A US national lab is teaching AI to be a scientist

Inside the Advanced Manufacturing Laboratory at Lawrence Livermore National Laboratory, automated instruments run experiments through the night. The work continues while the scientists are away from the bench, and the next time they return, the data is already there.

The lab is now closed-loop. AI proposes the experiment, runs it, measures the outcome, and decides what to try next, with no scientist at the bench.

Researchers now have "at their fingertips the ability to run dozens or hundreds, maybe thousands and — in my dream — millions of experiments," the source told the San Diego Union-Tribune.

The "dream" qualifier is deliberate. The ambition is scaled to the institution, not to a single paper.

Christopher Spadaccini, who leads LLNL's Materials Engineering Division, calls the moment the "tip of the iceberg" for AI integration with physical instrumentation. The materials science work is where the evidence sits. Metal alloys, battery chemistries, and chemical compounds are the on-site applications. The robot on the lab bench has a name taped to its side: "Peter Pipetter" — a small piece of branding for a piece of automated equipment that dispenses precise liquid volumes.

Biology and cancer treatments are named in the lab's forward-looking scope. The closed loop is not in those domains yet. The asymmetry is real: materials science is where the loop is operational; cancer biology is where the lab wants it to go.

The historical baseline helps explain what has changed. Between 2008 and 2014, early machine learning at LLNL sped up how experimental data was processed and helped narrow the search space for new materials. Scientists still ran the experiments themselves. The current wave is described as a step change beyond that prior period. AI no longer waits on the scientist. It runs the loop.

The closed loop changes the experimental scientist's job. The mechanical part of running an experiment, preparing samples, moving them between instruments, recording results, has been the bottleneck. The new setup hands that part to the machine. The human contribution moves upstream: choosing which questions to ask, designing the search space, interpreting anomalies the algorithm flags as out-of-distribution, and writing the follow-up experiments.

LLNL is a US national security laboratory with deep materials science expertise inherited from nuclear stockpile stewardship, which gives it the capital and the institutional reason to invest in autonomous science. The on-site instrument room is the visible artifact of that investment.

The lab treats the closed loop as a starting condition, not a finished one. The broader claim about AI running biology or drug-discovery experiments at the same fidelity has not yet been demonstrated on site.

For a working scientist, the practical question is whether the closed loop is robust enough to trust with materials that cost more to make than to model. For a national-security lab, the practical question is whether the rate of materials discovery can keep up with the demand for new alloys, new battery chemistries, and new energetic materials. The next milestone is the bridge from alloys to more complex chemistries, then to biology. Each step needs its own closed loop. The first step is where the work is now.