Quantum control learns from data to stabilise a trapped atom

Researchers at MIT have achieved the first real-time cooling of a single neutral atom using deep reinforcement learning, a technique where a neural network learns to control the atom through immediate feedback. The team demonstrated this capability with a neural-network policy trained on experimental measurements and run on the Quantum Machines OPX+ platform, enabling control decisions fast enough to adapt to changing atomic dynamics, the company says. Peters, co-author of the research and now a quantum engineer at Atom Computing, pointing to a new mode of control where models refine themselves during operation.

Real-Time Cavity Measurements Reveal Atom Motion

This capability bypassed traditional methods reliant on pre-programmed models by allowing the system to refine its control strategies online, directly from experimental measurements. Researchers monitored light leaking from the cavity to infer the atom’s state and then adjusted the depth of the optical tweezer, effectively removing energy from the atom’s motion without directly measuring its position or velocity. The approach achieved cooling through a learned response to the fluctuating light signal, a feat previously unattainable with fixed-sequence control systems.

This new mode of control relies on an agent that learns useful actions from noisy data, rather than requiring precise knowledge of the atom’s location or speed. The team deliberately withheld direct measurements of position and velocity, forcing the neural network to deduce the atom’s state solely from the cavity’s light transmission. More photons indicated a larger oscillation, while fewer signified a smaller one.

By monitoring this light, the agent learned to adjust the tweezer depth, an important step toward building quantum systems resilient to environmental disturbances. The entire feedback loop, including the neural network inference, ran entirely on the OPX+ during each experimental trajectory, demonstrating the hardware’s capacity for real-time processing. This initial loading success, achieved by responding to measured states in real time, increased the probability of trapping exactly one atom. Reliable, long-running quantum computers must adapt as operating conditions change.

Reinforcement Learning Adapts Trap Depth via OPX+

The Quantum Machines OPX+ executed a complete neural network within the real-time feedback loop, a feat enabling the first demonstration of deep reinforcement learning-based cooling of a single neutral atom, according to researchers. The team’s approach focused on manipulating the depth of an optical tweezer, a bowl-shaped potential well used to trap the atom, to remove energy from its motion and achieve cooling, according to Quantum Machines.

By dynamically adjusting the trap depth, the controller effectively alters the “bowl” from shallow and wide to deep and narrow, influencing the atom’s oscillation. Researchers in Prof. Vladan Vuletić’s group combined cavity-based measurements with adaptive feedback running on the pulse processing unit (PPU) inside the OPX+. This let them load a single atom quasi-deterministically, with 92(2)% fidelity, increasing the probability of ending with exactly one atom in the trap.

Unlike conventional methods that require precise knowledge of the atom’s initial state, the agent learned to respond to measurements of light leaking from the trapping cavity, inferring the atom’s condition and adjusting the tweezer depth in real time. This process bypassed the need for an accurate model of the atom’s velocity, a significant advantage given the complexities and potential inaccuracies inherent in modeling cold-atom experiments.

Experimental drift, nonlinearities and measurement noise, often problematic for traditional feedback cooling, are instead learned and compensated for directly by the agent through interaction with the system. The researchers initially trained the agent using simulated data. Visualizations of the experiment reveal the dynamic nature of the control; without feedback, the trap depth remains constant, while with feedback, the network commands adjustments to the depth every 27.3 microseconds, as indicated by the recorded photon detection rates.

Deep RL Achieves Feedback Cooling of Single Atom

The team initially trained the neural network using simulated data, providing a starting point for the control strategy before transferring it to the live experiment. The core of this work lies in the ability to refine models online during operation, a departure from traditional feedback cooling which often relies on pre-defined models of atomic behavior. The researchers report, emphasizing the potential of this approach for tackling complex control challenges in quantum systems.

The system’s architecture allows neural inference to run entirely within the feedback loop, enabling rapid control decisions that respond to changes in the system before they significantly impact performance, the company says. This manipulation of trap depth, guided by the reinforcement learning agent demonstrates a new mode of real-time quantum control.

The researchers observed that without feedback, the trap depth remained constant, resulting in a static tweezer beam. This sequence of decisions, made on a single atom, illustrates the speed and responsiveness of the learned control policy. The team’s work demonstrates that a learned policy can carry out feedback cooling in a shot-noise-limited experiment on fast timescales., Quantum Machines reports.

Stay current

See today’s quantum computing news on Quantum Zeitgeist for the latest breakthroughs in qubits, hardware, algorithms, and industry deals.

Avatar of Rusty Flint

Rusty Flint

Rusty is a quantum science nerd. He's been into academic science all his life, but spent his formative years doing less academic things. Now he turns his attention to write about his passion, the quantum realm. He loves all things Quantum Physics especially. Rusty likes the more esoteric side of Quantum Computing and the Quantum world. Everything from Quantum Entanglement to Quantum Physics. Rusty thinks that we are in the 1950s quantum equivalent of the classical computing world. While other quantum journalists focus on IBM's latest chip or which startup just raised $50 million, Rusty's over here writing 3,000-word deep dives on whether quantum entanglement might explain why you sometimes think about someone right before they text you. (Spoiler: it doesn't, but the exploration is fascinating)

Latest Posts by Rusty Flint: