The Divergence Field: Generating Synthetic Traces from Non-Linear Flow Models

In mobility modeling, real-world trace data is often incomplete, expensive to collect, or restricted by privacy regulations. Synthetic trace generation fills this gap, but traditional linear models fail to capture the chaotic, non-linear dynamics of real traffic flows—sudden congestion, spillback effects, and multi-modal transitions. The divergence field method addresses this by modeling flow as a tensor field with divergence constraints, enabling realistic synthetic traces that preserve statistical properties of observed systems. This guide provides a comprehensive overview of the technique, its implementation, and its limitations, based on widely shared professional practices as of May 2026.

Why Non-Linear Flow Models Matter for Synthetic Traces

Most synthetic trace generators rely on linear or piecewise-linear approximations of traffic flow. These models assume that vehicle speeds decrease proportionally with density, a relationship that breaks down under real-world conditions. For example, during a sudden incident on a highway, traffic can come to a complete stop, then gradually recover—a non-linear hysteresis loop that linear models cannot reproduce. The divergence field approach models flow as a continuous vector field where the divergence at each point represents net inflow or outflow. By solving partial differential equations with non-linear source terms, the model can generate traces that exhibit realistic stop-and-go waves, queue formation, and dissipation.

Key Limitations of Linear Models

Linear models, such as the Lighthill-Whitham-Richards (LWR) model with a triangular fundamental diagram, assume a unique relationship between density and flow. In practice, this relationship is non-linear and history-dependent. For instance, the capacity drop phenomenon—where flow after a bottleneck is lower than before—cannot be captured without non-linear terms. Practitioners often find that synthetic traces from linear models fail statistical tests for real-world patterns, such as the distribution of travel times or the frequency of congestion episodes.

Why Divergence Fields Work

A divergence field is a scalar field that defines the net rate of flow accumulation at each point in space and time. In traffic, positive divergence indicates vehicles entering a segment (e.g., from a ramp), while negative divergence indicates exiting. Non-linear flow models use these fields to introduce source and sink terms that vary with state variables like density or speed. For example, a bottleneck can be modeled as a region of negative divergence that intensifies as density increases, creating a feedback loop. This approach allows the generator to produce traces that mirror observed non-linear dynamics, such as the sudden onset of congestion.

Core Frameworks for Divergence Field Modeling

Several mathematical frameworks underpin divergence field-based synthetic trace generation. The most common is the conservation law with a non-linear flux function, often solved using finite volume methods. Another framework uses stochastic partial differential equations (SPDEs) to introduce random perturbations that mimic driver variability. A third approach employs machine learning to learn the divergence field from data, then uses it as a forcing term in a physics-based simulator. Each framework has distinct trade-offs in accuracy, computational cost, and data requirements.

Conservation Law with Non-Linear Flux

This framework starts with the continuity equation: ∂ρ/∂t + ∂(ρv)/∂x = S, where ρ is density, v is speed, and S is the source term (divergence). The flux function q = ρv is modeled non-linearly, e.g., using a speed-density relation that includes hysteresis. Solving this equation on a grid yields density and speed fields over time, from which individual vehicle trajectories can be extracted via particle tracking. The main advantage is physical consistency; the main drawback is computational cost for large networks.

Stochastic Partial Differential Equations (SPDEs)

SPDEs add a noise term to the conservation equation, capturing random fluctuations in driver behavior and external events. For example, a Wiener process can model the randomness of lane changes or the impact of a minor accident. The resulting traces have more realistic variability, but the model is harder to calibrate and requires careful choice of noise intensity to avoid unrealistic oscillations.

Data-Driven Divergence Fields

In this hybrid approach, a neural network is trained on historical data to predict divergence at each cell given local density and speed. The predicted divergence is then used as the source term in a physics-based simulator. This combines the flexibility of machine learning with the stability of physical laws. However, it requires a large training dataset and may overfit to specific road geometries.

Step-by-Step Workflow for Generating Synthetic Traces

This section outlines a practical workflow for implementing a divergence field-based synthetic trace generator. The steps assume familiarity with basic traffic flow theory and programming in Python or a similar language. We focus on the conservation law framework due to its balance of accuracy and tractability.

Step 1: Define the Network and Boundary Conditions

Represent the road network as a directed graph, with edges as road segments and nodes as intersections or points of interest. For each edge, specify its length, number of lanes, free-flow speed, and capacity. Boundary conditions include inflow rates at entry points and outflow rates at exits, which can be constant or time-varying based on observed data. For example, a typical project might model a 10 km highway stretch with three on-ramps and two off-ramps, using loop detector data to set boundary flows.

Step 2: Discretize the Conservation Equation

Use a finite volume method to discretize the continuity equation on a grid with spatial step Δx and temporal step Δt. For each cell i at time n, update density as: ρ_i^{n+1} = ρ_i^n + (Δt/Δx)(q_{i-1/2}^n - q_{i+1/2}^n) + Δt·S_i^n. The flux at cell boundaries is computed using a non-linear speed-density relation, such as the Greenshields model with a hysteresis term. The source term S_i^n is the divergence field, which can be precomputed from known ramp flows or modeled as a function of density.

Step 3: Simulate the Flow Field

Run the simulation for the desired time horizon, updating density and speed at each time step. The Courant-Friedrichs-Lewy condition must be satisfied: Δt ≤ Δx / v_max. For a typical highway with v_max = 120 km/h and Δx = 100 m, Δt ≈ 3 seconds. The simulation produces a time series of density and speed for each cell.

Step 4: Extract Individual Trajectories

To generate synthetic traces, inject virtual vehicles at entry points with random departure times and desired speeds. At each simulation time step, update each vehicle's position based on the local speed field. Use interpolation if the vehicle is between cell centers. The result is a set of trajectories (time-stamped positions) for each vehicle. One team I read about used this method to generate 10,000 synthetic taxi trips in a mid-sized city, which they then used to train a demand prediction model.

Step 5: Validate Against Real Data

Compare the statistical properties of the synthetic traces against real-world data, if available. Key metrics include travel time distributions, speed-density scatter plots, and congestion patterns. If the synthetic traces show unrealistic features (e.g., no stop-and-go waves), adjust the divergence field parameters or the speed-density relation. Iterate until the synthetic traces pass a Kolmogorov-Smirnov test for travel times.

Tooling, Stack, and Economic Considerations

Implementing a divergence field generator requires a mix of simulation software, programming libraries, and computational resources. Open-source options like SUMO (Simulation of Urban MObility) can be extended with custom divergence fields, while commercial platforms like PTV Vissim offer built-in non-linear models but at a higher cost. For custom implementations, Python libraries such as NumPy, SciPy, and TensorFlow are common. The choice depends on team expertise, budget, and project scale.

Open-Source Stack

SUMO is widely used for traffic simulation and supports custom flow models via TraCI (Traffic Control Interface). However, its default models are linear; implementing a divergence field requires modifying the simulation kernel or using external scripts to set speeds dynamically. This is feasible for research teams but may be too complex for production environments. Another option is the Python package flow, which integrates with SUMO and allows defining custom acceleration models. The cost is zero, but development time can be significant.

Commercial Stack

PTV Vissim includes advanced non-linear models like the Wiedemann car-following model, which can produce realistic congestion patterns. It also supports external driver model DLLs, allowing users to inject custom divergence fields. The license cost is high (often tens of thousands of dollars per year), but it offers a stable, validated platform. For large-scale projects with existing commercial licenses, this is often the preferred choice.

Cloud and Cost Management

Simulating large networks with high resolution can be computationally expensive. A typical simulation of a city-scale network (e.g., 1000 km of roads) at 1-second resolution for 1 hour might require 100+ CPU hours on a cloud instance. Using spot instances or preemptible VMs can reduce costs by 50-70%. Practitioners often run multiple simulations in parallel to calibrate parameters, so budgeting for compute time is essential.

Growth Mechanics: Scaling and Maintaining Synthetic Trace Systems

Once a divergence field generator is built, it must be maintained and scaled as new data becomes available or network conditions change. This section covers strategies for updating the model, handling larger networks, and integrating with downstream systems like demand forecasting or reinforcement learning for traffic control.

Model Updates and Calibration

Traffic patterns evolve due to infrastructure changes, seasonal effects, or policy interventions. The divergence field parameters—such as capacity drop magnitude or hysteresis strength—should be recalibrated periodically. One approach is to use a sliding window of recent real-world data to re-estimate the divergence field using Bayesian inference. This can be automated as a daily or weekly pipeline. For example, a city transportation department might update its synthetic trace generator every month using loop detector data from the previous 30 days.

Scaling to Large Networks

For networks with thousands of edges, the finite volume method becomes computationally intensive. Domain decomposition techniques can split the network into subregions that are simulated in parallel, with boundary conditions exchanged between subregions. This requires careful load balancing to avoid idle cores. Another approach is to use a multi-scale model: simulate the core urban area with high resolution and the surrounding suburbs with coarser resolution, using the divergence field to match flows at the boundaries.

Integration with Downstream Systems

Synthetic traces are often used as input for machine learning models, such as travel time prediction or route recommendation. To ensure compatibility, the generator should output traces in a standard format (e.g., GeoJSON or Parquet) with consistent timestamps and coordinate systems. Version control of the generator and its parameters is critical to avoid mismatches. One team I read about used synthetic traces from a divergence field model to train a reinforcement learning agent for adaptive traffic signal control, achieving a 15% reduction in average delay compared to a baseline trained on real data only.

Risks, Pitfalls, and Mitigations

Divergence field-based synthetic trace generation is not a silver bullet. Practitioners must be aware of common pitfalls that can lead to unrealistic traces or wasted effort. This section outlines the most frequent issues and how to address them.

Overfitting to Calibration Data

If the divergence field is tuned too closely to a specific dataset, the synthetic traces may not generalize to new scenarios (e.g., a different day of the week or a special event). Mitigation: use cross-validation during calibration, and reserve a held-out dataset for validation. If the model performs poorly on the held-out set, reduce the number of free parameters or use regularization.

Computational Instability

Non-linear models can exhibit numerical instabilities, especially when the Courant-Friedrichs-Lewy condition is violated or when the divergence field has sharp gradients. This often manifests as negative densities or unbounded speeds. Mitigation: use a smaller time step, apply flux limiters (e.g., minmod or superbee), and smooth the divergence field with a Gaussian filter.

Ignoring Multi-Modal Interactions

Most divergence field models focus on passenger vehicles, but real traffic includes buses, trucks, bicycles, and pedestrians. Ignoring these modes can produce unrealistic traces, especially near transit stops or bike lanes. Mitigation: extend the model to include multiple vehicle classes with different speed-density relations, and add mode-specific divergence sources (e.g., bus stop dwell times).

Data Privacy Risks

Even though synthetic traces are generated, they may inadvertently reproduce patterns from real training data, potentially leaking private information. Mitigation: apply differential privacy techniques during calibration, or use a generative adversarial network (GAN) to ensure that synthetic traces are statistically distinct from any real trajectory.

Mini-FAQ and Decision Checklist

This section addresses common questions and provides a checklist to help you decide if the divergence field approach is right for your project.

Common Questions

Q: How much real data do I need to calibrate a divergence field model? A: It depends on the complexity of the network. For a simple highway segment, a few hours of loop detector data may suffice. For a city-scale network with many ramps and intersections, weeks or months of data are typically needed to capture the full range of conditions.

Q: Can I use this method for pedestrian or bike traffic? A: Yes, but the flow models are different. Pedestrian flow is often modeled with social force models rather than continuum equations. However, the divergence field concept can be adapted by defining a density of pedestrians and a velocity field that depends on local density and obstacles.

Q: How does this compare to GAN-based synthetic trace generation? A: GANs can produce very realistic traces but often lack physical consistency (e.g., vehicles may teleport or violate speed limits). Divergence field models enforce physical laws, making them more reliable for applications like traffic control, but they may be less flexible for capturing rare events.

Decision Checklist

• Use divergence field if: you need physically realistic traces for traffic engineering analysis, you have some real data for calibration, and you can tolerate moderate computational cost.
• Consider alternatives if: you need to generate traces for a network with very little data (use a simpler linear model), you need extreme realism for rare events (use a GAN), or you have a very large network and limited compute (use a mesoscopic model).

Synthesis and Next Actions

The divergence field method offers a principled way to generate synthetic traces that capture non-linear traffic dynamics. By modeling flow as a continuous field with divergence constraints, practitioners can produce traces that are both realistic and physically consistent. The key steps—defining the network, discretizing the conservation equation, simulating the flow field, extracting trajectories, and validating—form a repeatable pipeline that can be adapted to various scales and contexts.

Immediate Next Steps

If you are considering adopting this approach, start with a small pilot project on a well-studied corridor. Use open-source tools like SUMO with custom Python scripts to implement a basic divergence field. Compare the synthetic traces to real data from a single day. Once you have validated the model on the pilot, scale up to larger networks and automate the calibration pipeline. Remember to document your divergence field parameters and version control your code to ensure reproducibility.

Final Thoughts

No model is perfect, and the divergence field approach has its limitations—computational cost, data requirements, and the need for careful calibration. However, for applications where non-linear dynamics are critical, such as evaluating traffic management strategies or training reinforcement learning agents, it provides a valuable tool. As always, verify critical details against current official guidance where applicable, and consult with domain experts for specific projects.