Project Aims to Put AI to Work Reducing its Own Energy Needs

It's no secret that artificial intelligence uses a lot of electricity.

A standard ChatGPT query consumes approximately 0.34 watt-hours - roughly 10 times more than a Google search. According to the Pew Research Center, U.S. data centers consumed 183 terawatt-hours of electricity in 2024 - more than 4% of the country's total electricity consumption - roughly equivalent to the annual electricity demand of the entire nation of Pakistan.

Researchers at UC Merced are taking part in a systemwide effort to develop more efficient computing. The multidisciplinary project involves principal investigators from five UCs (Santa Barbara, Merced, San Diego, Irvine and Berkeley) as well as scientists from Lawrence Livermore and Los Alamos national laboratories.

The UC Office of the President issued a $6 million grant for the project; of that, $810,000 will come to UC Merced.

"The rapid growth of artificial intelligence is accelerating both the opportunities for, and threats to, the United States' longstanding economic leadership," Theresa Maldonado, UC vice president of Research & Innovation, said in announcing the grant. "To help keep America in the lead, the University of California is scaling up its commitments in critical emerging areas of scientific research, such as AI, and moving forward with unprecedented speed to fund targeted research that fosters innovation."

Chemical and materials engineering Professor Elizabeth Nowadnick is leading UC Merced's team, which is looking into topological materials. These materials, in which electrons move in unusual ways, might provide more efficient platforms for next-generation computers.

Nowadnick said the UC Merced team has two goals:

• Use AI to advance the discovery of topographical materials, developing a framework for the rapid advancement of other functional materials, such as magnets, superconductors and related quantum materials.

• Employ AI to develop new modalities of switching, or turning electronic devices on and off, advancing possible next-generation rapid, low-power electronic computing devices. That could reduce the computational cost of AI operations.

"We are investigating topological materials whose electronic structures can be rapidly switched with minor disturbances, meaning that the cost of each switching operation is minimized," Nowadnick said. "Because the changes are electronic and do not involve large structural changes to the material in question, such switching may be extremely rapid."

By identifying the most promising candidates before laboratory testing, this AI approach could dramatically accelerate the discovery of new chip materials that switch faster while consuming less energy.

Nowadnick has one postdoctoral researcher, Kuntal Talit, and two physics Ph.D. students, Haseeb Ahmad and Tharushi Ekanayake, working on the project. They are using a powerful computational tool called density functional theory (DFT) to simulate the properties of the topological materials and determine how they might behave in a computer system.

They are also collaborating with computer scientists at UC San Diego to build an AI system that automates DFT calculations, so users can interact with the DFT code using natural language, similar to how people interact with ChatGPT. This agentic, or autonomous, system, which the team has named TritonDFT, can orchestrate the entire DFT workflow, which spans distinct areas of expertise, including physics, DFT details and high-performance computing. This has the potential to increase the speed at which users can perform DFT calculations and makes performing such calculations more accessible to non-experts.

"This special initiative brings together the unmatched academic expertise of UC faculty, world-class scientific talent and capabilities at our national labs, and the strategic leadership within our system to accelerate the scale, reach and impact of AI-powered scientific discovery research that benefits the nation," said June Yu, vice president of UC National Laboratories.