Finally, AI Does Science on Its Own

The rise of “Artificial Scientists” marks a new era for humanity.

Younes Henni, PhD
3 min readOct 14, 2021
Image from Wikimedia Commons.

“Such an experiment wouldn’t exist if the AI hadn’t found it.”

That’s what a group of physicists at Toronto University in Canada concluded.

Doing science has always been a skill unique to humans. Of course, companies like Facebook, Google, and Netflix built AIs that could:

  • Tell dogs from cats.
  • Beat world champions in many games.
  • And read medical scans better than any doctor.

But MELVIN (the codename of a new AI software) is more special. It decides how to run lab experiments. And not only any lab experiments but quantum physics experiments.

If you think high-school physics is hard. Quantum physics is 1000 times more complicated. For an AI to “think” and “help” humans do quantum physics is a breakthrough many have waited for.

This remarkable discovery happened almost by chance. Physicists were trying to “create” complex structures using particles of light (aka photons). Sometimes, when two photons (or more) talk to each other, they act as a single structure. If you wiggle one photon, the other(s) would respond instantly. In physics, this phenomenon of “connected structures” has the ugly name: entanglement.

Entangled particles are the building blocks of quantum computers. Understanding them, handling them, and making them on demand is critical to mass-producing “personal quantum computers”.

Sadly, entanglement is hard to make. The more photons you want to “entangle”, the harder it gets.

Before MELVIN, researchers spent months trying to entangle three and four photons. They designed crazy-looking experiments with cutting-edge equipment: lasers, spectrometers, holograms, and so on. But each time, they failed. “I thought, ‘This is absolutely insane. Why can’t we come up with a setup?’” recalls one of the scientists.

One day, the researchers had a light-bulb moment. What if we wrote a program that takes a goal as input and spits out the steps needed to achieve that goal. MELVIN was born.

In a few hours on its first try, MELVIN found the answer that escaped scientists for months. “That was a crazy day. I could not believe that it happened,” recalls one of the scientists. MELVIN would tell the physicists what to do and how to do it, and they would simply follow its steps.

Some of MELVIN suggestions would puzzle his creators. One day, it asked them to place crystals (like diamond or sapphire) somewhere unusual. In all their experiments, the physicists always put the crystals at the start of the setup. It is from these crystals that the entangled photons would come.

The result of the trick was no less than incredible. The physicists not only entangled four photons, but the final structure had the highest degree of entanglement ever.

“In all my decades of research, no human dreamt such an idea,” recalls one researcher. He adds, “It’s a gorgeous example of the explorations these thinking machines can take us on.”

Algorithms like MELVIN are still far from having truly unique ideas or building theories from scratch. Still, “I do think one day they’ll get there.” Concludes one physicist.

MELVIN is helping physicists do science experiments in ways they never thought of. Imagine what its AI successors would do.

To read more about MELVIN, check the study on ArXiv.



Younes Henni, PhD

Physicist • Soft Dev • ☕ Junkie • I bring you the latest in science, tech, health, economics & personal growth. To read all: