Physics specialists from Peking University have created an artificial intelligence system called AI-Newton, capable of independently deriving fundamental laws of physics based solely on empirical data. This AI managed to recreate Newton’s Second Law, which establishes the relationship between force, mass, and acceleration, accumulating understanding in a manner similar to a human scientist. Artificial intelligence independently “discovered” Newton’s Second Law Researchers presented the AI with data obtained from physical investigations of systems exhibiting pendulum oscillations to determine if the system could derive key physical principles. stefilyn/Getty The method, known as symbolic regression, is a machine learning technique aimed at finding mathematical expressions that best represent the relationship between variables in a dataset. Unlike general-purpose neural networks, which function as a “black box,” symbolic regression generates formulas understandable to humans. This approach constructs an equation that best fits the observations by iterating through combinations of mathematical operations (such as addition, multiplication, exponentiation, taking the logarithm). For this reason, this method is highly valuable in scientific research where the interpretability of the results is critically important. Most existing AIs can identify patterns in information and predict outcomes, but they are incapable of formulating comprehensive scientific concepts, such as the laws of gravity. The developed innovation reverses this situation. “This system simulates the human research process, sequentially building a knowledge base consisting of concepts and established laws,” explains Yan-Qing Ma, a physicist at Peking University and one of the creators of this development. The AI’s ability to discover significant concepts opens the prospect for scientific discoveries without prior programmer intervention. AI-Newton utilizes symbolic regression: the model searches for the most suitable mathematical expression to adequately describe physical phenomena. The research team trained the system on the results of 46 simulated experiments involving sphere motion, spring action, object collisions, oscillations, and pendulums. Test data intentionally included errors to simulate real measurement conditions as closely as possible. The collected results are presented as a preprint on the arXiv platform. Can AI operate at the level of a human physicist? Final testing results. The capabilities of AI-Newton were tested on a set of 46 experiments related to Newtonian mechanics. To better illustrate, some of the most challenging configurations have been omitted in this image. AI-Newton successfully identified fundamental generalized laws, including the law of conservation of energy and Newton’s Second Law, although with a constant discrepancy—by a factor of 2. Final testing results. The capabilities of AI-Newton were tested on a set of 46 experiments related to Newtonian mechanics. To better illustrate, some of the most challenging configurations have been omitted in this image. AI-Newton successfully identified fundamental generalized laws, including the law of conservation of energy and Newton’s Second Law, although with a constant discrepancy—by a factor of 2. AI-Newton processed time-series data on a sphere’s position and managed to derive an equation describing its velocity. The system saved this knowledge and applied it in subsequent tasks, where it successfully determined the sphere’s mass by appealing to Newton’s Second Law. This is undoubtedly a significant breakthrough. However, in another study, scientists from Harvard and the Massachusetts Institute of Technology (MIT) investigated the capabilities of large language models such as GPT and Claude on similar tasks. These models learned to accurately predict planetary trajectories (Kepler’s laws), but it turned out they were unable to derive Newton’s law of universal gravitation itself, which determines these orbits. Instead of the law of gravitation, the models generated formulas devoid of physical meaning. Keyon Vafa, a computer science specialist, notes: “A language model tuned to predict the outcomes of physical experiments is incapable of encoding concepts in the simplest and most concise way. Instead, it finds some completely non-human method for approximating physical solutions.” To achieve truly autonomous scientific discovery, AI must be integrated into all stages of research—from selecting a relevant problem and planning experiments to verifying hypotheses; however, a long way must still be traveled to achieve this level.
About the Author
