A collaborative team led by Michigan Technological University’s Associate Professor Susanta Ghosh has found that artificial intelligence models need less data than previously thought to produce accurate predictions in the development of novel materials.
Developing new materials often starts with understanding their electronic structure, or how electrons are arranged and behave within them. Calculating these electronic structures typically requires expensive simulations in supercomputers. That’s led researchers to train artificial intelligence models on millions of simulated data points to make faster predictions. But according to a recent study led by Associate Professor Susanta Ghosh in Michigan Tech’s Department of Mechanical and Aerospace Engineering, that much training data isn’t necessary.
Collaborating with researchers from the University of California, Los Angeles (UCLA), Ghosh and his team discovered that AI models actually need a lot less data to make accurate predictions. In a study recently published in npj Computational Materials, the research team found that AI models could discard 90% of their training data with virtually no loss in accuracy.
Even after removing 99% of the training data, most predictions stayed highly accurate, while the models trained three or more times faster. The new findings apply to an expansive range of materials, including metals, water and complex alloys.
How can an AI model make accurate electronic predictions with less data? Ghosh says the answer is by using a minimal dataset carefully selected by researchers that is sufficient to capture the diversity of the data.

“Much of the electronic structure data repeats the same information. A small, well-chosen sample captures nearly everything the AI needs to learn.”
Susanta GhoshAssociate ProfessorDepartment of Mechanical and Aerospace Engineering
These findings challenge the necessity of extensive datasets for electronic structure predictions, and could make AI-driven materials development faster and less costly. Ghosh, who is a faculty member of the Center for Artificial Intelligence in Michigan Tech’s Institute of Computing and Cybersystems, believes this could have a large impact in the field of materials development and modeling.
“By identifying representative electronic-structure data and reducing training time by at least threefold, this work can accelerate AI-driven discovery of complex alloys across vast composition spaces,” he said.
Ghosh’s study was conducted in collaboration with a team led by Assistant Professor Amartya Banerjee from UCLA’s Department of Materials Science and Engineering and MTU alumnus Ponkrshnan Thiagarajan ’23, who completed his Ph.D. in Mechanical Engineering-Engineering Mechanics at Tech and is currently a postdoctoral fellow at the Hopkins Extreme Materials Institute. Research was supported by the U.S. Department of Energy, the National Science Foundation, and the University of California Office of the President.
Michigan Technological University is an R1 public research university founded in 1885 in Houghton, and is home to nearly 7,500 students from more than 60 countries around the world. Consistently ranked among the best universities in the country for return on investment, Michigan’s flagship technological university offers more than 185 undergraduate and graduate degree programs in science, technology, engineering, mathematics, computing, forestry, business, health professions, robotics, psychology, social sciences, humanities, and the arts. The rural campus is situated just miles from Lake Superior in Michigan’s Upper Peninsula, offering year-round opportunities for outdoor adventure.