Physics-Informed Deep Learning For Traffic State Estimation: A Survey and the Outlook

Xuan Di, Rongye Shi, Zhaobin Mo, Yongjie Fu

Abstract

For its robust predictive power (compared to pure physics-based models) and sample-efficient training (compared to pure deep learning models), physics-informed deep learning (PIDL), a paradigm hybridizing physics-based models and deep neural networks (DNN), has been booming in science and engineering fields. One key challenge of applying PIDL to various domains and problems lies in the design of a computational graph that integrates physics and DNNs. In other words, how physics are encoded into DNNs and how the physics and data components are represented. In this paper, we provide a variety of architecture designs of PIDL computational graphs and how these structures are customized to traffic state estimation (TSE), a central problem in transportation engineering. When observation data, problem type, and goal vary, we demonstrate potential architectures of PIDL computational graphs and compare these variants using the same real-world dataset.

Introduction

Traffic State Estimation (TSE) is the critical task of inferring traffic variables—such as density, speed, and flow—across road segments using sparse sensor data. Historically, researchers have relied on two distinct paths: physics-based models (e.g., Lighthill-Whitham-Richards) and data-driven deep learning models. While physics-based models are interpretable and data-efficient, they often fail to capture complex, non-linear human driving behaviors. Conversely, pure data-driven models require massive datasets and can produce physically inconsistent results, such as negative traffic density.

The positioning of PIDL as a hybrid between pure physics and pure machine learning.

Physics-Informed Deep Learning (PIDL) occupies the "sweet spot" between these two extremes, leveraging scientific theory to regularize neural networks. This survey explores how to effectively "inject" physical laws into neural networks to achieve high accuracy and physical consistency, particularly in "small data" regimes where sensors are limited.

Methodology: The Hybrid Computational Graph

The core of the proposed approach is the Hybrid Computational Graph (HCG). This framework integrates two primary components to ensure models remain grounded in physical reality while benefiting from the flexibility of deep learning:

  1. Physics-Uninformed Neural Network (PUNN): A standard deep network that maps spatio-temporal coordinates $(x, t)$ to traffic states like density ($\rho$) or speed ($u$).
  2. Physics-Informed Computational Graph (PICG): A structure that encodes traffic flow physics (e.g., the Conservation Law) as a regularization term within the loss function.
The Hybrid Computational Graph (HCG) framework, integrating data-driven observations with physics-based constraints.

Key Innovations

  • FD Learner (FDL): Rather than using a rigid, pre-defined mathematical relationship between density and flow (the Fundamental Diagram), the authors employ a neural network surrogate to learn this relationship from data while adhering to broader conservation laws.
  • Uncertainty Quantification (UQ): The framework extends to probabilistic models, benchmarking variants like Physics-Informed GANs (PhysGAN) and Normalizing Flows (PhysFlow) to provide confidence intervals for traffic predictions.

Experimental Setup and Results

The authors validated PIDL using both synthetic circular "ring road" data and the NGSIM (Next Generation SIMulation) dataset, which contains high-fidelity vehicle trajectories from the US 101 highway in Los Angeles.

Key Findings

  • Superior Accuracy: On the NGSIM dataset, PIDL variants significantly outperformed both the Extended Kalman Filter (a physics-heavy baseline) and pure Neural Networks. The PIDL-LWR-FDL variant achieved the lowest Relative Error (RE) for density (21.2%) compared to the pure NN's 42.8%.
  • Data Efficiency: PIDL models with only three loop detectors achieved higher accuracy than pure NNs with significantly more data points, demonstrating the power of the physics prior in sparse data environments.
  • Physics as Regularization: The optimal balance between data-driven loss ($\alpha$) and physics-informed loss ($\beta$) is dynamic; as data availability increases, the model should rely more on observations and less on the physics prior.

Outlook and Future Work

While PIDL shows immense promise, several challenges remain:

  • Discontinuities: Standard PIDL gradients struggle with "shockwaves" (sudden density changes). Future work may involve adding viscosity terms to smooth these transitions.
  • Scalability: Current research focuses on single road links. Scaling HCGs to city-wide networks with complex intersections remains a computational hurdle.
  • Transfer Learning: Investigating whether a PIDL model calibrated for one city can be transferred to another by simply adjusting physical parameters is a key area for future exploration.

Summary updated

Abstract

For its robust predictive power (compared to pure physics-based models) and sample-efficient training (compared to pure deep learning models), physics-informed deep learning (PIDL), a paradigm hybridizing physics-based models and deep neural networks (DNN), has been booming in science and engineering fields. One key challenge of applying PIDL to various domains and problems lies in the design of a computational graph that integrates physics and DNNs. In other words, how physics are encoded into DNNs and how the physics and data components are represented. In this paper, we provide a variety of architecture designs of PIDL computational graphs and how these structures are customized to traffic state estimation (TSE), a central problem in transportation engineering. When observation data, problem type, and goal vary, we demonstrate potential architectures of PIDL computational graphs and compare these variants using the same real-world dataset.
BibTeX
@inproceedings{di2023physics,
  author = {Xuan Di and Rongye Shi and Zhaobin Mo and Yongjie Fu},
  title = {{Physics-Informed Deep Learning For Traffic State Estimation: A Survey and the Outlook}},
  booktitle = {Algorithms},
  year = {2023},
  url = {https://openreview.net/forum?id=jdFiK3gOGJj},
}
Figure