Browse
Log in / Register
Share
April 26, 2026 — Data fusion, neural networks, and sentiment analysis may sound like science-fiction buzzwords, but they are now doing heavy lifting in transportation research at the U.S. Department of Energy’s National Laboratory of the Rockies (NLR).
At the lab’s high-performance computing center, researchers are pairing the Kestrel supercomputer with machine-learning models to tackle five concrete problems: smoothing busy-city traffic, preparing for connected vehicles, scheduling airport shuttles, explaining real-world truck platooning losses, and reading public attitudes toward transit. The projects show how “super intelligence” is moving from academic pilots into operational tools that can cut costs and emissions.
Anyone who has sat through a red light with no cross traffic knows the problem with fixed-timing signals. Newer adaptive signals can react to nearby vehicles, but NLR researchers wanted to push them further using reinforcement learning—a type of AI that improves through trial and error.
Qichao Wang, NLR senior computational transportation scientist, and his team built a virtual road network on Kestrel. They simulated thousands of scenarios with synthetic sensor data and connected-vehicle messages, then trained a model to estimate real-time congestion and predict how alternative signal settings would change the flow. The team also integrated a traffic-signal control module based on NEMA standards, making the system compatible with existing intersection hardware.
The result is a controller that can run on connected-vehicle data alone, without roadside sensors—a major advantage for cities that cannot afford dense infrastructure retrofits.
In a pilot along a busy Chattanooga, Tennessee, highway, the algorithm reduced travel delays by 32% and fuel use by 27%. The improvement was obvious to drivers, Wang recalled: on the first day of the test, a local engineer received a call from a friend who commuted on that road, saying traffic had never been so smooth.
Self-driving cars rely on their own sensors; connected vehicles are different. They talk wirelessly to traffic lights, road signs, and nearby vehicles. Before any of that appears on public roads, researchers need to understand how thousands of these vehicles would interact.
NLR built a V2X digital twin—a replica of a real road system—and simulated 4,000 connected vehicles. A faithful physics-based simulation of one second of operations would have taken longer than the age of the universe, even on Kestrel. The team rewrote the workflow as a parallel algorithm, cutting the time to about three hours per simulation step. Still too slow.
Their solution was to train an artificial neural network on the data produced by that slower simulation. Once the network learned to mimic the physics-based model, it could run a step in one-tenth of a second.
“We essentially trained the model to develop intuition, like humans develop intuition about what outcomes might follow a certain event,” Wang said. That intuition lets the network predict vehicle behavior almost instantly, reducing computational runtime to nearly zero.
The result makes it practical to test hundreds of V2X scenarios in the lab before spending money on field trials.
Airport shuttle buses face a chaotic mix of passenger demands, terminal schedules, and rental-car center peaks. When NLR partnered with Dallas Fort Worth International Airport (DFW), the goal was to find the best routes and dispatch strategies without waiting hours for simulations.
A conventional microscopic model running one day of DFW shuttle operations took eight hours or more on Kestrel. The research team built a new package called ASPIRES (Airport Shuttle Planning and Improved Routing Event-driven Simulation). Instead of simulating every moment, ASPIRES treats the day as a sequence of events—a bus arriving, boarding, departing—and jumps between them, skipping dead time. Combined with a genetic algorithm that searches for near-optimal routes, the tool simulates a full day in about two seconds.
Accuracy also improved. Traffic models calibrated for real-world operations often aim for 85% to 90% accuracy; ASPIRES captured 98% of the observed DFW operations.
Using those results, NLR identified better dispatching schedules and helped support a fleet modernization that saved DFW more than $1 million in the first two years.
Truck platooning links trucks electronically so they can drive closely together, cutting aerodynamic drag and fuel consumption. Test-track and wind-tunnel experiments have shown 6%–8% fuel savings, but real-world roads do not deliver the same results.
To find out why, NLR researcher Tyler Christeson built an automated pipeline to process thousands of hours of dashcam video captured by long-haul trucks. He then repurposed an existing computer-vision model to scan frames, identify cars, construction zones, and other obstacles, and estimate following distances.
The initial hypothesis—passenger cars cutting between platoon trucks—did occur, but Christeson’s analysis showed it was not the main culprit. Highway construction, slowing traffic, and other real-world obstacles were equally disruptive. Together, these factors cut the expected fuel benefit roughly in half.
The vision model also created an efficiency gain for NLR’s research team. Instead of having a person watch and manually code thousands of hours of footage—an expensive and error-prone process—the computer could detect events reliably and quickly. The same detection capability could eventually help automated vehicles, delivery drones, and vans judge their distance from obstacles in urban settings.
Surveys are expensive, slow, and often too small to capture regional differences. NLR’s Sentiment Analysis for Mobility Behavior Assessment (SAMBA) tool takes a different route.
SAMBA combines data science and social science methods to analyze public comments, social media posts, and other text about transportation options. Researchers can deploy it to gauge attitudes on airport mobility, public transit, and automated vehicles, revealing concerns and preferences without conducting a single interview.
For transit agencies, such tools could complement traditional rider surveys and help planners respond to public concerns faster.
What stands out about NLR’s work is the common procedure: create a detailed digital simulation, generate large volumes of data on a supercomputer, then train a neural network to compress the computation into near-instant predictions. That recipe has now been applied to traffic signals, connected-vehicle radio interactions, shuttle operations, and even public opinion.
The Chattanooga demonstration, in particular, marks one of the few published field trials where an AI-based signal controller produced measurable delay and fuel reductions on a real highway. As cities look for cost-effective climate action, these results offer evidence that AI can move beyond theory.
No single pilot solves every congestion problem. But the combination of speed, scalability, and accuracy demonstrated at NLR suggests super intelligence will play a growing role in the way the U.S. builds, operates, and understands its transportation systems.
