An AI reconstructed centuries of classical mechanics using only simulated experimental data, starting with zero physics knowledge on October 1, 2026.
AI-Newton began with only two variables—the position of objects and time—and no prior understanding of mass, force, energy, or gravity. Its training data contained no physics textbooks or established laws, leaving the system in total darkness regarding existing scientific theory.
The AI successfully built the framework of Newtonian mechanics by processing a series of simulated experiments. It first invented the concept of mass, then observed momentum, and identified the principle of energy conservation before deriving the law of universal gravitation.
AI-Newton Analyzes 46 Experiments With Gaussian Noise
Researchers provided AI-Newton with 46 sets of classical mechanics experiments, ranging from simple free-moving balls, spring vibrations, and collisions to complex systems involving projectiles, inclined planes, and multiple objects pulling on one another.
The team added Gaussian noise to the simulation data to force the system to find patterns within imperfect information. This data was generated via differential equations to mimic the imperfections of real-world measurements.
System Builds Concept and Law Libraries
AI-Newton accumulated knowledge in two internal repositories: a concept library and a law library, both of which were empty at the start. The AI defined velocity by analyzing changes in position over time and identified concepts like kinetic energy and elastic potential energy by extracting conserved quantities across different experiments.
The system then employed symbolic regression to determine the mathematical relationships between these concepts.
- Expansion: When a law failed to explain a new experiment, the AI attempted to add new terms to broaden the law’s applicability.
- Refinement: To prevent the knowledge base from becoming bloated, the system used a simplification algorithm based on differential algebra to remove redundant formulas.
The AI registered approximately 90 physics concepts and 50 general laws on average. A single round of this process typically required dozens of hours and thousands of attempts.
AI Derives Formulas Without Human Labels
AI-Newton produced mathematical formulas without human labels. The research team only identified the results as Newton’s second law
or energy conservation
after comparing the AI’s output formulas to known human physics.
This system reproduced discoveries through symbolic reasoning and iterative self-correction within a controlled simulated environment, whereas humans spent centuries extracting mass and force from observations of falling bodies and planets.
The ability to invent transferable concepts without human supervision suggests that artificial intelligence may eventually be able to derive fundamental laws of nature independently by analyzing complex data patterns within digital simulations.