Researchers at the University of Delaware report that an AI-agent workflow ran quantum measurements on ultrathin semiconductor devices for campaigns lasting up to six days. It chose successive tests, analyzed their results and prepared reports while a fixed program limited the commands reaching the cryogenic equipment.
The team tested molybdenum disulfide, or MoS₂, in devices one or two atomic layers thick. Quantum transport measurements track how electrical resistance changes with gate voltage and magnetic field, revealing which electronic states are occupied. Researchers built the devices and loaded them into a dilution refrigerator before giving the system a brief measurement goal. Most reported measurements were made at 10 millikelvin.
The agent could choose a sweep, but not override hardware limits
The FermiLink system gave a measurement agent one question at a time, such as where electrical conduction begins. The agent chose a parameter sweep; an analysis agent processed the result and decided whether to continue, repeat a measurement or stop. FermiLink monitored the instruments during long runs and handed the finished data to the next agent.
The agents could issue measurement requests only through a fixed Python acquisition script. Its voltage limits, sweep rates, magnetic-field settings and temperature range were hard-coded, so an agent could not change those safeguards while running a campaign.
Prepared workflow mapped electronic states and stopped on unreliable peaks
For a monolayer device, the workflow combined oscillations in resistance with Hall measurements, which estimate carrier density. It assigned the observed electronic bands and derived a conduction-band spin–orbit splitting of about 13 millielectronvolts, an energy difference between states linked to electron spin. On a second, lower-quality monolayer device, it could not identify reliable oscillation peaks and ended the planned sequence with a negative report.
On bilayer devices, the workflow mapped how occupied states changed with carrier density and electric field, including whether electrons favored one layer or were shared across both. The researchers had already written the step-by-step decision rules and rehearsed them on earlier data. The demonstration therefore shows a system carrying out and adapting a prepared experimental workflow, rather than devising a measurement strategy from scratch. Its ability to transfer to unfamiliar devices remains untested here.
Sources
- Bauer et al., paper abstract and submission record
- Full paper: experimental workflow, measurements and limitations
- Illustrative lead image: Onri Jay Benally, gold dilution-refrigerator rendering, CC BY-SA 2.0; Wikimedia's resized file is used without further changes. It does not show the study's equipment.
