Our group has recently started using deep learning to optimise the performance of our quantum-information experiments. The problem we tackled first was improving the number of atoms captured and cooled in a magneto-optic trap (MOT). Even with decades of research and engineering there is no proven recipe to maximise atom number and minimise temperature. Improving the MOT has mostly been limited to intuition based on approximations and trial-and-error.
Our machine learning system was found to catch and cool twice as many atoms in half the time compared to the best human effort. The solution found by the machine was also unexpected as it pushed the magnetic and optical fields around rapidly during the experiment in a way that no sensible human would have tried. (Reseach article here.)
Buoyed by this success we are focussed on a couple of questions:
1) What is the physics behind the machine learning MOT solution?
2) What else can be improved by our machine learning system?
The answer to the first question is going to be hard to find since the regime in which the atoms are cooled and trapped is not amenable to any sort of modelling owing to the number of atoms and the complex interactions during the trapping phase of the experiment. By exploring the machine learning solutions we may, however, be able to find important timescales and relate them to known atomic parameters and in this way we could gain some new insights into the physics of this system.
The answer to the second question is “lot of things”. We plan to implement our machine learning system on all kinds of physical systems, both in the quantum and classical regimes. This could include applications to optical quantum memory, atomic magnetometry, 2-qubit gates and much, much more.