On the edge of the known: Will AI be able to discover new physics
Artificial intelligence is becoming an important tool for exploring the universe: neural networks help scientists analyze large amounts of data and find hidden patterns. A recent study has shown that transfer learning can speed up the search for unknown physical phenomena and reduce the cost of complex simulations. However, the excess of accumulated knowledge sometimes prevents AI from discovering new physics. About how scientists are looking for a balance between experience and the ability of algorithms to make discoveries, see the Izvestia article.
How AI helps to find new laws
Modern cosmology is trying to answer questions that still remain a mystery to scientists. To do this, researchers create complex computer models that simulate the development of the universe under different conditions. However, such calculations require huge computing resources and can take a lot of time.
A new study published in the Journal of Cosmology and Astroparticle Physics has shown that artificial intelligence can accelerate this process. Scientists have found that the transfer learning method allows neural networks to study complex cosmological models faster and find signs of new physics.
The essence of the approach is that AI is not trained immediately on the most complex simulations. First, the neural network obtains basic knowledge on simpler models based on the modern standard cosmological model ΛCDM, which describes the expansion of the universe and the distribution of galaxies. Then the algorithm is trained on more complex variants, where unknown physical processes may manifest themselves.
The experiment showed that in some cases transfer learning can reduce the number of expensive computer simulations by more than tenfold. This can significantly speed up the search for new physical phenomena, such as those related to dark matter, dark energy, massive neutrinos, or possible deviations from known laws of nature.
According to Grigory Rubtsov, Deputy Director for Scientific Work at the Institute for Nuclear Research of the Russian Academy of Sciences, Corresponding Member of the Russian Academy of Sciences, such technologies have already become a full-fledged tool of modern science.
— What is popularly called artificial intelligence, and in scientific research is called machine learning, is becoming a full-fledged tool with many practical applications. We can say that scientists have another method in their hands that significantly complements the existing ones and makes it possible to increase the accuracy of the analysis," the expert noted.
However, the study also revealed an important limitation of the new approach. Sometimes the knowledge already gained can prevent artificial intelligence from recognizing truly new phenomena. If an unknown physical effect turns out to be similar to an already familiar pattern, the neural network may interpret it through old models and not notice a fundamentally new discovery.
This effect is called "negative transference." It is similar to the situation when a person, faced with an unusual phenomenon, tries to explain it only through already known experience. In the study, scientists discovered such a problem when analyzing models with massive neutrinos: some signs of new physical processes turned out to be similar to changes associated with the σ8 parameter of the standard cosmological model.
As Grigory Rubtsov noted, the problem is natural, since the search for new phenomena has always been one of the most difficult tasks of science.
— It is known from the history of physics that people can also miss such phenomena in some cases. What's important for AI is that it's basically a tool that learns from training datasets. Simple machine learning models cannot output anything other than what was in the training data. More complex ones have variability. There are also models that find yet unknown patterns, but this task is many orders of magnitude more difficult," he explained.
How AI is Changing Scientific Search
Research in the field of cosmology shows a broader trend. Artificial intelligence is gradually becoming not just a data processing tool, but an assistant to scientists in the search for new knowledge. If today technology helps to model the development of the universe faster, then tomorrow similar methods can accelerate the development of drugs, the creation of new materials, forecasting climate change and solving complex engineering problems.
As Grigory Rubtsov explained, since many scientific tasks have a similar structure, solutions developed for one area can be successfully applied in another.
— These methods are certainly in demand in medicine, industry and other industries. The tasks are very similar, so developments often move from one area to another," the expert said.
He calls medicine one of the examples. Artificial intelligence is already helping doctors analyze medical images and examination results. These technologies become especially valuable in the diagnosis of rare diseases, when it is almost impossible to accumulate a large amount of data for training algorithms. A similar principle can be used in other scientific disciplines, where researchers have to look for rare or poorly studied phenomena.
Yuri Borodachev, head of the research center in the field of artificial intelligence in the field of Transport and logistics at the National Research Nuclear University MEPhI, emphasized in an interview with Izvestia that the main value of AI is not that it replaces humans, but that it allows researchers to focus on really complex tasks.
— Now artificial intelligence is perceived more as a limited but very useful assistant. It really increases labor productivity, especially where there are understandable algorithms and repetitive processes," the expert noted.
One of the main advantages of neural networks is the ability to work quickly with a huge number of options.
— If a scientist is faced with a complex problem that has thousands of possible solutions, a person can spend years just trying to figure out the basic approaches. And the neural network can offer several of the most promising options, show possible search directions," explained Borodachev.
Therefore, the impact of such technologies goes far beyond cosmology. As the volume of scientific data grows, artificial intelligence becomes a universal tool that helps researchers find the information they need faster, compare the results of different studies, and notice connections that would be difficult for a person to discover on their own. Ultimately, this allows us to accelerate scientific discoveries, which means that new technologies, treatments, and engineering solutions that directly affect people's quality of life can be implemented faster.
Will AI be able to become the main space explorer
In the coming years, the role of artificial intelligence in cosmology may grow significantly. New telescopes and scientific projects will collect unprecedented amounts of data that cannot be processed manually. Under these conditions, it is algorithms that will help scientists find hidden patterns faster, test hypotheses, and look for signs of new physics.
According to Yuri Borodachev, the development of such technologies will change the very work of a researcher.
— More and more of the mechanical and routine work of the researcher will be transferred to artificial intelligence. The person will be more involved in the formulation of tasks, setting up experiments and analyzing the results obtained," the expert believes.
However, despite the rapid development of technology, experts do not expect that artificial intelligence will be able to completely replace scientists. Grigory Rubtsov believes that algorithms remain a tool, and the determining role still belongs to man.
— The boundary runs in the same place as the boundary between the carpenter's work and the plane — at the point of contact of the hand and the handle. It can be assumed that the plane will be automatic and there is no need to hold it, but even in this case, there is a person operating this machine somewhere. A person determines the meaning and content of the machine's operation, and the same thing happens with the automation of scientific research," the expert explained.
As Rubtsov emphasized, artificial intelligence is indeed capable of accelerating individual discoveries, but scientific progress itself is determined not only by computing capabilities, but also by the ideas of researchers, and sometimes by a successful combination of circumstances. Moreover, excessive trust in algorithms can have the opposite effect.
The development of artificial intelligence raises new ethical issues. If algorithms begin to manage not only digital data, but also real objects, such as autonomous vehicles or robotic systems, the question of responsibility for decisions arises.
According to Yuri Borodachev, even the most advanced autonomous systems should remain under human control.
— There should be a conditional "red button" that will allow you to stop the system if it does not start working as expected, — the expert concluded.
Переведено сервисом «Яндекс Переводчик»
