Macroscopic Fluid Models for Network Traffic Control

37 concepts34 papers

Imagine trying to optimize the routes, travel costs, and emissions of millions of vehicles across a metropolitan area. If you attempt to track every individual vehicle—its precise acceleration, lane changes, and reaction times—the computational complexity scales directly with the number of agents. In large-scale networks, this microscopic approach becomes computationally intractable for real-time control and highly sensitive to minor structural errors, often resulting in artificial gridlocks during simulation.

To solve this, a powerful paradigm shifts the perspective from individual agents to aggregate flows. By treating traffic as a continuous, compressible fluid, researchers can model traffic density, velocity, and flow using partial differential equations (PDEs). This macroscopic abstraction makes the complexity of optimization independent of the number of vehicles, enabling city-wide traffic management, fleet rebalancing, and dynamic pricing.

The Fluid Analogy and Conservation Laws

The foundation of macroscopic traffic modeling relies on the analogy between a stream of vehicles and fluid dynamics. Instead of tracking discrete coordinates, the state of a road network is defined by continuous fields: density $\rho(x,t)$ (vehicles per unit length), velocity $v(x,t)$, and flow $q(x,t)$ (vehicles per unit time), bound by the fundamental identity $q = \rho v$.

Because vehicles are neither created nor destroyed along a homogeneous road segment, the system must satisfy a hyperbolic conservation law:

$$\frac{\partial \rho}{\partial t} + \frac{\partial q}{\partial x} = 0$$

This formulation, famously known as the Lighthill-Whitham-Richards (LWR) model, allows researchers to apply classical fluid mechanics to transportation networks. For instance, Legesse Lemecha Obsu et al. use scalar hyperbolic conservation laws to optimize traffic through complex junctions like roundabouts by dynamically adjusting right-of-way priority parameters. To evaluate how these continuous streams split and merge at intersections, Samuel Hammond et al. utilize Riemann solvers within Godunov-type numerical schemes, treating traffic bottlenecks as wave propagation problems.

By modeling traffic as a continuous fluid field, engineers can design active management strategies—such as variable speed limits—to smoothly regulate outflow and dissipate congestion waves, as demonstrated by Maria Laura Delle Monache et al..

Smoothing Shocks with Hamilton-Jacobi Formulations

While standard fluid models are highly efficient, they possess a major mathematical drawback: shock waves. In traffic, a shock wave manifests as a sharp, discontinuous boundary between free-flowing traffic and a bumper-to-bumper jam. Mathematically, these discontinuities make classical differentiation impossible, complicating gradient-based optimization and control.

To circumvent this, researchers "lift" the conservation laws into a higher-dimensional, smoother representation using cumulative vehicle counts, $M(x,t)$, which measure the total number of vehicles that have passed a point $x$ by time $t$. Because density is the spatial derivative of cumulative count ($\rho = -\partial M / \partial x$) and flow is the temporal derivative ($q = \partial M / \partial t$), the conservation law can be rewritten as a Hamilton-Jacobi equation.

Microscopic:  [Car 1] -> [Car 2] -> [Car 3]  (Scales with N)
                               │
                               ▼ (Spatial Aggregation)
Macroscopic:  Density Field ρ(x,t)           (Discontinuous Shocks)
                               │
                               ▼ (Integration)
Integrated:   Cumulative Count M(x,t)        (Lipschitz Continuous & Smooth)

This transformation fundamentally changes the mathematical behavior of the problem. As Nicolas Laurent-Brouty et al. demonstrate, while density models suffer from discontinuous shocks, cumulative vehicle functions are Lipschitz continuous. This continuity guarantees that the mathematical models remain well-posed and stable, even when accounting for finite buffers and traffic spillback at junctions.

Furthermore, John Lygeros et al. show that traffic systems modeled via cumulative vehicle counts exhibit state-monotone dynamics. This structural property allows complex, non-convex freeway control problems to be solved exactly using efficient convex relaxation techniques, ensuring that the mathematically simplified solutions remain physically valid in real-world environments.

Scaling Fleet Optimization Independent of Agent Count

The primary practical advantage of fluidic abstraction is that the computational cost of optimization no longer depends on fleet size. This scale-invariance is crucial for managing Autonomous Mobility-on-Demand (AMoD) systems and large-scale ride-hailing platforms.

If an operator tries to route and match thousands of autonomous taxis individually, the resulting mixed-integer linear program quickly becomes intractable. However, by employing a mesoscopic fluidic approximation, the fleet can be modeled as a continuous flow moving through a time-expanded network. Ramón Iglesias et al. leverage this abstraction to preemptively rebalance empty vehicles toward forecasted demand, reducing passenger wait times by up to 90% while keeping optimization complexity independent of the total number of vehicles.

This fluidic framing allows operators to co-optimize entirely different infrastructure layers. For example, Alvaro Estandia et al. couple a fluidic vehicle routing model with an optimal power flow framework, making the joint optimization of electric vehicle fleets and power distribution networks scalable.

Similarly, Ioannis Paschalidis et al. integrate demand management (pricing) and supply management (rebalancing) into a single continuous flow network, maximizing system profits by treating customer arrivals and vehicle redistributions as balanced fluid flows.

Siloed Optimization:   [Pricing Engine] ──(Indirect Signal)──> [Rebalancing Engine]

Integrated Fluid Flow: [Customer Flow] <───(Coupled State)───> [Vehicle Fleet Flow]

Regional Accumulation and Data-Driven Control

To scale control to entire cities, researchers often abstract individual roads into broader geographic zones. Instead of modeling every intersection, a city can be partitioned into homogeneous regions where traffic behavior is governed by an aggregate relationship known as the Macroscopic Fundamental Diagram (MFD), which relates the total number of vehicles in a region (accumulation) to the region's structural throughput.

This regional abstraction simplifies high-dimensional network dynamics into low-dimensional accumulation tracking. Alessio Rimoldi et al. combine this regional MFD framework with Data-enabled Predictive Control (DeePC). Because the traffic state is reduced to aggregate regional densities, the authors can control urban congestion using a purely data-driven framework, bypassing the need to derive explicit, highly complex physical model equations for every street.

For simulation and machine learning testbeds, regional modeling eliminates the computational bottlenecks of traditional micro-simulators. Zijian Hu et al. developed a hybrid simulator that pairs link-based flow models with regional "Bathtub" models. By determining vehicle speeds from aggregate regional density rather than individual link capacities, their simulator can model city-wide traffic involving millions of vehicles in under a minute, providing a high-speed environment for training multi-agent reinforcement learning agents without triggering artificial gridlocks.

Bridging the Micro-Macro Divide

While macroscopic models excel at scale, traffic is ultimately driven by discrete agents. A key focus of modern research is understanding how microscopic behaviors manifest as macroscopic fluid properties.

To bridge this gap, researchers use kinetic theory—similar to statistical mechanics in physics—to derive fluid equations from driver interaction rules. Giuseppe Visconti et al. apply a Fokker-Planck approximation to stochastic driver behaviors, demonstrating that macroscopic phenomena, such as the scattered "multivalued" diagrams observed in real highway sensor data, naturally emerge from microscopic randomness.

Microscopic Stochasticity ──(Fokker-Planck Limit)──> Macroscopic Fluid Properties
(Individual Driver Rules)                            (Scattered Flow-Density Curves)

This multi-scale understanding allows for the design of targeted interventions where a small number of controlled agents influence the global fluid flow. For instance, Han Wang et al. demonstrate that a single autonomous vehicle acting as a traffic actuator can dissipate stop-and-go waves across an entire highway segment.

When explicit multi-scale modeling is required, Sanghyun Son et al. provide a solution via a differentiable hybrid traffic simulator. By coupling anisotropic fluid models (where conditions ahead dictate flow) with agent-based tracking and deriving analytical gradients for the entire pipeline, their framework enables rapid, gradient-based optimization for urban planning and network control.

Key Foundational Work

Dynamic Congestion and Tolls with Mobile Source Emission (2012) — contributed 1 concept elements to this direction.
Simulating smart cities with DEUS (2012) — contributed 1 concept elements to this direction.
Numerical approximation of continuous traffic congestion equilibria (2009) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_macroscopic_traffic_abstraction_58c750d3,
  title = {macroscopic traffic abstraction},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/macroscopic-traffic-abstraction/rdr_58c750d32f6742369eacc583af250f09},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_macroscopic_traffic_abstraction_58c750d3_v1,
  title = {macroscopic traffic abstraction},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/macroscopic-traffic-abstraction/rdr_58c750d32f6742369eacc583af250f09/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}

See something wrong? Report a mistake