Blog

On October 15, millions of people will participate in the Great California ShakeOut. Cyrus Shahabi, Helen N. and Emmett H. Jones Professor in Engineering and professor in USC Viterbi and USC Stevens‘ Ming Hsieh Department of Electrical and Computer Engineering and Thomas Lord Department of Computer Science, studies how AI-based traffic systems respond to disruption in urban road networks. His research combines deep learning, spatial data management and large-scale traffic modeling to improve real-time situational awareness during emergencies in Los Angeles.

When a major earthquake hits Los Angeles, the road network does not just get congested. It becomes a different road network. Certain roads will close entirely for safety reasons. Two-way streets may be converted to one-way evacuation corridors. The map that traffic systems were built on — the one with sensors at known intersections, edges connecting known nodes — is no longer the map.

That shift is not primarily a civil engineering problem. It is a computer science problem. And it is one our current AI tools are not equipped to solve.

Most AI-based traffic forecasting, including my own most-cited work, is built on graph neural networks. The “graph” can represent the traffic network in different ways, for example, intersections as nodes and roads as edges, or sensors as nodes and the road-network distances between them as weighted edges, with relationships learned between specific points over time.

The fundamental assumption is that the graph doesn’t change often (e.g., new roads aren’t added very frequently). Traffic changes much more frequently than the network. You train a model on that fixed structure, and at inference time — when you actually use it — that structure is still there.

After an earthquake, that assumption fails. If you change the graph, what the model learned is no longer correct. And if you want to retrain it on the new network, that takes a long time. In a disaster, you don’t have that time.

The traditional answer is traffic simulators, which can model a changed network from scratch. But simulators are slow and also require a human in the loop — someone who takes real-world data, sets the parameters, and runs the simulation. That process is time-consuming, and the results are only as good as the parameters someone set under pressure, with incomplete information.

A Graph That Doesn’t Depend on the Road

The model my team has been developing, called DeepStateGNN, takes a different approach. Instead of anchoring our model to specific physical sensors or intersections, we created what we call meta-nodes: abstract representations of semantic categories. One meta-node might capture sensors on north-south freeways with multiple lanes. Another captures how sensors across the city tend to behave in rain.

The result is a much smaller, fixed graph — not a map of the road network, but a map of what road conditions tend to mean. When sensors fail or disappear, the meta-nodes absorb the change without requiring the graph to be rebuilt. When new sensors are added, they contribute to existing meta-nodes rather than adding new nodes to the graph.

We originally built this for scalability and sensor robustness. What we found surprised us. When we tested DeepStateGNN on sensors it had never seen during training, it became significantly more accurate, not less, despite being much smaller and faster than competing models. If a new sensor appears on a three-lane road in Hollywood, the model draws on what it knows about Hollywood roads and how three-lane roads behave, and stitches together a strong prediction for that new location. We even fed the strongest competing models the same semantic information about the new sensors. They still could not match DeepStateGNN’s accuracy.

The Earthquake Application

That generalization is what makes the earthquake scenario interesting. A surrogate model of this kind could evaluate more than 1,000 road configurations per second — at least four orders of magnitude faster than traditional traffic simulators. No human configuring parameters. Data in, data out.

The challenge after a major disruption is rarely finding one obvious response. Every action changes conditions elsewhere in the network, and the number of possible combinations quickly becomes too large to evaluate manually. If a validated version of this system were connected to live road and traffic feeds, emergency managers could test hundreds of responses — signal changes, turn restrictions, lane reversals, temporary closures — in seconds rather than waiting for lengthy simulations. That could help preserve access to hospitals and affected neighborhoods while avoiding actions that simply move congestion elsewhere. Experienced operators would remain essential. The system would simply let them ask many more “what if?” questions while there is still time to act.

What we are proposing is a validated research prototype, not a system an emergency operations center could turn on tomorrow. Demonstrating that it remains accurate during real disruptions, under road configurations it has never encountered, is the first step. Deployment would then require earthquake-specific validation, live traffic and closure feeds, integration with agency systems, and exercises with the relevant agencies.

The research is intended to close the computational gap. A subsequent pilot would need to close the operational and institutional gap.

Published on October 13th, 2026

Last updated on October 13th, 2026

This article may feature some AI-assisted content for clarity, consistency, and to help explore complex scientific concepts with greater depth and creative range.