Inside the lab where AI runs the experiments while scientists sleep

On a Saturday inside the Advanced Manufacturing Laboratory in Livermore, Calif., the only steady sound is the hum of whirring automated instruments. One of them is labeled, in a bit of lab humor, "Peter Pipetter." Across the bench, a robotic arm tends a centrifuge. Over the noise, someone talks about a different kind of throughput: not samples per hour, but experiments a scientist no longer has to be in the room to run.

"We're entering this exciting era, and we're forging that era here at the (Advanced Manufacturing Laboratory) for the next generation of experimenters," a researcher said. "Whether it's in batteries, whether it's in biology, whether it's in alloys, (scientists) now have at their fingertips the ability to run dozens or hundreds, maybe thousands and – in my dream – millions of experiments."

That arc — from a single hand-pipetted sample to a system that runs around the clock — is the structural claim behind Project ARMOR. The program, formally Advanced Robotics for Materials Manufacturing Optimization and Research, is LLNL's attempt to wire together the three pieces that used to be separate jobs: a planner that decides what to test next, a robot that runs the test, and an AI that reads the result and feeds it back into the planner. The bench it lives on is a materials, chemistry and biology testbed housed in the Lab's engineering directorate.

The unit of inquiry is what changes, not the science. In the old model, a graduate student picked one variable, ran one experiment, recorded one result, and started over. Between 2008 and 2014, an earlier generation of machine learning at LLNL cut some of that drudgery by helping scientists rank which experiments were worth running, but the loop still closed in a human head. What ARMOR is testing, in the words of Materials Engineering Division leader Christopher Spadaccini, is the next step: a closed loop in which AI proposes the next experiment, the robot runs it, and the data comes back without anyone being at the bench.

"We're at the tip of the iceberg right now. We're just learning what this can do," Spadaccini said. "That interface with hardware and the physical world is really exciting, and I think it's set to explode."

The "interface" is the load-bearing word. ARMOR sits next to an older sibling, the Alloy Prediction and Experimentation (APEX) platform, which LLNL has been running for several years to search for new metal alloys. APEX is the proof of concept: an autonomous loop that uses AI to design an alloy, a robot to make and test it, and AI again to pick the next composition. The design space for alloys is so large that, as APEX principal investigator Mason Sage has put it, running one experiment every second since the Big Bang would not come close to covering it. ARMOR is the more general-purpose version of that idea, aimed at batteries, biology and chemistry as well as alloys.

The shift shows up first in small, concrete places. Someone recalled a chemist in the materials science lab who used to spend four to six hours pipetting samples into a tube rotator. That time budget capped how many hypotheses she could test in a week. Since adopting the new workflow, she now orders a series of experiments to run Friday afternoon and comes back to full results on Monday. The headline number is the same as in the 2008–2014 wave — more experiments per scientist per week — but the mechanism is different. The earlier generation ranked candidate experiments for a human to run. The new one runs them while the human is at home.

That distinction is what Spadaccini means by "autonomy," as opposed to plain automation. An automated lab executes a pre-written script; an autonomous one decides what the script should be. It is a meaningful line, and it is also where the honest caveats live. Most scientific instruments were designed for human hands, not robotic ones, and wiring them into an AI loop often means writing custom software or retrofitting legacy hardware so the autonomous system can talk to a centrifuge, a tube rotator or a 3D printer. That integration work, in Sage's phrase, is the lab's "secret sauce" — the part that does not show up in a press photo.

Staff scientist Rodrigo Telles, who builds the connective tissue between algorithms and the robots running multi-phase experiments, described the failure mode in plain terms. A robot arm cannot just reach into a centrifuge whenever it pleases; the sequence — load, spin, retrieve, hand off — has to be specified, and only a human knows which microwell plate is supposed to be where. Without that scaffolding, Telles said, the autonomous equipment is "like a band without a conductor." The AI does not yet know that the centrifuge has to finish before the arm reaches in.

That gap is also why the lab talks about the work as a 2026 inflection rather than a finished system. Spadaccini's "tip of the iceberg" framing is forward-looking, and the broader LLNL writeup is careful to flag the cybersecurity, safety and organizational work that still has to land before a fully autonomous pipeline is normal. An AI that controls a robot that handles sensitive materials is, as a lab official noted in that same piece, a category of risk that does not exist when the same AI is summarizing a document.

What does exist now is a working loop on a narrower set of problems. APEX has been running autonomously on alloy discovery for years; ARMOR is extending that pattern across more domains; and the HPCwire industry writeup treats the program as a marker for where national-lab experimentation is headed next. The Daily Press's local report is the first time the lab has put a face, a quote and a working name on the program in the same story.

The reason the change feels larger than a tooling upgrade is the part someone keeps coming back to. Many material properties — corrosion behavior, fatigue life, the way a new alloy responds to radiation — can only be reliably measured by running the experiment, which is why the physical bench has been the "gold standard" of materials science for a century. An AI that can run a million of those measurements in a year does not replace the gold standard. It moves the unit of scientific inquiry from a single experiment, designed and run by one person, to a continuously running system, designed by a person and run by a loop. Whether that loop holds at scale — across biology, batteries and chemistry, not just alloys — is the open question the program is built to test, and the one the next 12 months of ARMOR runs are meant to answer.