Scientists give opinions on the 2024 Nobel Prize in Physics
Scientists John Hopfield and Geoffrey Hinton are recognized “for fundamental discoveries and inventions” that enable machine learning using artificial neural networks.
The methods they developed formed the basis for the creation of machine learning algorithms.
“The awarding of the Nobel Prize in Physics this year is a vivid demonstration of the role of fundamental physics in the creation and development of modern technologies,” notes First Vice-Rector Dmitry Tayursky. “Using concepts and methods of statistical physics, the laureates have practically shown that computers can simulate such functions as memory and learning.”
KFU employees describe the research they are conducting using neural networks.
“John Hopfield proposed a neural network, which later became known as the Hopfield network. This uses one of the models of statistical physics, namely that each node in the neural network corresponds to a spin. In general, we have some spin system. This system is developed in such a way that in the end it comes to the desired state with minimum energy,” says Chair of the Department of Computational Physics and Modeling of Physical Processes Anatolii Mokshin. “In turn, Geoffrey Hinton on the basis of the Hopfield network develops another method, which later gets the name of Boltzmann machine. It is a kind of Hopfield network where stochastic equations are added.”
According to the scientist, Kazan Federal University conducts various studies with artificial neural networks, which are used as a tool to solve a wide variety of problems.
“Fundamental and applied research is conducted at the Department of Computational Physics. If we talk about applied research, we are engaged in solving problems related to the search for materials with the necessary physical and chemical properties. Fundamental research is aimed, in particular, at determining the properties of materials under extreme conditions, for example, at ultra-high pressures and temperatures. In addition, we carry out research related to solving so-called inverse problems in physics. In particular, we search for the interaction energy between molecules and atoms on the basis of experimental data on neutron or X-ray diffraction. In addition, we are engaged in so-called computer-aided material design, which involves determining the equilibrium crystal structure of a material based only on its known composition,” adds Mokshin.
The Institute of Geology and Petroleum Technologies actively uses artificial intelligence to automate tasks in the oil and gas industry, says Deputy Director for Innovation Vladislav Sudakov.
“For example, a system based on machine learning is used to detect seismic events that may be related to field development. It warns the population about possible fluctuations of the Earth’s crust,” Sudakov says. “In Tatarstan, over 50 years of oil field development, a large array of information about wells has been accumulated. This allows us to create such neural networks that are able to predict where and what kind of oil has been localized and to plan effective technologies for the production of these oil reserves.”
The decision-making system based on machine learning and neural networks, the scientist believes, allows for efficient and very fast acquisition of a wide variety of information about each well. Cognitive technologies tell geologists what decisions to make when examining cores, cuttings and digital images of thin sections.
Hadi Mohammadi (Iran), Lead Research Associate of the Laboratory of Computer Design of New Materials and Machine Learning (LCDNMML), believes that machine learning plays an important role in research today, especially in organizing and analyzing huge amounts of data.
“Predicting the properties of new materials through fast and cheap calculations has become possible thanks to such algorithms,” he says. “Today, based on a large set of crystal structure-property type data, it is possible to predict what new materials that do not yet exist will be like. Or, given a dataset of structures with the right properties, machine learning algorithms can be used to predict the structures of new materials with predetermined properties. The possibilities of this approach are practically limitless. I would like to note that machine learning algorithms are based on fundamental principles of statistical physics and mathematical apparatus”.
Irina Gumarova, Head of the LCDNMML, said that the laboratory conducts research using machine learning algorithms to create interatomic interaction potentials for various substances, “This year our laboratory staff published an article in the journal Computational Materials Science on the creation of a neural network potential for aluminum. We are now in the process of validating an algorithm to build a potential for gold nanoparticles. Such potentials are important for modeling the properties of these materials within molecular dynamics algorithms to predict a wide range of applied properties.”
