Physicists Utilize AI for Rapid, Precise Control of Electromagnetic Fields in Quantum Research
May 9, 2023 -- In quantum research, customized electromagnetic fields are used to precisely control particles. Physicists at Forschungszentrum Jülich and TU Wien have now shown that machine learning is an excellent way to do this.
This project required close cooperation between theoretical physics and experimental physics. The two first authors Martino Calzavara (1st from right) and Yevhenii Kuriatnikov (2nd from right) discuss final details of the setup with Maximilian Prüfer (far lleft). Credit: TU Wien.
Tiny particles can be manipulated with electromagnetic fields: They can be captured, held in place or moved to a specific location. However, it is difficult to find out what shape these electromagnetic fields should have exactly and how they should be controlled during the experiment. This often requires lengthy series of experiments with numerous measurements.
Scientists at Forschungszentrum Jülich and TU Wien have now been able to show that this task can be completed much faster than before with the help of learning algorithms - and with the same precision. To do this, teams in Jülich and Vienna developed a neural network tailor-made for exactly this application with the fastest possible learning curve. The result was published in the scientific journal "Physical Review Applied" and will now be used in very different quantum experiments.
The precise control and manipulation of optical fields is needed in many research areas, with applications ranging from microscopy to quantum simulators. So-called iterative control algorithms are used to improve the light field step by step and come as close as possible to the target. In the process, a new experiment is carried out after each change step. The necessary series of experiments can take weeks, and a slight change to the desired light field means that one has to start all over again.
Artificial intelligence (AI) has now been used for precisely this task. The AI learns to correctly imitate the behaviour of the physical system. Thus, the algorithms can try out at speed how different changes to the experiment will affect the current situation, without the need for long, elaborate series of experiments. This means that a large number of experiments can now be carried out that would previously only have been possible with much greater effort or not at all.
Further Details
Researchers led by Tommaso Calarco from the Peter Grünberg Institute at Forschungszentrum Jülich, together with Jörg Schmiedmayer's research group at the Vienna Center for Quantum Science and Technology (VCQ) at TU Wien, have developed a neural network whose structure is precisely adapted to the physical task that needs to be solved. Only with this physics-inspired neural network was it possible to obtain excellent predictions by the neural network with experimentally manageable amounts of data.
The strategy proved successful: a camera was used to measure the light field, and the neural network was trained with these images. Over time, it learned which changes in the experiment affect the quantum particles and in what way - without having to program the physical formulas that describe this relationship. In a certain sense, the artificial intelligence develops a kind of "knowledge" of the system.
Original Publication
Optimizing Optical Potentials With Physics-Inspired Learning Algorithms
M. Calzavara, Y. Kuriatnikov, A. Deutschmann-Olek, F. Motzoi, S. Erne, A. Kugi, T. Calarco, J. Schmiedmayer, M. Prüfer
Phys. Rev. Applied (28 April 2023), DOI: 10.1103/PhysRevApplied.19.044090
Source: Jülich Supercomputing Centre
