2

Imagine that prior to the time of Copernicus, scientists could run machine learning algorithms (somehow).

Would it be possible to discover the truth of the geocentric model with their use?

Would the machine learning algorithms assist in discovering Kepler's laws of planetary motion?

Would they be beneficial for Newton in his discovery of the three laws of motion?

The broad goal of this question is to discern how ML techniques are helpful in the scientific process.

It is inspired by this stackexhcange reply.

kjetil b halvorsen
  • 77,844
Sam
  • 537
  • 1
    I very much doubt it, statistical/ML models are geocentric models in a way - all they do is function approximation. – user2974951 Nov 23 '23 at 09:36

1 Answers1

2

The discovery of physical laws can be aided by ML

While ML models can, at this point, not ask the right questions or collect the data etc., it is absolutely possible to use machine learning to aid the (re-)discovery of natural laws.

One method that's frequently brought up in this context is symbolic regression, which optimizes model fit to the data by combining well-understood mathematical functions. For an example have a look at this article applying symbolic regression to re-discover Kepler's law on planetary motion. A search on google scholar should give many more methods using Kepler's data and other classical physics-related examples.

Apart from that, ML tools are useful in all kinds of function approximation tasks that arise in the context of physics research. For some current examples have a look at this article discussing the usefulness of CNNs for gravitational wave detection, or this article using autoencoders in a similar context.

Scriddie
  • 2,244
  • 6
  • 13
  • 1
    Rediscovering is quite different from discovering or aiding in discovering. One might say that the trick is setting up the right starting environment by using domain knowledge of an already-known process, and only then let the model crunch out the numbers. – user2974951 Nov 23 '23 at 11:17
  • 1
    I agree that the trick is in setting up the problem correctly for ML to be useful, and it's not ML discovering physics somehow all by itself. Nonetheless, with the ubiquity of function approximation, I think it's more than fair to say that ML techniques can aid the discovery of physical laws. I've updated the answer with two examples from current research. – Scriddie Nov 23 '23 at 11:42