Schuman’s team controls tricky engines with neuromorphic network

Human drivers know intuitively that the amount of gas their car engines need depends on the circumstances. You push the gas pedal down further when merging onto a highway than moving forward in the drive-thru line.

The enormous combustion engines powering locomotives and marine cargo ships also require different amounts of energy in different situations. Dual-fuel engines can increase the efficiency of these large machines by injecting diesel only when energy needs are high; otherwise, the engines can run on gasoline or renewable methanol.

However, making an artificial intelligence (AI) model that accurately determines when the diesel is needed is extremely difficult.

“You need to use the limited information available from the engine at each cycle to make an informed decision about how to change the fuel injection to improve combustion efficiency,” said Katie Schuman, an assistant professor in the Min H. Kao Department of Electrical Engineering and Computer Science (EECS) at the U niversity of Tennessee, Knoxville.

For the last several years, Schuman, her graduate students, and their collaborators at Oak Ridge National Laboratory (ORNL) have been working on a neuromorphic (biological brain-inspired) computing system that determines the best fuel injection parameters at a lower power cost than competing models.

It also has the unique ability to self-improve, adjusting to changes in engine function on a millisecond scale.

Karan Patel, one of Schuman’s PhD students and the lead author of the team’s paper, presented the system at the 2025 Neuro Inspired Computational Elements Conference in March. He also helped deploy the system for its first real-world test on a million-dollar engine at ORNL.

“Seeing our complete system adapt to a real engine and actually improve performance was extremely gratifying,” Patel said. “After months of development, everybody on the team was excited to see it working.”

Starting the Neuromorphic Engine

While working at ORNL, Schuman collaborated with transportation researchers Brian Kaul and Bryan Maldonado to create a neuromorphic, real-time evolution system for diesel engine control.

Kaul and Maldonado had been frustrated by the lack of agility in the modeling approaches they had applied to fuel injection. Algorithms that performed well in the lab would fall short when they encountered real-world conditions outside their training parameters—a significant problem since the very act of running the engine changes the engine’s characteristics over time.

“In the transportation, you have many dynamic problems that typical machine learning (ML) approaches can’t account for,” Schuman said. “Being able to adapt to those changes is one of the key benefits of neuromorphic computing.”

After Schuman joined the UT faculty in 2022, the team secured a seed grant from the UT-ORNL Innovation Institute (UTORII) to expand their work to dual-fuel combustion.

“The UTORII grant was a great opportunity to push our work forward in a new domain with a more complex engine,” said Schuman.

A Dynamic System with Real Results

Each cycle in a combustion engine may last for only one thousandth of a second, but every previous cycle has an impact on the current situation. That makes the problem of engine control very similar to the problems our brains tackle every moment, Schuman said.

“Time matters in the way that the network processes data,” she explained. “What particular sensor readings you get, and in what particular order, is important for what decision you should make next.”

The software system Schuman’s team created, Neuromorphic Optimization using Dynamic Evolutionary Systems (NODES), utilizes a brain-inspired ML technique called online learning: rather than training on a static set of data, NODES processes a continuous stream of incoming data to train incrementally over time.

Every cycle, NODES evaluates signals from engine sensors, creates a series of fuel-injection control networks, and deploys the one that has the greatest chance of improving engine efficiency. The best-performing networks are stored in a component aptly called the Leaderboard.

“If the current environment changes and the deployed network wasn’t trained for that environment, NODES can refer to the Leaderboard to deploy a new network that can handle that environmental change,” Patel explained. “The Leaderboard basically provides the training algorithm with a good starting point, allowing the system to train better networks in the same amount of time.”

The hardware NODES runs on is also neuromorphic, based on the brain’s ability to simultaneously process and use information.

Patel’s novel reconfigurable hardware system, designed alongside ORNL embedded systems expert Brett Witherspoon, runs on two very small and power-efficient CPUs. One of them, dubbed FireBox, is responsible for controlling the engine. The other one hosts NODES and continually improves the control network based on real-time engine conditions.

When an optimal network is ready, it takes command of FireBox. Meanwhile, NODES immediately begins creating a new network based on the information coming from the engine.

This unique approach not only adapts the system to new sensor data but to an engine’s current operating conditions—which it did successfully this May.

NODES was constructed using data from Kaul and Maldonado’s engine taken in 2024, and over the last year its operating characteristics had changed, as everyone knew they would.

The system adapted to the engine’s current conditions within a day.

“We had simulated the system, but now we can show that it actually does the thing that we said it was going to do,” Schuman said. “Deploying a model like this could have a huge impact for the shipping industry and other sectors that use these very big engines. That’s really exciting.”