhood. Right panel: inflow branch channels. This model exhibits the largest ERF among the three variants, consistent with the extended spatial dependencies required for accurate prediction at the coarsest temporal resolution. 23 [PITH_FULL_IMAGE:figures/full_f… view at source ↗
Figure 14
Figure 14. Figure 14: Wall-clock time comparison for completing 50,000∆t of spinodal decomposition simulation. Comparison between the GPU-accelerated phase-field solver and FluxNet-D models at three temporal resolutions across system sizes from 2562 to 40962 grid points. Both axes use logarithmic scales. The 1000∆t model achieves 17.3× speedup at 10242 resolution, with efficiency gains increasing at larger domain sizes. 24 [P… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

33 extracted references · 7 linked inside Pith

  1. [1]

    W., Hamrick, J

    Battaglia, P. W., Hamrick, J. B., Bapst, V., Sanchez-Gonzalez, A., Zambaldi, V., Malinowski, M., Tacchetti, A., Raposo, D., Santoro, A., Faulkner, R., et al. Relational inductive biases, deep learning, and graph networks. arXiv preprint arXiv:1806.01261, 2018

    Pith/arXiv arXiv 2018

  2. [2]

    Scheduled sampling for sequence prediction with recurrent neural networks

    Bengio, S., Vinyals, O., Jaitly, N., and Shazeer, N. Scheduled sampling for sequence prediction with recurrent neural networks. Advances in Neural Information Processing Systems, 28, 2015

    2015

  3. [3]

    Message passing neural PDE solvers

    Brandstetter, J., Worrall, D., and Welling, M. Message passing neural PDE solvers. arXiv preprint arXiv:2202.03376, 2022

    Pith/arXiv arXiv 2022

  4. [4]

    Cahn, J. W. and Hilliard, J. E. Free energy of a nonuniform system. I . interfacial free energy. The Journal of Chemical Physics, 28 0 (2): 0 258--267, 1958

    1958

  5. [5]

    Phase-field models for microstructure evolution

    Chen, L.-Q. Phase-field models for microstructure evolution. Annual Review of Materials Research, 32 0 (1): 0 113--140, 2002

    2002

  6. [6]

    S., Riley, P

    Gilmer, J., Schoenholz, S. S., Riley, P. F., Vinyals, O., and Dahl, G. E. Neural message passing for quantum chemistry. In Proceedings of the International Conference on Machine Learning, pp.\ 1263--1272. PMLR, 2017

    2017

  7. [7]

    Gupta, J. K. and Brandstetter, J. Towards multi-spatiotemporal-scale generalized PDE modeling. arXiv preprint arXiv:2209.15616, 2022

    Pith/arXiv arXiv 2022

  8. [8]

    Deep residual learning for image recognition

    He, K., Zhang, X., Ren, S., and Sun, J. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pp.\ 770--778, 2016

    2016

  9. [9]

    and Mitsume, N

    Horie, M. and Mitsume, N. Physics-embedded neural networks: Graph neural PDE solvers with mixed boundary conditions. Advances in Neural Information Processing Systems, 35: 0 23218--23229, 2022

    2022

  10. [10]

    Karlbauer, M., Praditia, T., Otte, S., Oladyshkin, S., Nowak, W., and Butz, M. V. Composing partial differential equations with physics-aware neural networks. In Proceedings of the International Conference on Machine Learning, pp.\ 10773--10801. PMLR, 2022

    2022

  11. [11]

    A., Alieva, A., Wang, Q., Brenner, M

    Kochkov, D., Smith, J. A., Alieva, A., Wang, Q., Brenner, M. P., and Hoyer, S. Machine learning--accelerated computational fluid dynamics. Proceedings of the National Academy of Sciences, 118 0 (21): 0 e2101784118, 2021

    2021

  12. [12]

    LeVeque, R. J. Finite Volume Methods for Hyperbolic Problems, volume 31. Cambridge University Press, 2002

    2002

  13. [13]

    F ourier neural operator for parametric partial differential equations

    Li, Z., Kovachki, N., Azizzadenesheli, K., Liu, B., Bhattacharya, K., Stuart, A., and Anandkumar, A. F ourier neural operator for parametric partial differential equations. arXiv preprint arXiv:2010.08895, 2020

    Pith/arXiv arXiv 2010

  14. [14]

    Learning chaotic dynamics in dissipative systems

    Li, Z., Liu-Schiaffini, M., Kovachki, N., Azizzadenesheli, K., Liu, B., Bhattacharya, K., Stuart, A., and Anandkumar, A. Learning chaotic dynamics in dissipative systems. Advances in Neural Information Processing Systems, 35: 0 16768--16781, 2022

    2022

  15. [15]

    Lighthill, M. J. and Whitham, G. B. On kinematic waves II . a theory of traffic flow on long crowded roads. Proceedings of the Royal Society of London. Series A. Mathematical and Physical Sciences, 229 0 (1178): 0 317--345, 1955

    1955

  16. [16]

    PDE -refiner: Achieving accurate long rollouts with neural PDE solvers

    Lippe, P., Veeling, B., Perdikaris, P., Turner, R., and Brandstetter, J. PDE -refiner: Achieving accurate long rollouts with neural PDE solvers. Advances in Neural Information Processing Systems, 36: 0 67398--67433, 2023

    2023

  17. [17]

    Neural entropy-stable conservative flux form neural networks for learning hyperbolic conservation laws

    Liu, L., Zhang, L., and Gelb, A. Neural entropy-stable conservative flux form neural networks for learning hyperbolic conservation laws. arXiv preprint arXiv:2507.01795, 2025

    Pith/arXiv arXiv 2025

  18. [18]

    and Hutter, F

    Loshchilov, I. and Hutter, F. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017

    Pith/arXiv arXiv 2017

  19. [19]

    Understanding the effective receptive field in deep convolutional neural networks

    Luo, W., Li, Y., Urtasun, R., and Zemel, R. Understanding the effective receptive field in deep convolutional neural networks. Advances in Neural Information Processing Systems, 29, 2016

    2016

  20. [20]

    Deep energy-based modeling of discrete-time physics

    Matsubara, T., Ishikawa, A., and Yaguchi, T. Deep energy-based modeling of discrete-time physics. Advances in Neural Information Processing Systems, 33: 0 13100--13111, 2020

    2020

  21. [21]

    Learning mesh-based simulation with graph networks

    Pfaff, T., Fortunato, M., Sanchez-Gonzalez, A., and Battaglia, P. Learning mesh-based simulation with graph networks. In Proceedings of the International Conference on Learning Representations, 2020

    2020

  22. [22]

    V., and Nowak, W

    Praditia, T., Karlbauer, M., Otte, S., Oladyshkin, S., Butz, M. V., and Nowak, W. Finite volume neural network: Modeling subsurface contaminant transport. arXiv preprint arXiv:2104.06010, 2021

    Pith/arXiv arXiv 2021

  23. [23]

    Raissi, M., Perdikaris, P., and Karniadakis, G. E. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics, 378: 0 686--707, 2019

    2019

  24. [24]

    Richards, P. I. Shock waves on the highway. Operations Research, 4 0 (1): 0 42--51, 1956

    1956

  25. [25]

    A reduction of imitation learning and structured prediction to no-regret online learning

    Ross, S., Gordon, G., and Bagnell, D. A reduction of imitation learning and structured prediction to no-regret online learning. In Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, pp.\ 627--635. JMLR Workshop and Conference Proceedings, 2011

    2011

  26. [26]

    Learning to simulate complex physics with graph networks

    Sanchez-Gonzalez, A., Godwin, J., Pfaff, T., Ying, R., Leskovec, J., and Battaglia, P. Learning to simulate complex physics with graph networks. In Proceedings of the International Conference on Machine Learning, pp.\ 8459--8468. PMLR, 2020

    2020

  27. [27]

    Learning under laws: A constraint-projected neural PDE solver that eliminates hallucinations

    Singha, M. Learning under laws: A constraint-projected neural PDE solver that eliminates hallucinations. arXiv preprint arXiv:2511.03578, 2025

    arXiv 2025

  28. [28]

    The lattice B oltzmann method: Principles and practice

    Timm, K., Kusumaatmaja, H., Kuzmin, A., Shardt, O., Silva, G., and Viggen, E. The lattice B oltzmann method: Principles and practice. Cham, Switzerland: Springer International Publishing AG, 2016

    2016

  29. [29]

    Torquato, S. et al. Random Heterogeneous Materials: Microstructure and Macroscopic Properties, volume 16. Springer, 2002

    2002

  30. [30]

    Towards the ultimate conservative difference scheme

    Van Leer, B. Towards the ultimate conservative difference scheme. V . a second-order sequel to G odunov's method. Journal of Computational Physics, 32 0 (1): 0 101--136, 1979

    1979

  31. [31]

    Learning the solution operator of parametric partial differential equations with physics-informed DeepONets

    Wang, S., Wang, H., and Perdikaris, P. Learning the solution operator of parametric partial differential equations with physics-informed DeepONets . Science Advances, 7 0 (40): 0 eabi8605, 2021

    2021

  32. [32]

    and Shu, C.-W

    Zhang, X. and Shu, C.-W. On positivity-preserving high order discontinuous G alerkin schemes for compressible E uler equations on rectangular meshes. Journal of Computational Physics, 229 0 (23): 0 8918--8934, 2010

    2010

  33. [33]

    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 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...