Applying Machine Learning to stirring a quantum fluid
A machine learning breakthrough redefines control of Bose-Einstein Condensates (BECs), the quantum state frequently termed the fifth state of matter.
A groundbreaking study led by EQUS PhD student Simeon Simjanovski (UQ) demonstrated a novel approach to controlling quantum superfluids, utilising machine learning to fine-tune fluid currents in a toroidal Bose-Einstein Condensate.
Superfluids and BECs: stirring exotic quantum materials
Superfluids and Bose-Einstein Condensates (BECs) are fascinating states of matter that exhibit quantum properties on a macroscopic scale. These fluids can flow without resistance under certain conditions, making them ideal systems for studying fundamental quantum phenomena.
And when confined to a toroidal (i.e. doughnutshaped) container, BECs reveal even more intriguing properties. Their flow, known as a ‘persistent current’, is quantised—restricted to specific, discrete speeds due to the nature of their quantum wavefunction.
Simeon explains, “You can think of it like water swirling around in a circular swimming pool that only flows at certain fixed speeds. The challenge is figuring out how to stir the system to achieve a specific flow speed without introducing unwanted oscillations or disturbances.” This study aimed to address that challenge using advanced machine learning techniques.
Control matters: ‘perfect’ experiments and future technologies
Precise control of superfluid currents has significant implications for both fundamental research and practical applications. A clean, controlled flow is essential for creating reliable initial conditions in experiments, enabling physicists to subsequently explore complex quantum behaviours free of interference from ‘messy’ starting conditions.
Additionally, such control would be pivotal for developing superfluid-based atomtronics, ultra-sensitive rotation sensors, or quantum computing components.
Improving on human intuition with machine learning
Traditionally, determining how to stir a superfluid to achieve a desired flow required trial-and-error experimentation.
This time-consuming approach is difficult to optimise, requiring systematically ‘walking’ the entire parameter space of the experiment – i.e. exploring the full range of variables such as stirring speed and intensity – to identify optimal conditions to achieve clean, stable flows.
To overcome these challenges, the EQUS team turned to machine learning, which could find the best solution from a limited ‘experience’.
“Machine learning took over the drudgery of that trial and error,” says collaborator Prof Matt Davis, an EQUS Chief Investigator. “We programmed the system to test different stirring techniques and evaluate their effectiveness, allowing the computer to find the best solutions.”
The surprising effectiveness of a swift kick
One of the team’s counterintuitive findings was the surprising effectiveness of multiple stirring strategies. Rather than there being one universal ‘best’ way to stir a system, the machine learning system would identify a new optimum each time.
“One of the most surprising results was that the optimal stirring method wasn’t always gentle or gradual,” corresponding author A/Prof Tyler Neely, EQUS Associate Investigator (UQ), notes. “A quick and aggressive ‘kick’ was the best way to achieve the desired flow!”
This discovery underscores the power of machine learning to identify solutions that could elude even the most experienced researchers.
Applications in future technologies, and ‘cleaner’ research
The EQUS study represents a milestone in the quest to harness the peculiarities of quantum mechanics for practical use. By applying machine learning to the control of superfluid currents, the team have opened new avenues for exploration in both applied quantum technologies and fundamental physics.
The work could pave the way for more robust quantum sensors that require precisely tuned initial states. Such sensing devices could revolutionise fields like navigation, geology and medical imaging by offering unprecedented sensitivity and accuracy.
In addition, the methods developed in this study could serve as a foundation for more complex quantum experiments. Simeon points out that their work focused on a relatively simple system – a single toroidal BEC. “Future studies could involve more intricate geometries, like multiple rings or interconnected networks,” he says.
“Each new configuration presents its own challenges and opportunities for discovery.”
Such clean, controllable systems can serve as simulators for complex solid-state systems, enabling clearer insights into their quantum mechanical behaviour.
Generating precise initial states is a cornerstone for advancing EQUS’ Quantum Engines and Instruments program’s mission of developing scalable quantum technologies.
“This work shows how machine learning can make experimental protocols more efficient and robust. It’s not just about optimising experiments; it’s about rethinking how we approach control in quantum systems,” says one of the authors, EQUS Chief Investigator Prof Halina Rubinsztein-Dunlop (UQ).
This story was first published in the 2024 EQUS annual report.