Physics-Aware Variational Autoencoder for Urban Travel Demand Calibration

TMLR Paper6196 Authors

Abstract:

Urban mobility digital twins are revolutionizing how cities manage increasingly complex transportation systems, enabling real-time optimization across multiple stakeholders, services, and dynamic operations. Central to these digital twins is the origin-destination (OD) calibration problem—estimating travel demand patterns that produce realistic traffic simulations matching observed conditions. However, existing calibration methods face critical limitations: they require a prohibitively large number of expensive simulation runs and struggle with high-dimensional city-scale networks.
To mitigate these issues, we introduce ControlVAE, a novel physics-informed neural network approach for sample-efficient OD calibration. Our method leverages traffic flow patterns, embedded in an auxiliary differentiable physics model, to directly calibrate an interpretable neural representation of the OD matrix from observed data. Specifically, we develop a conditional variational autoencoder framework with a controllable cross-attention mechanism that incorporates this traffic simulation model via differentiable physics knowledge. Our experiments on realistic high-dimensional traffic networks, including the Munich network with 5,329 OD pairs, demonstrate superior sample efficiency, requiring 75\% fewer simulation evaluations than standard baselines like SPSA. In addition, ControlVAE reduces the Normalized Root Mean Squared Error (RMSN) by up to 40\% compared to traditional transportation approaches, confirming that the physics-informed deep learning formulation provides a practical advantage over existing OD calibration methods.

Submission Length: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Tianbao_Yang1

Submission Number: 6196