The quantum world defies logic: wrap your brain around instantaneous messaging between distant particles, or cats that are alive and dead at the same time.
- Human intuition is often used to optimise laboratory experiments that require trapping cold atoms
- A neural network came up with a non-intuitive way to double the number of captured atoms
- The underlying physics behind the technique and why it works is a mystery
Perhaps physicists should leave human intuition at the laboratory door when designing quantum experiments too.
An Australian crew enlisted the help of a neural network — a type of artificial intelligence — to optimise the way they capture super-cold atoms.
Usually, physicists smoothly tune lasers and magnetic fields to gradually coax atoms into a cloud, according to study co-author Ben Buchler from the Australian National University.
“It’s like a slow hug,” he said.
Much to the team’s surprise, the neural network came up with a more dynamic technique: it pushed atoms around quickly by rapidly changing the laser and field strength.
This more vigorous approach, published in Nature Communications, trapped twice as many atoms in half the time when compared to traditional methods devised by humans.
Why trap cold atoms?
In the past few decades, cold clouds of atoms have formed the foundations for advances in precision measurement, optical atomic clocks and quantum processing.
And in many cases, the colder the cloud, the better. That’s because a warm atom is a jiggly atom, and this poses a problem for physicists.
When atoms interact with one another, they create noise in the system.
And when you’re talking about clouds of billions of atoms, that can be a lot of unwanted noise.
So physicists get their atoms to literally chill — and stop moving around so much — using lasers, said Aaron Tranter, a PhD student at the Australian National University and study co-author.
Picture an atom moving towards a laser. When photons hit the atom, they exert a tiny amount of force on it.
This is a bit like a bunch of ping pong balls bouncing off a rolling bowling ball. With enough ping pong balls, the bowling ball will come to a stop.
Laser cooling can chill atoms to a fraction over absolute zero (minus 273 degrees Celsius).
Next, physicists must collect as many atoms into a cloud as possible. For experiments that rely on cold atoms absorbing and emitting light, a bigger cloud means a stronger signal, Mr Tranter said.
The problem is finding the best way to adjust the lasers and magnetic fields to achieve this.
Artificial intelligence over human logic
Fine-tuning these experiments usually falls to a PhD student, who might spend years painstakingly tweaking various parameters for only marginal gains.
In this study, the researchers looked to much faster artificial intelligence to optimise their work.
What are neural networks?
What are neural networks?
- Neural networks are a type of artificial intelligence
- They comprise computing systems inspired by the way the brain processes information
- Information moves through a series of connected layers of algorithms called nodes
- The more layers, the “deeper” the network
- Neural networks are trained by feeding them many examples of the task we want them to learn
- The network adjusts the strength of the connections between the nodes until it learns to extract important information
- Eventually, the neural network becomes adept at that particular task
So when Mr Tranter, Dr Buchler and their team fed training data to a neural network and ran it almost 600 times, a radically different method popped out.
Rather than a slow embrace, the neural network devised a way that involved fast changes in the laser and magnetic fields, shoving the atoms more actively towards the centre of the trap.
“We ran it over and over again. But this vigorous corralling was a constant feature of the system,” Dr Buchler said.
And it worked a treat, capturing more atoms in less time.
The neural network the researchers used isn’t specific to atom traps — or even physics experiments, Mr Tranter added.
“It could also be used to improve any system that is traditionally tuned by tedious trial and error … [such as] industrial and manufacturing processes.”
Tyler Neely, a physicist at the University of Queensland who wasn’t involved in the study, said using a neural network is powerful “because it doesn’t care about the actual mechanisms that allow it to get the result”.
Dr Neely intends to use a neural network in his own cold atomic experiments.
“We’d like to have atoms spinning in a ring, in a ‘ring trap’,” he said.
“The way you spin it up is, again, one of those things where you can do it a multitude of ways.
“So the idea is that we’re going to implement the algorithm to find the best way of doing that because it should be much more efficient.”
With unexpected results come questions
This raised the question: why do cold clouds respond so well to constant corralling?
“We don’t know for sure why it works,” Mr Tranter said.
“The next step is to try to understand why it works so well.”
This is easier said than done. Even though the atoms are cold, there are still lots of them.
“And where you have a lot of atoms, it’s very difficult to model the way they interact with each other,” Mr Tranter said.
For Dr Neely, the mysterious physics behind the neural network’s technique “is one of the other interesting things about this work”.
“Because it turns out that this crazy way works so well, it means something is going on with the atoms that we currently don’t understand.”
This content was originally published here.