A team of researchers in China has developed an AI system called AI-Newton that can autonomously discover fundamental principles of physics from experimental data—without human guidance. The model, designed to mimic the way humans develop scientific understanding, was trained on data from 46 simulated physics experiments involving pendulum-like motion, oscillations, springs, collisions, and the free movement of objects. These experiments included realistic noise and statistical errors to reflect real-world measurement limitations. The system uses a method known as symbolic regression, which searches for the most accurate mathematical equations to describe physical phenomena. Rather than simply fitting data, AI-Newton builds a knowledge base step by step, identifying key concepts like velocity and mass, and then using them to derive more complex laws. In one test, the model was given data on a ball’s position over time and successfully deduced the equation for velocity. It then used that knowledge, along with Newton’s second law, to calculate the ball’s mass—demonstrating a form of logical progression similar to human scientific reasoning. Yan-Qing Ma, a physicist at Peking University and a key developer of the model, says the breakthrough lies in AI-Newton’s ability to generate and store useful scientific concepts, enabling it to make discoveries without pre-programmed rules. The results, while not yet peer-reviewed, suggest a new path for AI to contribute to scientific discovery by identifying not just patterns, but underlying laws. In contrast, earlier attempts to use AI for physics have had limited success. In 2019, researchers at ETH Zurich created “AI Copernicus,” a neural network that could predict planetary orbits from Earth-based observations. However, the model required human experts to interpret the resulting equations and link them to real physical laws. Similarly, a 2023 study by researchers at MIT and Harvard tested large foundation models like GPT, Claude, and Llama on predicting planetary motion. When trained on orbital data, these models could only predict trajectories and failed to derive meaningful laws of gravity. Instead, they produced nonsensical or inconsistent force equations. AI-Newton’s approach stands out because it doesn’t just predict outcomes—it learns to construct scientific concepts and apply them iteratively. This capability could one day allow AI to assist in uncovering new physical principles, especially in complex systems where human intuition may be limited. While still in early stages, the model represents a significant step toward AI that doesn’t just analyze data, but actively participates in the scientific process.