In a darkened room somewhere on the University of Washington's Seattle campus, a highly trained physicist is turning a dial - then measuring, and then turning it again, possibly for hours. The device on the other end of the tuning knob is a mouthful: a quantum-noise-limited parametric amplifier. This is the crucial component that takes the impossibly faint signal from a dark matter detector and makes it loud enough to hear on the radiofrequency spectrum.
Just as a musician amplifies the sound of a guitar so an audience member in the back row of a concert can hear the music, physicists use these specialized amplifiers to hear faint signals from quantum computers and dark matter detectors, among other sensitive electronics. Cooled to a fraction of a degree above absolute zero, these amplifiers boost signals while adding almost no noise of their own.
The amplifiers are also, in the words of Pacific Northwest National Laboratory (PNNL) physicist Christian Boutan, "notorious for being difficult to tune up."
Boutan thinks he has a solution: deploying an AI agent that will train itself to auto-tune the amplifier without human intervention, reducing dead time associated with axion dark matter searches and superconducting qubit-based quantum computing as well as buying time for highly trained physicists to use their training more productively.
Now he and his team will be pushing his plan forward with support from the Genesis Mission, the Department of Energy's ambitious plan to double the productivity and impact of American science and engineering through artificial intelligence. Boutan's work fits that bill perfectly.
A Dominant Source of Wasted Time
Nowhere are the stakes clearer than with the Axion Dark Matter eXperiment (ADMX) G2 flagship search for axion dark matter.
Why search for dark matter- Researchers at the National Aeronautics and Space Administration (NASA) have described dark matter as "the invisible glue that holds the universe together."
ADMX's strategy to detect it is to place a microwave cavity inside a strong magnetic field, wait for dark matter axions to convert into detectable microwave photons, and read the output with ultrasensitive low-noise quantum electronics.
Progress in a search like this is measured by how quickly the instrument can sweep across the range of possible axion masses. Every hour spent not scanning is an hour of discovery space left unexplored. So tuning up the amplifiers is "the dominant source of waste of time," Boutan said.
In other words, this PNNL Genesis Mission-funded project, called Autonomous Quantum Amplifier Workflow Optimization and Learning Framework (AQUA-WOLF), is aimed squarely at the largest single drag on the speed of a dark matter discovery, for the ADMX-based search.
This same bottleneck looms for quantum computing. These amplifiers, Boutan noted, are key to doing fast readout of superconducting qubits. As quantum computers scale from handfuls of qubits to thousands, inefficient calibration procedures start becoming a hard ceiling on progress.
How an AI Agent Would Stay in Tune
Boutan explained that using a technique called reinforcement learning, the agent will be trained by being "rewarded" for obtaining the same measure as physicists who tune by hand: gain. The gain, measured in decibels, reflects how much the device boosts the strength of an incoming signal.
Initially, the team will input the goal and allow the agent to turn the dials autonomously, rewarding it in proportion to the gain it achieves. Eventually, the project scope will broaden to include multi-objective tuning, where other considerations such as low noise or stability will be added to the reward. The team will experiment with several algorithms within a chosen framework and then host a contest among agents to down-select to the most promising approach. With the recent advances in AI, they expect all of this advance work to only take a couple of months, once they get started.
The Stretch Goal: An AI That Knows Physics
The project's most ambitious aim is to build realistic physics into the agent by encoding the system's Hamiltonian, which is "basically a description of the energy of the system," said Erik Lentz, a PNNL physicist who is charged with teaching the AI agent to use physics principles in its reasoning and decision-making. Lentz is part of a collaborative group of AQUA-WOLF researchers. The full PNNL team includes Lentz and Boutan, along with Aric Davison, Geoffrey Dolinger, Trent Hartman, Noriaki Kono, Stephen Jones, and Malachi Schram.
Because all parametric amplifiers share an underlying structure, an agent that understands how it can be expected to behave could be plugged into an unfamiliar device and get straight to work. Today's conventional algorithms explore the parameter space "in a kind of inefficient way," said Lentz.
Real-World Test Cases
Once the framework is ready, it will be validated on real hardware at multiple institutions and for different device types - including a broadband amplifier developed at the National Institute of Standards and Technology alongside collaborators at Harvard. The test will evaluate how well the framework works under different conditions and will be conducted by new users not involved in the AI agent's development. In parallel, the PNNL team will use transfer learning to adapt the approach to a second amplifier type, Josephson parametric amplifiers, building toward a system that could, with future funding, be aimed at experiments like ADMX.
Once the framework has been validated, the team plans to release the resulting agent as open-source, with datasets contributing to shared scientific infrastructure being built as part of the Genesis Mission.
The goal isn't to remove scientists from the lab. It's to stop them from spending their hours on a tedious chore so that they can spend more time discovering what holds the universe together.