Rohit Kannan receives CAREER award to explore new approaches to energy and infrastructure problems
Kannan, assistant professor of industrial and systems engineering, received over $600,000 from the National Science Foundation for the five-year project.
Electricity doesn’t flow through a power grid the same way every day — or even every minute. To keep the lights on, engineers must repeatedly solve complex planning problems that determine which generators operate and how electricity moves efficiently through the system.
Rohit Kannan, assistant professor in the Grado Department of Industrial and Systems Engineering, is working to make those planning decisions faster and more reliable using machine learning.
A National Science Foundation Faculty Early Career Development (CAREER) award totaling more than $600,000 will support Kannan’s research on using machine learning to solve complex problems that arise in critical infrastructure systems such as power grids, chemical refineries, and gas and water networks.
“Many machine-learning approaches can produce fast answers, but they don’t always guarantee that the solution is optimal,” Kannan said. “In critical infrastructure systems, that matters. We want to accelerate the solution process without giving up the mathematical guarantees that the solution is the best possible one.”
Critical decisions for critical structures
Behind critical infrastructure systems are planning problems that require engineers to make many connected decisions at once, from turning equipment on or off to deciding how much energy or material should move through a network. In power systems, for example, engineers may need to make these decisions repeatedly — sometimes every 15 minutes — as conditions such as weather, demand, and equipment availability change throughout the day.
“If I’m a power grid operator or a refinery operator, I have to solve the same problem over and over again to determine how to schedule power generation or refinery operations for the next time period,” Kannan said. “The structure of the problem remains the same, only some of the data changes. That’s where there is potential to learn how to solve these problems better and faster.”
That potential comes in the form of machine learning. Currently, even the best algorithms struggle to solve large-scale problems — such as the ones needed to operate statewide power grids — efficiently. Kannan’s research uses machine learning to help these systems make faster decisions by learning from past solutions so engineers can get an answer sooner and feel confident it’s the best one.
Familiarity with uncertainty
Kannan is not new to tackling complex planning problems. During his doctoral research at the Massachusetts Institute of Technology, he developed algorithms and software to solve the same kinds of interconnected challenges that arise in infrastructure systems. As a postdoctoral researcher at Los Alamos National Laboratory, he began exploring how machine learning could help speed up those solutions without sacrificing quality.
"From energy systems to manufacturing, many engineering challenges depend on solving complex optimization problems where accuracy matters,” said Harsha Nagarajan, a staff scientist at Los Alamos National Laboratory. “During his postdoctoral work at Los Alamos National Laboratory from 2021 through 2023, Dr. Kannan introduced a 'strong partitioning' approach and demonstrated how machine learning can adapt optimization strategies to specific problems, resulting in faster open-source tools that maintain provable accuracy.”
As part of the CAREER project, Kannan is also looking beyond speeding up today’s planning tools to address another major challenge in machine learning: data. While machine learning models often rely on large amounts of training data, real-world optimization problems are not always available at that scale. Kannan is exploring whether machine learning can be used to generate synthetic examples to train and test future models.
“Machine learning is very data-intensive, and we don’t always have enough real problems to train on,” Kannan said. “So one thing we’re exploring is whether machine learning itself can help us generate new problems that look like the real world.”
While machine learning has gained traction in simpler optimization settings, its use in more realistic, nonlinear problems remains relatively underexplored — and Kannan sees an opportunity to close that gap.
“Recently, we’ve shown that machine learning has significant potential in these more complex, nonlinear problems,” Kannan said. “The goal now is to design new machine learning approaches that can be integrated into existing tools so that we can solve real-world problems more efficiently.”
