AI-Newton Learns Physics Laws on Its Own

Most artificial intelligence (AI) models can find patterns in data and
make predictions, but they have trouble using that data to create big
scientific ideas, like the laws of gravity. Now, a team in China has
built a system called AI-Newton that can study experimental data and
automatically discover important physics rules, such as Newton’s second
law, which explains how force and mass affect acceleration.
The model works somewhat like a human scientist. It slowly builds a
collection of ideas and laws, says Yan-Qing Ma, a physicist at Peking
University who helped create the system. Because it can find useful concepts
on its own, AI-Newton might one day make new scientific discoveries without
humans telling it what to look for.
Keyon Vafa, a computer scientist at Harvard University, says that AI-Newton
uses a method called symbolic regression. This means the system searches for the
best mathematical equation to explain how something in the physical world behaves.
This approach is useful for scientific discovery because it pushes the AI to
form clear concepts.
The Peking University team tested AI-Newton using a simulator that created
data from 46 physics experiments involving moving balls, springs, collisions,
and systems that vibrate or swing like a pendulum. The simulator also added
random errors to make the data more realistic.
For example, AI-Newton received data about a ball’s position at different
times and was asked to create a mathematical equation connecting time and
position. It successfully produced an equation for velocity. It then used
this new knowledge in later tasks, where it correctly figured out the mass
of the ball using Newton’s second law. These findings have not yet been peer
reviewed.