A look at How Oak Ridge National Laboratory is leveraging Google

The US Department of Energy's Genesis Mission, which unites DOE National Labs, industry and academia to harness AI for breakthroughs, is looking to integrate cloud, models, and supercomputing.

On July 22, 2026, Google announced a mission-wide commitment of $40 million in AI tokens and cloud credits for Genesis researchers. Google also committed to providing Gemini seats and tokens for one year to thousands of users across DOE’s national laboratories.

Google is also contributing a series of frontier AI models and services to the Genesis Mission including:

• AlphaEvolve, a Gemini-powered coding and discovery agent.
• AlphaFold 3, a model that predicts the structure and interactions of biomolecules.
• AlphaGenome, a tool for understanding how DNA variation affects biology and disease.
• WeatherNext, a family of weather prediction models.
• And AlphaEarth Foundations, a foundational AI model for mapping and understanding Earth.

Constellation Insights caught up with Oscar Hernandez Mendoza, PhD, Senior Computer Scientist for Oak Ridge National Laboratory, at Google Public Sector Summit 2026 to talk shop. Here's a look at some of the highlights:

The scientific method comes first. Hernandez said the collaboration with Google revolves around using AI to propose ideas and hypotheses. Scientists need verifiable artifacts, reproducible experiments and domain-specific evaluation to test AI-generated proposals. This work spans hypothesis generation and scientific application optimization, with the aim of reaching answers faster.

"We are using AI for scientific discovery, where we use the scientific method to validate what AI is proposing," said Hernandez.

More from Google Public Sector Summit 2026:

• NOAA's huge data, cloud migration to Google Public Sector sets up AI plans
• How Google Public Sector’s RISE unit works: Show don’t tell
• Google Public Sector customers: Data foundation eases AI bottlenecks
• Accenture Federal Services, Google Public Sector roll out Gemini Launchpads
• Google Public Sector launches Gemini for Government Services, sees AI agents as integration accelerator
• Security at machine speed has its growing pains

Hernandez added that checking where AI models fail matters as much as where they succeed. "We want to understand not only when AI works, but also when it doesn't because we can make improvements," he said.

Agentic AI is also "opening the doors for understanding the data" and can "accelerate scientific discovery."

AI could improve the efficiency of supercomputing in many ways. Hernandez said power and operating constraints make efficiency increasingly important as supercomputing scales. AI can help screen candidate algorithms and code optimizations, provided the results meet scientific accuracy requirements. "We're hitting the power wall and the operating envelopes for supercomputers," said Hernandez. "AI can help us to be more efficient in terms of time and energy to solution.”

The Genesis Mission spans 297 projects. DOE reported a first-year portfolio of 297 projects on October 8, 2026. Hernandez described research spanning additive manufacturing, fusion and water systems. "We have mission critical applications and we need to make sure that the hypothesis is validated," said Hernandez.

Institutional knowledge and shared skills. Hernandez described a complementary goal in this research: capturing scientific expertise in reusable agent skills and sharing appropriate tools and best practices with the community. "There is a lot of information that's usually just the knowledge of scientists. Part of the Genesis Mission is producing a set of skills that can be shared with the community along with best practices," he said.

Cloud, on-premises and tokens. Hernandez described using cloud AI services alongside Frontier. For example, code changes proposed by AI models are tested on Frontier for accuracy and performance. The choice of cloud and on-premises resources depends on the scientific workload and applicable DOE and ORNL requirements. "Initially we test things at a small scale using cloud resources and then deploy them on leadership-class supercomputers," said Hernandez.

Several Genesis scientific workflows aim to connect laboratory instruments, AI and supercomputing to analyze data and test predictions against simulations and experiments.

He added that Oak Ridge National Laboratory is assessing the token and computing budgets needed for its scientific workloads, supported by Google’s mission-wide commitment. “Google has provided $40 million in token and cloud credits. We see this $40 million investment as an exploratory phase to understand how we're going to use these resources going forward," said Hernandez.

Guardrails and a scientific AI harness. Hernandez said Oak Ridge National Laboratory is focused on the proper guardrails for tools such as Google DeepMind’s AlphaEvolve. The team has connected AlphaEvolve to Frontier to test AI-generated code for accuracy and performance, keeping a record of the results.

"Autonomous AI and self-improvement programs are still in the early stages. They require a lot of guidance. Without it, they may explore irrelevant problems or produce incorrect results," said Hernandez.

Hernandez’s team has built an evaluation harness that tests AI-generated code and preserves the results for inspection and reproducibility. "We need to build a harness system so we can log what AI proposes, the experiments we run and the human role," said Hernandez. "Just like a scientific paper where you have citations, we should be able to attribute and verify what AI does and reproduce the results."

What does success look like? "It's a multidimensional outcome. Obviously, we want to reduce the time to scientific discovery for the problems we’re trying to solve, but time and energy to solution matters as well," said Hernandez.

He added that the aim is to accelerate discoveries, improve supercomputer utilization and reduce energy use per validated result. These outcomes require measurement against scientific and computational baselines. "Besides the time, this is really about using fewer resources to solve a problem," said Hernandez. "This partnership is opening a door for this type of exploration."