AI-Newton reconstructs the concepts behind classical mechanics

A curve can fit one experiment without explaining the next. AI-Newton is designed to make the harder step: propose concepts that connect several sets of measurements, then use them to build more general physical laws.[1][2]

The system recovered Newton’s second law, energy conservation and universal gravitation from simulated mechanics data with statistical errors, according to a paper accepted by Physical Review Research on 9 October 2026. An earlier version appeared on arXiv in April 2025. The test uses familiar laws as an answer key for assessing how a machine builds transferable explanations.[1][2]

That makes the work a study of scientific reasoning. The achievement lies in linking experiments through interpretable concepts, rather than returning a separate mathematical fit for each dataset.

A formula is not yet an explanation

Suppose an algorithm watches a succession of pendulum experiments and finds a highly accurate polynomial fit. It may reproduce those measurements extremely well while failing to describe a satellite or a falling stone. A transferable physical law requires the right variables and relationships, not just numerical fit. In human science, concepts such as force, mass and acceleration provide common language across different phenomena.[1][2]

The AI-Newton approach uses a knowledge representation organised around physical concepts and an iterative process that proposes and generalises them. The public authors describe it as operating without supplied prior physical knowledge. That claim needs precise framing: the programme does not start without any design assumptions at all. Its representation, search strategy and notion of explanatory simplicity are created by researchers. What is withheld is a list of the particular physics laws it is asked to recover.

The experiments used data generators that simulate physical systems and add statistical errors. This lets the researchers test whether the concepts remain recoverable when measurements are imperfect, while retaining a known answer for evaluation. The result establishes performance in those simulated settings; applying the system to measurements from physical instruments would introduce additional sources of error and unmodelled influences.[1][2]

Moving from an equation to a reusable concept

Newton's second law, conventionally summarised as force equals mass times acceleration for a constant-mass system, can connect many ordinary motion experiments. Conservation of energy is a broader organising principle, although its expression depends on the system and which exchanges count as part of the energy ledger. Universal gravitation links the motion of distant bodies to mutual gravitational attraction.

These relationships provide clear tests because their meaning is already understood. A system can be examined for whether it recovers the right quantities and extends them across the mechanics experiments, instead of being rewarded only for matching a curve.[1][2]

Where concept discovery could help

Researchers sometimes build known conservation or symmetry constraints directly into a model. AI-Newton explores the complementary problem of recovering useful concepts from data. The first approach exploits established knowledge; the second could help organise a domain where an adequate explanation is still missing.[1][2]

The decisive next test is whether the learned concepts produce useful predictions beyond the examples that gave rise to them. That is where a collection of fitted expressions begins to earn the status of a physical explanation.

Report an error or suggest a correction