Physcis doesn’t need to be about the real world
Nothing can confirm a scientific theory more than its predictive power. Indeed, if a theory can foresee the outcome of an experiment previously never realised, it proves it capacity to address the problem on a fundamental level. The process of elaborating such theory is rather mysterious, after all in some sense we end up with more information about the world than we started with . Even in the booming domain of deep-learning, which can be seen as a simplified prototype of human generalization process, it is not yet clear how and even if machine learning algorithms are able to generalize .
Despite this lasting mystery, many physical theories are strongly believed to have near absolute predictive power for the class of problems to which they apply. Notably, classical mechanics is generally seen as safe for prediction about macroscopic properties of objects. Starting from the hypothesis that a theory is correct, it is then natural to try to imagine some of its exotic consequences and implications. Such imaginative efforts lead for example to the Michelson and Morley experiment which failure ruled out the aether theory of light , a result now considered to be an important milestone in the history of physics.
This is the normal course of physics, what I want to talk about here is something else, something I would call “Fantasy Physics”: the application of well tested physics theory to absurd scenarios. Examples of such fantasies can be found Randall Munroe’s well known “What if?” series and are often invented and discussed at YPF events, most likely around 2 a.m. It is one more reason to sign up for the next YPF event if you haven’t already.
The exercise of solving improbable scenarios is not only enjoyable , it is also deeply instructive. It requires research, imagination and the ability to synthetize a vast knowledge of physics and select its most relevant part. One of the reasons that makes this exercise hard is that its answer does not exist yet, to have it one must elaborate it.
A good experience for sure, but worthless regarding the advance of science. Here is an example: imagine an advanced civilization of aliens, that for some arbitrary reason need a large amount of energy. Like a huge amount. They need so much that they decide to enclose a star in a gigantic sphere covered with some sort of solar panel, which allows to harvest a big fraction of the energy radiated from said star. It’s an interesting scenario, and using knowledge about gravity and materials, it is possible to deduce many properties of that sphere, one of them being that gravity would not maintain the star at its center, gravitational forces from the sphere being balancing inside it, and as such our aliens would probably choose another design for their massive power generator.
This is definitively food for thought for science fiction writers, but for scientist it is of little interest.
Or… is it?
A massive power harvesting structure of that sort would greatly change the spectrum of a star. If we can predict how the spectrum change, then we would be able to identify stars coated in such sphere. In fact, a megastructure of this magnitude is probably among the smallest proof of extra-terrestrial civilization we can currently observe from earth.
Obviously, I am not the first to think of that, and the idea of a spherical energy harvester around a star have been formalized more than 50 years ago by Freeman Dyson, who gave his name to hypothetical structure today known as “Dyson sphere”. Incidentally, the SETI program, which aims at finding extra-terrestrial intelligence, is looking for Dyson spheres and related megastructures.
Of course, the example of Dyson sphere is cherry picked, and most of the results we get from Fantasy Physics will be useless. But we should not consider Fantasy Physics to be one of the most complex useless activities, we should consider it to be one a the few enjoyable activities that may push science forward. The odd of this happening are of course very low, yet higher than for most alternatives.
 There is however nothing magical about the fact that humans are able to predict results beyond the actual data, we are simply sometimes right in-between our many errors. My favourite example of this effect is the one of Democritus who lived around 400 BC and predicted the existence of atoms, even though there would be no experimental evidence of their existence for more than two millennia.
But while Democritus claimed that matter was made of fundamental elements, Aristotle (among others) contradicted him by claiming that matter was continuous… one of them was bound to be correct.
 Zhang, Chiyuan, et al. "Understanding deep learning requires rethinking generalization." arXiv preprint arXiv:1611.03530 (2016).
 On a totally unrelated note, Michelson and Morley experiment is one of my favorites because it includes making a 1.6-ton sandstone block float on mercury.
 If you happen to enjoy such activities and want to share your results, the YPF will gladly publish them on this blog.