Deep learning as a tool for physics
Aqacia originally got its start optimising the complex world of experimental physics, in particular quantum optics. As part of the Australian Research Council Center for Quantum Computation and Communication Technology, the ANU quantum optics group led by Prof. Ping Koy Lam and Prof. Ben Buchler was the origin of the technology that drives Aqacia today. The research group focused on three main research themes, optical quantum memory, quantum information and optical levitation. The optical quantum memory program, spearheaded by Prof. Buchler, was where Dr Aaron Tranter developed the deep learning technology during his PhD.
Select Research Catalogue
Multi-parameter optimisation of a magneto-optic trap – (2018)
Seminal research paper outlining the application of the deep learning technology to the magneto-optical trap setup in the lab at the Australian National University.
Machine learner optimization of optical nanofiber-based dipole traps – (2022)
Second application of the DLO to the optimisation of an experimental system. A collaborative effort between the researchers at the Okinawa Institute of Science and Technology and the Australian National University.
Deep learning optimiser
The deep learning optimiser (DLO) was developed with one goal in mind: to optimise high dimensional physical systems. The ANU quantum optics group was working on optical quantum memories as a way to store quantum information carried by light. The platform of choice is something called a magneto-optic trap, a device capable of trapping and cooling atoms down to only a few hundred micro-kelvin! To get the best performance out of this device you need lots of atoms, something on the order of 10 billion.
The problem comes from controlling this device as it is very complicated. In fact, depending on how you define it there is a continuous space of possible control schemes, leading to an infinite dimensional problem! Clearly, this is too big. Instead, we decided the best thing to do would be to divide the space into small but manageable steps. Even taking relatively large steps we still ended up with 63 independent controls, too large for a human to stand there changing values.
This is where the DLO came in, a specialised tool designed specifically to operate with high dimensional and complex systems. The DLO works by using the power of artificial neural networks (ANNs) to provide an approximate representation of the landscape. In doing the DLO learns what controls work the best by continually predicting the best parameters and then subsequently testing them on the system. Applying this technique we were able to improve the setup to trap double the number of atoms in half the time.
Since then, the DLO has been applied to a number of problems including optical nano-fibre traps, single electron transistors, single photon generation and a myriad of other complex problems. The original research paper is open access and published in Nature Communications.
