Main article: Artificial Intelligence
How AI is used in science
Generative AI can have a fundamental impact on science, technology, accelerating technological progress. In 2023, Spydell Finance identified several key areas in this area.
Systematization and structuring of ultra-large arrays of information
How can a scientist find similar research materials? Through search, but the data can be irrelevant or obsolete. Indexing and analysis of thousands of scientific articles is needed to combine and integrate similar studies into a holistic picture.
Hypothesis generation
AI can be used to generate hypotheses that can be used to conduct scientific research. This could help scientists speed up the process of discovering new knowledge.
Ultra-fast search and processing of solution combinations to find the optimal research path
AI can help scientists design and optimize experiments by predicting the most promising areas of research, which reduces costs and increases the chances of success.
Simulation and Simulation
AI is able to create complex models and simulations that can predict the results of experiments and research, as well as help in understanding complex systems and processes.
Quickly find and correct errors in mathematical, physical models, or program code
Quickly finding and correcting errors in mathematical, physical models or program code will simplify and speed up the calculation process.
Analysis and interpretation of complex data in modeling complex systems
Analysis and interpretation of complex data in the modeling of complex systems, creating a more understandable and readable data structure.
2024: In Russia, the introduction of AI solutions in the field of science is only gaining momentum so far
The Institute of Statistical Research and Knowledge Economics HSE based on data statistics and the specialized survey "Doing in" science Russia analyzes the prevalence in scientific organizations and universities countries of practices of using AI solutions for research and development.
Modern AI-based technologies are changing the usual way of life in all areas of activity, and science is no exception. A survey conducted by the journal Nature (2023) showed that more than a quarter of scientists already using AI in their research expect this technology to become an integral tool for their field of science in the next 10 years, another 47% are confident that it will be very useful. A similar study by Oxford University Press suggests that it will not take so long to wait: 75% of scientists surveyed who publish their work in leading journals used various AI tools in 2024, including machine translation services (49%), chat bots (43%) and search engines (25%). According to respondents, AI-based solutions are useful at all stages of the research cycle and for a wide range of tasks: 41% of respondents used them to search for literature, approximately 35% - to summarize and/or edit the text (for example, the manuscript of an article), 25% each - to generate ideas, collect data and/or analyze them.
According to statistics, in Russia the introduction of AI solutions in the field of science is only gaining momentum. In 2023, about 5% of scientific organizations and about 10% of universities used AI for their own purposes, but these indicators do not fully reflect the real scale of the use of this technology by scientists, since they characterize only the practices of the organizations themselves, and not their employees.
In the future, we should expect an expansion of the introduction of AI in the field of science and higher education: every second organization sees prospects for the further use of appropriate tools in its activities here. In addition, almost 25% of scientific organizations and 38% of universities already using AI believe that such technologies will radically change the internal processes in science in the coming years; many of them consider intelligent decision support technologies to be the most promising for these tasks (33%).
Obviously, the possibility of realizing these expectations largely depends on the level of development of the necessary digital infrastructure. According to a survey of 719 scientific organizations and universities conducted by the ISIEZ NRU HSE as part of the project "Doing Science in Russia" (October - November 2024), access to AI systems for research and development is still difficult. The surveyed managers rated the provision of such systems of foreign development (ChatGPT, Trinka, Mendeley, Scite, Google Jax, etc.) by 2.71 points out of five possible, domestic (GigaChat, GitVerse, YaLM, SOVA, RAZUM AI, GOLEM, NeuroMark, AI BAUM PLATFORM, NiznW points, etc.) 2.60 The situation in universities is somewhat better than in other organizations (Chart 1).
Against the background of restrained assessments of the current situation, forecasts for the next three years look more optimistic: organizations of all types expect a significant increase in the practice of using AI systems for research and development. Of course, to ensure such dynamics, it is necessary to eliminate the barriers that prevent the spread of AI in science. Among the most significant of them, universities and scientific organizations note: lack of financial resources, lack of qualified personnel, underdeveloped ICT infrastructure, lack/low quality of big data for the introduction of AI. The influence of these restraining factors is indicated by half of universities and about 40% of scientific organizations.
Overcoming the barriers of the spread of AI in Russian science could be facilitated by a special program providing for the development of standards for research using AI; grants to young scientists and scientific teams studying and using AI in their activities (as a priority - for those areas of science where such technologies are rarely used); support for the development of AI applications for science tasks; compensation for the costs of universities and scientific organizations for the purchase of big data in order to train and develop generative models[1].
2022: AI simplified the solution of the well-known problem of quantum physics from 100 thousand equations to four
On October 4, 2022, it became known that with the help of artificial intelligence (AI), physicists were able to radically optimize the known quantum problem, which until recently implied the solution of 100 thousand different equations. Now it is enough to solve the four equations, and this is without any sacrifice in terms of the accuracy of the results.
As reported, the work, published in Physical Review Letters on September 23, 2022, could lead to changes in how scientists investigate systems containing multiple interacting electrons. If this solution can be scaled to other similar problems, it will be possible to create superconducting materials or means of environmentally friendly energy production.
 | We start with a large corpus of interconnected differential equations and then use machine learning to turn it into something so small that you can count it on your fingers. |  |
The problem, known as the Hubbard model, relates to the behavior of electrons moving inside a lattice structure. If two electrons occupy one point in the lattice, they interact. Hubbard's model is an "ideal" variant of several important classes of materials; with it, scientists gain insight into how the behavior of electrons provides the desired states of matter, such as superconductivity, in which electrons move without encountering resistance. The model is also used to work out various methods of working with more complex quantum systems.
The simplicity of Hubbard's model, however, is deeply deceptive, the publication writes Phys.org. Even when a modest number of electrons are calculated, and the most advanced computational approaches are used, the amount of computation itself remains large. The point is quantum coupling: after two electrons interact, they are coupled, and no matter how far apart they are subsequently, they cannot be considered as independent units. As a result, physicists have to take into account all electrons at once, and not each individually. And the more electrons are added to the system, the more clutches occur, and the higher the computational resources that are required to study such a system.
In such cases, physicists use renormalization groups - a mathematical apparatus that is used to detect changes in the system when modifying its properties, such as temperature, or the consequences of scaling.
However, even a renormalization group tracking all possible couplings between electrons without compromising accuracy would contain tens of thousands, hundreds of thousands, or even millions of individual equations to be solved.
Di Sante and his colleagues thought about the possibility of using a neural network in order to make a massive renormalization group more manageable. And they succeeded.
The neural network first indexed all connections in the full-size renormalization group, then reconfigured the strength of these connections until it revealed a narrowly limited set of equations that produce exactly the same result as the original renormalization group. The number of such equations was eventually reduced to four.
Neural network training required large computing resources: the program worked continuously for several weeks. However, now this neural network can be used to produce calculations in connection with other major physical and mathematical problems, without the need to start its training from scratch.
Di Sante and his associates also study what exactly their neural network "understood" about the system to which it was applied, in the hope of revealing patterns that were not obvious to physicists before.
The question remains how much this approach works with more complex quantum systems, for example, with materials in which electrons interact at long distances. According to Di Sante, there are very interesting opportunities to use this method in other areas where renormalization groups are used, including cosmology and neurology.
| | If the conclusions made in this work are not refuted, then perhaps we are talking about a global revolution in physics. A revolution that turned out to be achievable only thanks to machine learning and the ability characteristic of neural networks to identify hidden patterns that allow complex systems to be reduced to a reasonable number of parameters. So far, the capabilities of neural networks are at the initial stage of development, but there is reason to believe that in the future they will be able to solve some other problems and problems of physics that still remain unresolved, for example, Schrödinger equations, multiple problems of superfluidity[2] | |
2020: Sberbank attracts artificial intelligence to decipher the manuscripts of Peter the Great
On June 29, 2020, it became known that Sberbank decided to attract artificial intelligence technologies to decipher the manuscripts of Peter the Great. Read more here.
Notes
| The site content is translated by machine translation software powered by PROMT. The machine-translated articles are not always perfect and may contain errors in vocabulary, syntax or grammar. Read original article If you find inaccuracies or errors in the results of machine translation, please write to editor@tadviser.ru. We will make every effort to correct them as soon as possible. |