In many quantum physics experiments, electromagnetic fields are used to hold and manipulate particles. However, their fine adjustment has so far been very time-consuming. Physicists have therefore developed an adaptive artificial intelligence that can take over this task: The tailor-made neural network is trained with earlier variants of such experiments and learns which physical laws apply and how changes in the fields affect particle behavior. Initial tests showed that the AI system can correctly predict and reproduce the settings and results of such quantum physics experiments. Artificial intelligence could thus also facilitate research in quantum physics, as the scientists explain.
Quantum particles can be manipulated with electromagnetic fields: they can be captured, held or moved to a specific location. This is important, for example, for their use in quantum computers or in quantum physical measurements. "To control quantum particles, we use a combination of several electromagnetic fields," explains co-author Maximilian Prüfer from the Vienna University of Technology. “Electric current is sent through tiny structures, creating a magnetic field. In addition, we use light beams that can be manipulated in a targeted manner using lenses, mirrors and filters.” The shape and intensity of the light beam determine which forces the particles feel at which point. By adjusting the intensity distribution of the light, the particles can be specifically influenced.
Elaborate adjustment
However, it is not easy to determine what shape the electromagnetic particle traps should have for the desired effect and how to change them during the experiment. So far, this has usually involved lengthy test series with numerous measurements. "In principle, there are two different methods of controlling this light field," explains Prüfer. "You can calculate in advance what shape the field needs to have - but this is only possible if you really know all the details of the experiment, including all interfering effects. The result can therefore only be as precise as the calculation model used.” The second option is iterative control algorithms: a new experiment is carried out after each change step and the result is used to optimize the arrangement.
“In principle, such algorithms are only limited by the experimental measurement accuracy. However, this wonderful property has a price: Each improvement step requires its own attempt at the experiment,” explains co-author Andreas Deutschmann-Olek from the Vienna University of Technology. As a result, the measurements required for such a series of experiments can take weeks, and even a small change in the desired light field means that you have to start over. That is why the team led by first author Martino Calzavara from Forschungszentrum Jülich has now turned to artificial intelligence for help. "We have developed a neural network whose structure is precisely adapted to the physical task that needs to be solved here," explains Prüfer. "It was important to use our knowledge of the physical properties of the system and incorporate it into the artificial intelligence from the outset. We call this a physics-inspired neural network.”
AI system learns the principles of experiments
The research team trained its adaptive AI system with digital variants of the experiment based on all previous test data. A camera is used to measure where the particles are located. The neural network is trained with these images. Over time, it learns which changes to the experiment affect the quantum particles in which way - without the physical formulas that describe this relationship having been programmed into it beforehand. As a result, the artificial intelligence independently develops a certain “understanding” for the principles of the experiment. This is comparable to the ability of a human being to recognize which object or animal is depicted on a drawing based on just a few sketchy lines: the AI only needs a relatively small amount of information after training to be able to control the desired experiment to determine.
"We were able to show that the artificial intelligence actually learns to correctly imitate the behavior of the physical system," says Prüfer. In this way, the algorithms can try out at lightning speed how various changes to the experiment affect the current situation, without the need for long, time-consuming series of experiments. "The information collected from previous experiments is stored in a structured manner in the neural network and can thus be transferred to new situations," explains Deutschmann-Olek. Where in the past it might have taken a hundred experiments to find the right settings, today a small fraction of that is enough. It is now possible to carry out a large number of experiments that previously would only have been possible with much greater effort or not at all. "The use of machine learning in quantum physics research is currently on the rise," explains Prüfer. "We hope that our work will also provide insights into how a physical understanding together with the well-developed AI methods can improve experiments."
Source: Martino Calzavara (Jülich Research Center) et al., Physical Review Applied, doi: 1103/PhysRevApplied.19.044090