That the Earth and the other planets of the solar system revolve around the Sun seems obvious to us today, but it was a conclusion that astronomers reached after centuries of studies and animated discussions. A modern-day artificial intelligence has instead managed to understand by itself that the Sun is at the center of the solar system, based on the apparent movements of the Sun and Mars in the terrestrial sky. The test served to experiment with new machine learning systems that could one day help physicists discover new laws and properties, solving even very complex problems related to quantum mechanics.
The experiment was conducted at the ETH Zurich and its results were published in the scientific journal Physical Review Letters. The researchers' goal was to develop an algorithm capable of analyzing huge data sets from which to derive mathematical formulas, mimicking as much as possible the systems that physicists use to derive the equations that describe the phenomena they observe or to predict their course. To do this, they used a “neural network”, that is a machine learning system that simulates the functioning of the human brain.
In general, neural networks learn to interpret certain phenomena through the analysis of large data sets. Thanks to their use, various applications have been developed in recent years, for example for the recognition of objects in images or people's faces. The information derived from the data is then diffused into the millions of nodes of the neural network, which in the analogy with the human brain correspond to neurons, the single cells that make up the nervous tissue. The problem is that this information is not easily accessible and complicated to interpret.
The researchers then thought of simplifying the system by creating two smaller neural networks, linked together by a very low number of connections. One of these had the task of annalizing the dataset, while the second had to use the analyzes of the first to perform tests and make predictions. Faced with the need to use few links to communicate, the first network had to provide data in condensed form. Naturally the solution adopted by the researchers presented several other complications, but in principle we can say that it worked a bit like the dynamic that is established between teacher and student: the first has much more information, but transmits to the second a distillate with the things that are more relevant and understandable at that stage of learning.
Once the neural network was set up, the researchers fed the system a set of data on the movements of Mars and the Sun from an observation point on Earth. Mars, when observed in its motion relative to that of the Earth, is a bit strange: at certain times of the year it moves in reverse, as if it changes direction. Believing for centuries that the Earth was at the center of the Universe (geocentric system), astronomers explained the phenomenon by theorising that the planets moved by describing small circles (epicycles) and that for this reason they sometimes seemed to reverse their direction, from the point of observation. on earth. In the mid-sixteenth century, Nicolaus Copernicus found a more logical explanation: taking up hypotheses already formulated in the past, he guessed that it was the Sun that was at the center of the solar system (and of the Universe, but he was a little carried away) and that were therefore the planets to revolve around it.
The neural network of modern day researchers has done something similar, coming to produce formulas that correctly describe the orbit of Mars, and that make the conclusion that the planets of the solar system revolve around the Sun inevitable. This conclusion, over four centuries old. ago, revolutionized much of the knowledge on the Universe and paved the way for the subsequent discoveries of Galileo Galilei and then of Giovanni Keplero.
In their study, the researchers clarify that formally artificial intelligence indicated formulas and equations, which then required human intervention to be interpreted and verify that they described the correct planetary motions. However, the experiment showed that a neural network has the potential to analyze a system and describe it with the laws of physics. In the future, with more elaborate algorithms and greater computational capabilities, it will be possible to create systems to analyze complex physical phenomena, the properties of which still elude researchers.
For the next evolutions of their system, the Zurich researchers plan to add features that affect algorithms less and incentivize them to explore less orthodox solutions than the classic ones. They are confident that they will succeed by creating a neural network that can create new experiments and simulations on its own, with which to test their own hypotheses.
