ACCESS: AI Spots Temperature Risks in US Rail Infrastructure
March 24, 2025 -- Odds are your life is much improved by trains, even if you’ve never stepped foot on one. Everything from food to cars is transported via rail across the country. According to the Association of American Railroads (AAR), “Freight rail is part of an integrated network of trains, trucks and barges that ships around 59 tons of goods per American every year.”
Credit: Shutterstock
There isn’t a more cost-effective or efficient way to move goods across the country. On average, railroads are three to four times more fuel-efficient than trucks, which is why it’s so important to keep the trains safe and running on time.
Researchers at the University of Pittsburgh (Pitt) have been using ACCESS-allocated resources at the Pittsburgh Supercomputing Center (PSC) to train AI to detect when a railway might be in imminent danger of failure. Long stretches of continuous rail are particularly vulnerable to extreme temperature changes due to their inability to expand and contract in the same way smaller segmented sections of rail can. These longer sections of rail can buckle when heat causes them to expand beyond the space they’re set in, or even pull apart if the weather is severely cold.
Testing the railways for temperature vulnerabilities is expensive and time-consuming. It also requires closing down sections of rail, which would impact all the goods that need to be transported in a timely manner. Matthew Belding, a graduate student on the Pitt research team, has a possible solution: create an AI that could predict whether a section of rail is in danger of being damaged due to temperature.
The team ran simulations on PSC’s Bridges-2 supercomputer. This resource gave the team the power to run these simulations in a fraction of the time it would take on traditional computers.
"I know that some of the last simulations we did [on Bridges-2] were a little over 1,500 cases or simulations for specific resistances and specific temperature changes. That took right around two weeks," said Belding. "We estimated it for our computing resources, at least in the lab. It would have taken well over a year, a year and a half."
The Pitt research team tested their tool at the MxV Rail Facility in Colorado. Their AI tool aced the first test with accurate predictions, which gives the team confidence to move on to the next stage of their research. They plan to expand their simulations to include real-world rails that have far more variation in the way they’re constructed.
You can find in-depth details on the methodology of this research in the original article: Simulations on Bridges-2 Teaches AI Program to Prevent Buckling or Breaks in Railways.
Source: Megan Johnson, ACCESS
