A Chinese AI model taught itself basic physics — what discoveries could it make?

Majority artificial intelligence Models (AI) can reliably identify patterns in data and make predictions, but have difficulty using that data to develop broad scientific concepts such as the laws of gravity. Now a team in China has developed a system called AI-Newton, which, after receiving experimental data, can autonomously “discover” key principles of physics, such as Newton's second law, which describes the effect of force and mass on acceleration.

The model mimics the human scientific process, gradually building up a knowledge base of concepts and laws, says Yan-Qing Ma, a physicist at Peking University in Beijing who helped develop the system. The ability to identify useful concepts means the system could potentially discover scientific discoveries without prior human programming, Ma adds.

Keyon Vafa, a computer scientist at Harvard University in Cambridge, Massachusetts, explains that AI-Newton uses an approach called symbolic regression, in which the model searches for the best mathematical equation to represent physical phenomena. This technology is a promising method scientific discoveryhe adds, because the system is programmed in a way that encourages it to derive concepts.

The Peking University team used the simulator to obtain data from 46 physics experiments.¹ including the free movement of balls and springs, collisions between objects, and the behavior of systems exhibiting vibrations, oscillations, and pendulum-like motion. The simulator also intentionally introduced statistical errors to simulate real-world data.

For example, A.I.-Newton was given data about the position of a ball at a given moment in time and asked to come up with a mathematical equation that explained the relationship between two variables, time and position. He was able to give an equation for speed. He saved this knowledge for the next set of problems, during which he successfully calculated the mass of the ball using Newton's second law. The results have not yet been peer reviewed.

Planetary trajectories

Scientists have previously used artificial intelligence models to predict the orbits of planets. In 2019, researchers from the Swiss Federal Institute of Technology (ETH) in Zurich developedA.I. Copernicus', a neural network that used ground-based observations to derive formulas for planetary trajectories. In this case, people were needed to interpret the equations and understand how they related to the motion of the planets around the sun.

Vafa and his colleagues at the Massachusetts Institute of Technology in Cambridge conducted a similar experiment with multiple baseline models, a type of AI model trained on large datasets including large language models such as GPT, Claude and Lama.

They trained models to predict the positions of planets in solar systems, and then asked them to predict the forces that govern the planets' trajectories. In preprint²The researchers showed that when the models were trained on orbital trajectories, they were unable to use the acquired knowledge to solve any problems other than predicting planetary trajectories. In an attempt to turn orbital trajectory data into a law of force behavior, basic AI models produced a meaningless law of gravity.