Professor of Civil and Urban Engineering, NYU Abu Dhabi
We study how traffic moves, and we build math and AI tools to predict it, control it, and keep it safe. Our work spans traffic flow theory, traffic state estimation, traffic signal control, neural operators for physics, and the security of AI in autonomous vehicles.
Papers · Team · Google Scholar · ORCID · sej7@nyu.edu
News
Nov 2026: Appointed Area Editor of Transportation Science (Traffic, Demand, and Networks).
2026: “Fractional is better: Learnable derivative orders in neural operator learning” at ICML 2026.
2026: Our stability analysis of decentralized max-pressure signal control appears in Transportation Research Part C (ISTTT 26 Proceedings).
A trained network is a pile of nonlinear layers. We rewrite what it computes as linear operators, which lets us look inside and see what each layer is doing.
Diagram showing the layers of a neural network represented as linear operators.
Quick ways of training networks to resist attacks can collapse without warning. We traced the problem to the norm used in training. Choosing the ℓᵖ norm adaptively keeps training stable.
Plots of robust accuracy during fast adversarial training with fixed and adaptive norms.
We use matrix completion techniques to forecast sensor states second-by-second over short (10-20 second) horizons.
Each intersection looks only at the queues around it and picks its own signal timing. Paying attention to where cars sit in each queue moves traffic better than the standard rule. The two animations show the difference.
Travel time on a road is not one number. The same trip can take 10 minutes or 25, so we estimate the whole spread of possible times. We describe that spread as a mix of a few simple hump-shaped curves and keep only the handful that matter. The estimate updates on the fly as new data stream in from vehicles.
A crash sends waves of slowing traffic up and down the highway. A sensor notices the crash only when one of those waves reaches it. The detection time, t(x, ℓ), depends on where the crash happens (x) and where the sensors sit (ℓ). We place sensors so that, with high probability, any crash is spotted within a target time. The same waves also tell us where the crash happened.
Read the paper · Patent: sensor placement · Patent: incident location
Space-time diagrams showing waves from a highway incident traveling upstream and downstream until they reach the nearest sensors.