AI-Newton: Новая эпоха в открытии физических законов с помощью искусственного интеллекта AI-Newton: A New Era in Discovering Physical Laws with Artificial Intelligence

A team of researchers from China has created an AI system named AI-Newton, which can autonomously «discover» fundamental principles of physics, such as Newton’s second law, after processing experimental data. This information was reported by Nature.

The model mimics the scientific process undertaken by humans by gradually building a knowledge base of concepts and laws. According to physicist Yangqing Ma from Peking University, this capability could pave the way for scientific breakthroughs without requiring prior human programming.

Kayon Wafa, a scientist from Harvard University in Cambridge, explained that AI-Newton employs a technique called «symbolic regression.» This method seeks to identify the most suitable mathematical equation to represent physical phenomena.

This technique is viewed as a promising avenue for scientific discoveries, as the system is designed to encourage the derivation of concepts.

The Peking University team utilized a simulator to generate data from 46 physical experiments involving the free motion of balls and springs, collisions between objects, and the behavior of systems that exhibit vibrations, oscillations, and pendulum movements.

The simulator intentionally introduced statistical errors to simulate real-world data.

AI-Newton received data regarding the position of a ball at a specific moment and was tasked with formulating a mathematical equation that would explain the relationship between time and position.

The neural network successfully generated an equation for velocity and retained the knowledge for a subsequent set of tasks, in which it was required to calculate the mass of the ball using Newton’s second law.

The results have not yet undergone peer review.

Previously, researchers have utilized AI models to predict planetary orbits.

In 2019, scientists from the Swiss Federal Institute of Technology in Zurich developed AI Copernicus — a neural network designed to derive planetary trajectory formulas based on observations from Earth.

Wafa and his colleagues from the Massachusetts Institute of Technology in Cambridge conducted a similar experiment with various foundational models such as GPT, Claude, and Llama.

These models were trained to predict the positions of planets in solar systems and were then asked to forecast their movement paths.

The neural networks were trained on orbital motion data. However, they could not apply their knowledge to accomplish any other tasks beyond calculating planetary trajectories. When attempting to convert the information into a law governing the behavior of forces, the models produced an irrelevant law of gravity.

«A language model trained to predict outcomes of physical experiments will not program concepts in a straightforward and concise manner. It will find a fundamentally non-human approach to reach physical solutions,» Wafa stated.

David Powers, a computer and cognitive science expert from Flinders University in Adelaide, Australia, pointed out that models capable of deriving scientific laws are beneficial. However, for autonomous discoveries, AI needs to participate in other phases of the project: identifying problems, determining necessary experiments, analyzing obtained data, and formulating hypotheses.

«Experimental science involves identifying relevant variables and conducting systematic experiments to gather data and test predictions,» the expert noted.

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