Chinese Scientists Develop AI Model to Uncover Physical Laws

Chinese scientists create AI model to uncover physical laws.

A team of scientists from China has developed an AI system called AI-Newton, capable of independently “discovering” key principles of physics, such as Newton’s second law, after receiving experimental data. This was reported by Nature.

The model emulates the human scientific process—gradually building a knowledge base of concepts and laws. This skill could potentially lead to scientific discoveries without prior human programming, noted Yanqing Ma, a physicist from Peking University.

Keon Wafa, a scientist from Harvard University in Cambridge, explained that AI-Newton uses an approach called “symbolic regression.” It searches for the best mathematical equation to represent physical phenomena.

This technique is considered a promising method for scientific discoveries, as the system is programmed to encourage the derivation of concepts.

The team from Peking University used a simulator to generate data from 46 physical experiments related to the free movement of balls and springs, collisions between objects, and the behavior of systems demonstrating vibrations, oscillations, and pendulum movements.

The simulator intentionally added statistical errors to mimic real data.

AI-Newton received information about the position of a ball at a specific moment in time and was tasked with formulating a mathematical equation to explain the relationship between time and position.

The neural network was able to form an equation for velocity and retained the knowledge for the next set of tasks, which involved calculating the mass of a ball using Newton’s second law.

The results have yet to undergo peer review.

Planetary Trajectories

Previously, scientists have used AI models to predict planetary orbits.

In 2019, researchers from the Swiss Federal Institute of Technology in Zurich developed AI Copernicus—a neural network for deriving formulas of planetary trajectories based on Earth observations.

Wafa and his colleagues from the Massachusetts Institute of Technology in Cambridge conducted a similar experiment with several basic models like GPT, Claude, and Llama.

They were trained to predict the position of planets in solar systems and then asked to predict their paths of movement.

The neural networks were trained on orbital motion data. They were unable to apply the knowledge to any tasks other than calculating planetary courses. When attempting to transform the information into a law of force behavior, the models derived an unnecessary law of gravity for the task.

“A language model trained to predict the results of physical experiments will not program concepts in a simple and concise way. It will find some completely non-human approach to approximate physical solutions,” said Wafa.

David Powers, a specialist in computer and cognitive sciences from Flinders University in Adelaide, Australia, noted that models capable of deriving scientific laws are useful. However, for autonomous discoveries, AI needs to participate in other project stages: identifying problems, determining necessary experiments, analyzing obtained data, and creating hypotheses.

“Experimental science involves identifying variables of interest and conducting systematic experiments to obtain data and test predictions,” said the expert.

In March, researchers from the UK and Canada developed an AI model, Aardvark Weather, for weather forecasting.