pith. sign in

arxiv: 2604.07918 · v1 · submitted 2026-04-09 · 📡 eess.SY · cs.SY

Second Order Physics-Informed Learning of Road Density using Probe Vehicles

S. Betancur Giraldo , J. M{\aa}rtensson , M. Barreau This is my paper

Pith reviewed 2026-05-10 18:16 UTC · model grok-4.3

classification 📡 eess.SY cs.SY
keywords traffic density reconstructionphysics-informed learningsecond-order traffic modelprobe vehicle dataAw-Rascle-Zhang modelnonequilibrium trafficSUMO simulation
0
0 comments X

The pith

A second-order physics-informed learning method reconstructs road traffic density from sparse probe vehicle trajectories more accurately than first-order models, especially in changing conditions.

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

The paper develops a physics-informed learning framework that estimates traffic density along a road using position and speed reports from only a few probe vehicles. It pairs a first-order model to learn the equilibrium speed-density curve with a second-order Aw-Rascle-Zhang model that enforces additional dynamic relations between density, velocity, and relaxation to equilibrium. Tests on simulated traffic that includes both steady flow and sudden disturbances show the second-order version produces lower error and stays stable where simpler models break down. This matters because reliable density estimates from cheap, mobile sensors can support real-time traffic management without installing dense fixed detectors everywhere.

Core claim

The central claim is that a physics-informed neural network built around the second-order Aw-Rascle-Zhang traffic model, after an initial first-order stage that estimates equilibrium velocity, yields more accurate and robust reconstructions of density from sparse probe trajectories than first-order physics-informed approaches, with the advantage most pronounced in transient nonequilibrium regimes even though the equilibrium-velocity input can destabilize the learning process when traffic deviates strongly from steady state.

What carries the argument

A two-stage physics-informed neural network that first trains an equilibrium velocity function from a first-order traffic model and then embeds the second-order Aw-Rascle-Zhang equations to constrain density estimation from probe data.

If this is right

  • Density estimates remain accurate when traffic undergoes sudden changes rather than staying in steady flow.
  • The method works with only sparse trajectory data instead of requiring dense fixed sensors.
  • Higher-order traffic models can be embedded directly into learning pipelines to capture wave-like behavior that first-order models miss.
  • Pre-training the equilibrium velocity improves steady-state results but must be handled carefully to avoid instability in transients.

Where Pith is reading between the lines

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

  • The same two-stage structure could be tested on real GPS traces from connected vehicles to check whether simulation gains carry over to field data.
  • Because the second-order equations already encode relaxation dynamics, the framework might extend naturally to short-term prediction of density evolution rather than pure reconstruction.
  • Replacing the neural network with other function approximators while keeping the same physics constraints would isolate how much of the improvement comes from the model order versus the learning architecture.

Load-bearing premise

That the equilibrium velocity function learned in the first-order stage remains a valid and stable input for the second-order model even when traffic is far from equilibrium and rapidly changing.

What would settle it

In the same SUMO simulation scenarios, if the density reconstruction error of the second-order model exceeds the error of the first-order model during transient periods, the claim of greater robustness would be falsified.

Figures

Figures reproduced from arXiv: 2604.07918 by J. M{\aa}rtensson, M. Barreau, S. Betancur Giraldo.

Figure 1
Figure 1. Figure 1: Comparison of traffic density evolving in time and space between SUMO simulation (left) and reconstruction from trajectory data using LWR [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of density vs velocity for trajectories data points taken [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of traffic density evolving in time and space between SUMO simulation (left) and reconstruction from trajectory data using LWR [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Distribution of L2 error on ρ (top) and L2 error on v (bottom). learned). The boxplots reveal two key aspects of the learning behavior: accuracy and variability across runs. In the equilibrium regime, all model configurations exhibit relatively small variability, indicating stable training dynam￾ics. The differences between LWR and ARZ are moderate, and the boxplots do not indicate a uniform advantage of o… view at source ↗
read the original abstract

We propose a Physics Informed Learning framework for reconstructing traffic density from sparse trajectory data. The approach combines a second-order Aw-Rascle and Zhang model with a first-order training stage to estimate the equilibrium velocity. The method is evaluated in both equilibrium and transient traffic regimes using SUMO simulations. Results show that while learning the equilibrium velocity improves reconstruction under steady state conditions, it becomes unstable in transient regimes due to the breakdown of the equilibrium assumption. In contrast, the second-order model consistently provides more accurate and robust reconstructions than first-order approaches, particularly in nonequilibrium conditions.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper proposes a physics-informed learning framework that reconstructs traffic density from sparse probe-vehicle trajectories by combining a second-order Aw-Rascle-Zhang (ARZ) model with a preliminary first-order training stage that estimates equilibrium velocity. The method is evaluated on SUMO simulations in both equilibrium and transient regimes; the abstract reports that learning the equilibrium velocity improves steady-state reconstructions but produces instability in transients due to breakdown of the equilibrium assumption, while claiming that the second-order model nonetheless yields more accurate and robust density estimates than first-order baselines, especially under nonequilibrium conditions.

Significance. If the tension between reported instability in transients and the claimed robustness advantage in nonequilibrium traffic can be resolved with quantitative evidence, the work would offer a concrete demonstration of how higher-order traffic PDEs can be embedded in data-driven reconstruction pipelines, potentially improving handling of stop-and-go waves and other non-steady phenomena beyond what first-order models achieve.

major comments (2)
  1. [Abstract] Abstract: the claim that 'the second-order model consistently provides more accurate and robust reconstructions than first-order approaches, particularly in nonequilibrium conditions' is load-bearing for the central contribution, yet the same paragraph states that the learned equilibrium velocity 'becomes unstable in transient regimes due to the breakdown of the equilibrium assumption.' Transient flow is the canonical nonequilibrium case; the manuscript must therefore supply regime-specific error metrics (e.g., RMSE or density error histograms) that isolate the second-order contribution from the first-order equilibrium-velocity estimate and show where the advantage persists.
  2. [Evaluation on SUMO simulations] Evaluation section (SUMO experiments): because the equilibrium velocity learned in the first-order stage is an explicit input to the second-order ARZ PDE, any instability observed in transients directly conditions the second-order results. The paper should include an ablation that replaces the learned equilibrium velocity with a fixed literature value or with ground-truth equilibrium speed to quantify how much of the reported robustness is attributable to the second-order dynamics versus the quality of the first-stage estimate.
minor comments (1)
  1. [Abstract] The abstract would benefit from explicit numerical values (e.g., percentage error reduction or R² scores) rather than qualitative statements of 'more accurate and robust.'

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the valuable feedback. We address the major comments point-by-point below and will make the necessary revisions to the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim that 'the second-order model consistently provides more accurate and robust reconstructions than first-order approaches, particularly in nonequilibrium conditions' is load-bearing for the central contribution, yet the same paragraph states that the learned equilibrium velocity 'becomes unstable in transient regimes due to the breakdown of the equilibrium assumption.' Transient flow is the canonical nonequilibrium case; the manuscript must therefore supply regime-specific error metrics (e.g., RMSE or density error histograms) that isolate the second-order contribution from the first-order equilibrium-velocity estimate and show where the advantage persists.

    Authors: We acknowledge the potential confusion in the abstract. The instability is specific to the learned equilibrium velocity in transient regimes, but the second-order ARZ model still offers advantages in modeling non-equilibrium dynamics like stop-and-go waves. To resolve this, we will revise the abstract for clarity and add regime-specific quantitative metrics in the evaluation section, including separate RMSE values and error distributions for equilibrium and transient regimes. This will isolate the second-order contribution and demonstrate its robustness where the equilibrium assumption holds or partially holds. revision: yes

  2. Referee: [Evaluation on SUMO simulations] Evaluation section (SUMO experiments): because the equilibrium velocity learned in the first-order stage is an explicit input to the second-order ARZ PDE, any instability observed in transients directly conditions the second-order results. The paper should include an ablation that replaces the learned equilibrium velocity with a fixed literature value or with ground-truth equilibrium speed to quantify how much of the reported robustness is attributable to the second-order dynamics versus the quality of the first-stage estimate.

    Authors: We agree this ablation is important to clarify the sources of performance. We will include an additional set of experiments in the revised manuscript where the equilibrium velocity is set to a fixed literature value (such as the free-flow speed from the simulation setup) and to the ground-truth equilibrium speed derived from the SUMO data. We will report the resulting density reconstruction errors and compare them to the learned-velocity case and the first-order baselines. This will help attribute the robustness to the second-order physics. revision: yes

Circularity Check

0 steps flagged

No significant circularity; two-stage learning is data-driven with external simulation benchmarks

full rationale

The paper describes a two-stage Physics-Informed Learning process: a first-order stage fits the equilibrium velocity from trajectory data, which is then supplied as an input parameter to the second-order Aw-Rascle-Zhang PDE for density reconstruction. This is an explicit modeling choice and data-driven fitting step rather than a derivation that reduces the output (density field) to the input by algebraic construction or renaming. The abstract explicitly flags the equilibrium assumption's breakdown in transient regimes and reports comparative accuracy on SUMO simulations, which constitute an external benchmark independent of the fitted value. No self-citation chains, uniqueness theorems, or ansatzes imported from prior author work are invoked to force the result. The central claim of improved robustness therefore rests on empirical evaluation rather than tautology.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The central claim rests on the validity of the Aw-Rascle-Zhang second-order PDE system and on the assumption that a separately learned equilibrium velocity can be treated as a fixed input. No new entities are postulated.

free parameters (1)
  • equilibrium velocity
    Estimated in the first-order training stage from trajectory data; directly affects the second-order reconstruction and is the source of instability in transient regimes.
axioms (2)
  • domain assumption The Aw-Rascle-Zhang model equations accurately describe both equilibrium and non-equilibrium traffic flow.
    Invoked when the second-order model is used for reconstruction; the abstract notes breakdown when equilibrium assumption fails.
  • domain assumption Sparse probe trajectories contain sufficient information to reconstruct density when the physics model is enforced.
    Core premise of the physics-informed learning framework.

pith-pipeline@v0.9.0 · 5397 in / 1504 out tokens · 105615 ms · 2026-05-10T18:16:10.664733+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

23 extracted references · 23 canonical work pages

  1. [1]

    Ferrara, S

    A. Ferrara, S. Sacone, and S. Siri,Freeway traffic modelling and control. Springer, 2018

    work page 2018

  2. [2]

    Experimental features of the emergence of moving jams in free traffic flow,

    B. S. Kerner, “Experimental features of the emergence of moving jams in free traffic flow,”Journal of Physics A: Mathematical and general, vol. 33, no. 26, pp. L221–L228, 2000

    work page 2000

  3. [3]

    Review of road traffic control strategies,

    M. Papageorgiou, C. Diakaki, V . Dinopoulou, A. Kotsialos, and Y . Wang, “Review of road traffic control strategies,”Proceedings of the IEEE, vol. 91, no. 12, pp. 2043–2067, 2003

    work page 2043

  4. [4]

    Traffic state estimation on highway: A comprehensive survey,

    T. Seo, A. M. Bayen, T. Kusakabe, and Y . Asakura, “Traffic state estimation on highway: A comprehensive survey,”Annual Reviews in Control, vol. 43, pp. 128–151, 2017

    work page 2017

  5. [5]

    The cell transmission model: A dynamic represen- tation of highway traffic consistent with the hydrodynamic theory,

    C. F. Daganzo, “The cell transmission model: A dynamic represen- tation of highway traffic consistent with the hydrodynamic theory,” Transportation Research Part B: Methodological, vol. 28, no. 4, pp. 269–287, 1994

    work page 1994

  6. [6]

    Real-time freeway traffic state estimation based on extended kalman filter: Adaptive capabilities and real data testing,

    Y . Wang, M. Papageorgiou, and A. Messmer, “Real-time freeway traffic state estimation based on extended kalman filter: Adaptive capabilities and real data testing,”Transportation Research Part A: Policy and Practice, vol. 42, no. 10, pp. 1340–1358, 2008

    work page 2008

  7. [7]

    Incorporation of lagrangian mea- surements in freeway traffic state estimation,

    J. C. Herrera and A. M. Bayen, “Incorporation of lagrangian mea- surements in freeway traffic state estimation,”Transportation Research Part B: Methodological, vol. 44, no. 4, pp. 460–481, 2010

    work page 2010

  8. [8]

    Treiber and A

    M. Treiber and A. Kesting,Traffic flow dynamics: Data, models and Simulation. Springer, 2014

    work page 2014

  9. [9]

    On kinematic waves ii. a theory of traffic flow on long crowded roads,

    M. J. Lighthill and G. B. Whitham, “On kinematic waves ii. a theory of traffic flow on long crowded roads,”Proceedings of the Royal Society of London. A. Mathematical and Physical Sciences, vol. 229, no. 1178, pp. 317–345, 1955

    work page 1955

  10. [10]

    Shock waves on the highway,

    P. I. Richards, “Shock waves on the highway,”Operations Research, vol. 4, no. 1, pp. 42–51, 1956

    work page 1956

  11. [11]

    Physics- informed learning for identification and state reconstruction of traffic density,

    M. Barreau, M. Aguiar, J. Liu, and K. H. Johansson, “Physics- informed learning for identification and state reconstruction of traffic density,” in2021 60th IEEE Conference on Decision and Control (CDC), pp. 2653–2658, IEEE, 2021

    work page 2021

  12. [12]

    Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations,

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

    work page 2019

  13. [13]

    A control perspective on training pinns,

    M. Barreau and H. Shen, “A control perspective on training pinns,” arXiv preprint arXiv:2501.18582, 2025

    work page arXiv 2025

  14. [14]

    Closed-loop neural operator-based observer of traffic density,

    A. Harting, K. H. Johansson, and M. Barreau, “Closed-loop neural operator-based observer of traffic density,” in2025 IEEE 64th Con- ference on Decision and Control (CDC), pp. 5845–5852, IEEE, 2025

    work page 2025

  15. [15]

    A physics-informed deep learning paradigm for traffic state and fundamental diagram estimation,

    R. Shi, Z. Mo, K. Huang, X. Di, and Q. Du, “A physics-informed deep learning paradigm for traffic state and fundamental diagram estimation,”IEEE Transactions on Intelligent Transportation Systems, vol. 23, no. 8, pp. 11688–11698, 2021

    work page 2021

  16. [16]

    Resurrection of

    A. Aw and M. Rascle, “Resurrection of ”second order” models of traffic flow,”SIAM Journal on Applied Mathematics, vol. 60, no. 3, pp. 916–938, 2000

    work page 2000

  17. [17]

    A non-equilibrium traffic model devoid of gas-like behav- ior,

    H. Zhang, “A non-equilibrium traffic model devoid of gas-like behav- ior,”Transportation Research Part B: Methodological, vol. 36, no. 3, pp. 275–290, 2002

    work page 2002

  18. [18]

    A study of traffic capacity,

    B. D. Greenshields, “A study of traffic capacity,” inProceedings of the Highway Research Board, vol. 14, pp. 448–477, 1935

    work page 1935

  19. [19]

    Nonlinear follow-the- leader models of traffic flow,

    D. C. Gazis, R. Herman, and R. W. Rothery, “Nonlinear follow-the- leader models of traffic flow,”Operations research, vol. 9, no. 4, pp. 545–567, 1961

    work page 1961

  20. [20]

    Follow-the-leader models can be viewed as a numerical approximation to the lighthill-whitham-richards model for traffic flow,

    H. Holden and N. H. Risebro, “Follow-the-leader models can be viewed as a numerical approximation to the lighthill-whitham-richards model for traffic flow,”Networks and Heterogeneous Media, vol. 13, no. 3, pp. 409–421, 2018

    work page 2018

  21. [21]

    Multilayer feedforward networks are universal approximators,

    K. Hornik, M. Stinchcombe, and H. White, “Multilayer feedforward networks are universal approximators,”Neural networks, vol. 2, no. 5, pp. 359–366, 1989

    work page 1989

  22. [22]

    Microscopic traffic simulation using sumo,

    P. Alvarez Lopez, M. Behrisch, L. Bieker-Walz, J. Erdmann, Y .- P. Fl ¨otter¨od, R. Hilbrich, L. L ¨ucken, J. Rummel, P. Wagner, and E. Wiessner, “Microscopic traffic simulation using sumo,” in2018 21st International Conference on Intelligent Transportation Systems (ITSC), pp. 2575–2582, 2018

    work page 2018

  23. [23]

    On accessibility fairness in intermodal autonomous mobility-on-demand systems,

    M. Salazar, S. Betancur Giraldo, F. Paparella, and L. Pedroso, “On accessibility fairness in intermodal autonomous mobility-on-demand systems,”IFAC-PapersOnLine, vol. 58, no. 10, pp. 327–333, 2024

    work page 2024