Limited sensor coverage and fragmented, mode-specific modeling infrastructure hinder the holistic monitoring of modern transportation networks. The resulting data blind spots prevent current models from capturing dynamic, cross-modal dependencies, where a disruption in one mode, such as a metro closure, triggers cascading surges in others, forcing planners and Traffic Management Centers to rely on reactive, siloed strategies.
To improve the state of the art, this project proposes a Virtual Sensor paradigm driven by Physics-Informed Generative AI. By integrating fundamental transportation physics with generative deep learning, the framework synthesizes high-fidelity data for sensor-sparse regions by inferring correlations from existing sensing infrastructure, creating cost-effective virtual data streams that simulate physical sensors and provide more complete multimodal network data for real-time operations and long-term planning. The project also evaluates the policy and governance dimensions of integrating emerging AI use cases, such as AI-generated data, into DelDOT’s planning, design, and operations.
Major research outputs include the Virtual Sensor framework itself, a hybrid generative architecture that synthesizes realistic flow and demand data for sensor-sparse modes such as micromobility and pedestrians; models of interdependent dynamics that quantify cross-modal dependencies and cascading disruptions to improve estimation accuracy in undersensed regions; and real-world validation through partnership with the Delaware Department of Transportation to test the framework’s ability to generate accurate synthetic data. The methodology applies a Physics-Informed Deep Learning approach with Fourier feature embeddings, trainable physics-based adjustments, and interdependent demand-capacity coupling. The project also delivers an AI policy and governance analysis, including a decision tree or matrix to guide responsible integration of AI into DOT planning, design, and operations, supported by technical reports and code repositories.
Universities Involved
University of Delaware
Principal Investigators
Arde Faghri
Mark Nejad
Philip Barnes
Andrea Pierce
Expected Research Outcomes & Impacts
The project advances traffic reporting, management, and prediction by moving beyond siloed, single-mode approaches. Traditional methods trade accuracy for cost, requiring dense sensor networks for high-fidelity modeling while sensor-sparse regions rely on less accurate models; this project addresses that tradeoff by learning cross-modal interactions to better use available multimodal data and applying physics-informed generative learning to synthesize high-fidelity states in unmonitored areas.
It contributes to transportation research by improving the representation of existing networks and integrating spatially informed, cross-modal deep learning into traditional modeling frameworks without increasing data requirements, with applications including improved routing, infrastructure planning, and network modeling. Improved data coverage can also support safety outcomes by estimating conditions where physical sensors are absent and revealing how incidents and capacity reductions shift demand. The AI policy and governance framework will provide DelDOT and other agencies practical guidance for responsible AI adoption.
Subject Areas
Physics-Informed Machine Learning, Multimodal Networks, Virtual Sensors, Transit Corridors, AI Governance
