LLNL students, mentors put agentic AI tools to work at cross-laboratory clinic

Nearly 100 Lawrence Livermore National Laboratory (LLNL) students, mentors and staff joined colleagues from across the Department of Energy (DOE) complex recently for a hands-on clinic focused on using agentic artificial intelligence tools in real scientific and technical work.

Held in the Lab’s Research Library, the 2026 Cross-Laboratory AI Clinic combined shared presentations from DOE and Oak Ridge National Laboratory (ORNL) with local working sessions at LLNL. Participants could bring their own data or research problem or work through a challenge problem developed through LLNL’s Data Science Institute (DSI).

“We wanted to give students and staff an opportunity to get started using AI tools and learn about more advanced topics, like using coding agents,” said LLNL data scientist Jason Bernstein, director of the DSI Consulting Service and Open Data Initiative (ODI). “A lot of people are being told that they can use AI to solve their problems faster and more efficiently, but they don't know how to accomplish this. We wanted to push people forward a little bit in their AI journey.”

The event brought together participants from eight DOE laboratories and sites: LLNL, ORNL, Idaho National Laboratory, Lawrence Berkeley National Laboratory, the National Energy Technology Laboratory, the National Laboratory of the Rockies, Pacific Northwest National Laboratory and Thomas Jefferson National Accelerator Facility. Most of the LLNL attendees worked on the DSI data challenge problem, while others brought their own data or problem, learning how they could apply the tools to their everyday work.

The day began with a shared program that included DOE remarks, a prompt-engineering presentation, an agentic AI demonstration and an industry panel including representatives from NVIDIA, Google, OpenAI, AMD and Anthropic. The afternoon shifted to local working sessions at each participating site, followed by reports-out from the labs and closing discussion.

DOE Associate Principal Deputy Under Secretary for Science Kristen Ellis, who addressed attendees remotely, tied the clinic to the broader goals of the Genesis Mission and the need to prepare a workforce capable of using AI in scientific research.

“We are at a crucial moment with the Genesis Mission and AI generally, and we're going to need the talent through the national labs to help get involved in this AI revolution,” Ellis said.

Ellis added that technology alone will not be enough to achieve the mission’s goals and emphasized the value of building AI fluency among scientists and students.

“We're going to need people,” Ellis said. “The scientist of the future is not someone who's going to be replaced by AI, but the scientist who knows how to effectively collaborate with AI will have extraordinary advantages in speed, creativity and impact.”

At LLNL, organizers discussed the difference between conventional chatbots and coding agents. While a chatbot is primarily designed for conversation, a coding agent can inspect files, write and edit code, run programs and use external tools. That broader capability can make agents useful for automating scientific workflows, but it also requires careful review, secure handling of data and clear validation of results, they said.

For Kenneth Song, a master’s student in computer science at UC San Diego and an intern in LLNL’s Cybersecurity Summer Institute, the clinic was his first experience using tools like Open AI’s Codex.

“I've never really worked with agentic tools, so I thought this AI clinic would really be beneficial,” Song said. “I'm just experimenting with Codex and finding that to be pretty powerful too.”

The challenge problem gave participants a concrete scientific use case. Developed using data curated for DSI’s Data Science Challenge by Haichao Miao, a research scientist in LLNL’s Center for Applied Scientific Computing, the exercise asks participants to build autonomous AI tools for inspecting additively manufactured materials.

The dataset was derived from the AMD Lattices project led by Jenny Nicolino, a staff scientist in Physical and Life Sciences' Materials Science Division. It consists of X-ray computed tomography scans of lattice structures, including 3D-printed octet lattices whose internal networks may contain bent, broken, missing or unusually thin struts. Participants explored workflows that could convert volumetric scan data into usable formats, separate lattice material from the background, recover the structure’s network of nodes and struts and compare the results with ground-truth.

Rather than simply asking a model for an answer, the exercise demonstrated how an AI agent could coordinate a broader pipeline involving code, image processing, visualization and quality checks.

Research scientist Magi Yassa, a group leader in LLNL’s Materials Engineering Division, said her work involves materials development for additive manufacturing, and that the clinic helped expand her sense of what agentic tools could handle.

“Understanding the power of these tools helps us [scientists] to think outside of the box, and how we can utilize all of that to our advantage,” Yassa said.

Yassa works with large collections of microscopy images generated during resin development and screening. In one project, she explained, her team produced more than 1,000 scanning electron microscope images, creating a volume of data that is difficult to evaluate manually. She said AI-assisted workflows could help classify images, identify useful printing conditions and narrow the range of results requiring closer human inspection.

Irabiel Romero, a University of California, Merced doctoral student and DSI intern working on power-grid optimization, said he attended to learn more about agents, skills and subagents and how they might be used in his own research.

“If you just hear about these tools secondhand, you might not know how they work or if they’re even going to be good for you,” Romero said. “The clinic provided examples, and those examples gave us detailed steps what to do. And then we were able to get help from people if we needed the help.”

Romero said agentic tools could eventually help him sort through large numbers of simulation files and identify conditions that require closer attention. Instead of manually opening every output, an agent could surface the files most relevant to a particular technical question.

The challenge problem and its associated dataset are on the ODI website, which makes distinctive LLNL datasets available to the broader data science community. The initiative supports curriculum development, raises awareness of the Lab’s data science efforts and creates opportunities for future collaboration.

The clinic was led by Bernstein; data scientist Mary Silva from LLNL’s Global Security Computing Applications Division; LLNL data scientist Denvir Higgins; LLNL GenAI Ops developer Tyler Alcorn; and DSI administrator Sira Neily. It was sponsored by LLNL’s Computing directorate, the Lab’s Academic Engagement Office and DSI.

Beyond the technical instruction, Bernstein said one of the most valuable outcomes was the peer-to-peer learning that emerged during the working session. That collaborative model is one LLNL hopes to build on through future clinics, more targeted training and continued support from DSI.

“It wasn’t just the mentors helping the participants; it was the participants helping other participants, which is really nice to see,” Bernstein said. “The best way to learn is to teach other people.”