Oak Ridge National Laboratory is exploring how artificial intelligence (AI) can accelerate scientific discovery through the Department of Energy’s Genesis Mission, with researchers using the technology to optimize supercomputer applications and investigate challenges in climate modeling and energy planning.
Oscar Hernandez, a senior computer scientist at Oak Ridge, told MeriTalk in an interview that the effort aims to move beyond using AI to generate answers and toward helping scientists develop hypotheses, test ideas, and evaluate results.
“We want AI to help us propose hypotheses and then connect them all the way through implementation and evaluation – to see if they work or don’t work,” Hernandez said.
The Genesis Mission, launched in November 2025 and led by the Energy Department, brings together federal agencies and industry partners to apply advanced AI and computing capabilities to scientific challenges spanning energy, health, space, and other research areas.
Of the 300 projects currently operating under the mission, Oak Ridge leads nine and contributes to 32 more. Its projects address areas including energy systems, fission reactors, water resources, and scientific computing.
Using AI to improve supercomputing
One of Oak Ridge’s early efforts involves using AI to optimize scientific software running on the Frontier supercomputer.
Through an early-access program, researchers connected Google’s AlphaEvolve system to Frontier to explore ways to improve the performance and energy efficiency of scientific applications.
AlphaEvolve generates variations of existing code that researchers can test for potential improvements. Instead of relying solely on experts to identify and implement optimizations manually, the system allows AI to explore a wider range of possible solutions.
Hernandez said the approach has already shown potential to significantly reduce the time required for certain optimization tasks.
“What could take months or even potentially years – or require significant expertise – we were able to shorten to something like days,” he said.
Those gains depend on how researchers guide the system and evaluate its output.
Hernandez said his team found that carefully defining optimization goals was critical. By using evaluation methods and reward functions that emphasized correctness and encouraging mixed-precision techniques, researchers guided AlphaEvolve toward solutions better suited to their scientific computing needs.
“How you ask AI to accomplish a goal matters a lot,” Hernandez said.
Even with those improvements, AI-generated code must undergo rigorous testing and expert review before researchers can incorporate it into scientific applications.
Expanding AI across research disciplines
Beyond supercomputing, Oak Ridge is applying AI to research challenges involving additive manufacturing, climate and water-cycle modeling, and electricity planning.
In one project, researchers are studying future water availability in the southeastern United States to help inform electricity-generation planning.
The work could help scientists better understand how changing environmental conditions affect energy systems and the decisions needed to manage them.
Researchers are also examining whether AI models can make certain predictions faster than traditional large-scale simulations, potentially allowing scientists to explore more scenarios and identify promising areas for further study.
These applications reflect a broader goal of the Genesis Mission: using AI not only to speed existing research methods, but also to help scientists investigate questions that might otherwise require substantial time and computing resources.
Moving from AI-generated ideas to verifiable science
Another area of exploration involves developing AI “co-scientists” that can analyze scientific literature and propose hypotheses for researchers to investigate.
Hernandez described a longer-term approach in which AI supports multiple stages of research, from generating ideas to implementing experiments and evaluating outcomes, while scientists retain responsibility for determining what the evidence supports.
That work requires collaboration across scientific disciplines, bringing together researchers who understand the underlying problems with those developing and applying AI tools.
“It’s not only scientific discovery but also discovering each other – the expertise we have and how to work together,” Hernandez said.
However, as researchers expand AI’s role in scientific discovery, Hernandez said one of the biggest challenges is ensuring they can verify the growing volume of information and proposed solutions these systems produce.
“AI can generate so much that we’re starting to become the bottleneck in verifying it,” Hernandez said.
The challenge goes beyond determining whether an AI-generated result appears promising. Verification is critical to efforts to use AI across the research process, especially as the technology lets scientists generate and test ideas faster. Scientists must also assess whether a proposed experiment is safe and whether other researchers can reproduce the findings.
“We need to be strategic about how we verify the massive amount of ideas and information AI can generate at the scales we’re talking about,” Hernandez said.