Moscow University Physics Bulletin最新文献

Hydrodynamic modeling via numerical simulators of underground gas storages (UGSs) is an integral part of planning and decision-making in various aspects of UGS operation. Although numerical simulators can provide accurate predictions of numerous parameters in UGS reservoirs, in many cases this process can be computationally expensive. In particular, calculation time is one of the major constraints affecting decisions related to optimal well control and distribution of gas injection or withdrawal over the reservoir area. Novel deep learning methods that can provide a faster alternative to traditional numerical reservoir simulators with acceptable loss of accuracy are investigated in this paper. Hydrodynamic processes of gas flow in UGS reservoirs are described by partial differential equations (PDEs). Since PDEs involve approximating solutions in infinite-dimensional function spaces, this distinguishes such problems from traditional ones. Currently, one of the most promising machine learning approaches in scientific computing (scientific machine learning) is the training of neural operators that represent mappings between function spaces. In this paper, a deep learning method for hydrodynamic modeling of UGS is proposed. A modified Fourier neural operator for hydrodynamic modeling of UGS is developed, in which the model parameters in the spectral domain are represented as factorized low-rank tensors. We trained the model on data obtained from a numerical model of UGS with nonuniform discretization grid, more than 100 wells and complex geometry. Our method demonstrates superior performance compared to the original Fourier neural operator (FNO), with an order of magnitude (50 times) fewer parameters. Tensor decomposition not only greatly reduced the number of parameters, but also increased the generalization ability of the model. Developed neural operator simulates a given scenario in a fraction of a second, which is at least (10^{6}) times faster than a traditional numerical solver.

In this article, we talk about generating data obtained in the Telescope Array experiment. For this we are using Wasserstein’s generative adversarial networks. Wasserstein’s generative adversarial networks were trained on data obtained using the Monte Carlo method. To improve the quality of the generation, we add the loss function for the generator, which is based on the physics of the process of spreading an extensive air shower. In the future, this network can be used to search for anomalies and for faster data generation, compared with algorithms based on the Monte Carlo method.

Modern Imaging Atmospheric Cherenkov Telescopes (IACTs) generate a huge amount of data that must be classified automatically, ideally in real time. Currently, machine learning-based solutions are increasingly being used to solve classification problems. However, these classifiers require proper training data sets to work correctly. The problem with training neural networks on real IACT data is that these data need to be prelabeled, whereas such labeling is difficult and its results are estimates. In addition, the distribution of incoming events is highly imbalanced. Firstly, there is an imbalance in the types of events, since the number of detected gamma quanta is significantly less than the number of protons. Secondly, the energy distribution of particles of the same type is also imbalanced, since high-energy particles are extremely rare. This imbalance results in poorly trained classifiers that, once trained, do not handle rare events correctly. Using only conventional Monte Carlo event simulation methods to solve this problem is possible, but extremely resource-intensive and time-consuming. To address this issue, we propose to perform data augmentation with artificially generated events of the desired type and energy using conditional generative adversarial networks (cGANs), distinguishing classes by energy values. In the paper, we describe a simple algorithm for generating balanced data sets using cGANs. Thus, the proposed neural network model produces both imbalanced data sets for physical analysis as well as balanced data sets suitable for training other neural networks.

In the end-cap region of the SPD detector complex, particle identification will be provided by a Focusing Aerogel RICH detector (FARICH). FARICH’s primary function is to separate pions and kaons in final open charmonia states (momenta below 5 GeV/(c)). The optimization of detector parameters, as well as a free-running (triggerless) data acquisition pipeline to be employed in the SPD necessitate a fast and robust method of event reconstruction. In this work, we employ a Convolutional Neural Network (CNN) for particle identification in FARICH. The CNN model achieves a superior separation between pions and kaons compared with traditional approaches. Unlike algorithmic methods, an end-to-end CNN model is able to process a full 2-dimensional detector response and skip the intermediate step of computing particle velocity, solving the particle classification task directly.

This study addresses the problem of environmental monitoring of air in cities and industrial areas, which consists in identification of gases and volatile organic compounds using metal oxide (MOX) semiconductor gas sensors. To provide selectivity in the detection of certain gases, the laboratory-made MOX gas sensors are operated in a modulated working temperature mode in combination with signal processing and machine learning approach to establish the response models. Six types of nonlinear operating temperature conditions—the so-called heating dynamics—were applied to twelve sensors with sensing layers of different chemical composition. Nine gases (CO, CH({}_{4}), H({}_{2}), NH({}_{3}), NO, NO({}_{2}), H({}_{2})S, SO({}_{2}), formaldehyde) in six different concentrations each were used as polluting admixtures to dry clean air. Due to the high complexity of the model describing the processes of interaction between gases and sensors, machine learning methods (logistic regression, random forest and gradient boosting) based on the use of physical experiment data were used to process the sensor response. Optimal heating dynamics and optimal machine learning methods for gas identification have been determined.

In this study we considered an inverse problem of optical spectroscopy. It consists in determining concentrations of the ingredient ions of multicomponent water solutions by their spectra. The problem of describing the spectra of multicomponent solutions is nonlinear and has no adequate mathematical model. Because of this, machine learning methods using experimental data were chosen to solve this problem. At the same time, inverse problems of spectroscopy are characterized by high input dimensionality with a large number of features, more or less relevant. In their turn, some of the relevant features are redundant due to their multicollinearity. This is caused by the fact that the characteristic lines have a width of several spectrum channels. Presence of redundant features leads to a deterioration in the quality of machine learning solution of the problem. Thus, there is a need for a feature selection procedure that takes into account both their relevance and redundancy, as well as their nonlinear relationship with the determined parameters. In this study, we considered a feature selection procedure based on the iterative selection of features with the highest relevance to the target variable and on the elimination of features with a high relationship with each other. In this selection process, we used a trained neural network to analyze weights and determine feature importance in a nonlinear way. We also used the Pearson correlation coefficient to measure how features are related to one another. Finally, we compared the quality of a neural network solution using spectroscopic data of the full set of input features and of its subsets. These subsets were compiled using the selection procedure under consideration. We also used traditional methods for selecting significant input features as baseline methods.

The generation of crystals with tailored properties is a significant challenge in both scientific research and practical applications. Due to the vast configuration space of crystalline structures, finding precise solutions to such problems is computationally intensive. In this study, we propose a method for generating defect configurations in MoS({}_{2}) crystals aimed at producing crystals with specific characteristics, focusing on formation energy and HOMO-LUMO energy levels as key examples. The approach leverages symbolic regression techniques, trained on datasets of two-dimensional materials with defects, to predict crystal properties. We introduce methods for identifying defect configurations with both minimal and specific formation energies, as well as for optimizing HOMO-LUMO energy levels. The main advantages of this approach are its efficiency and accuracy in generating valid and optimized crystal structures.

A new SPHERE-3 telescope is being developed for the study of the cosmic ray spectrum and mass composition in the 5–1000 PeV energy range. Registration of extensive air showers using reflected Cherenkov light method applied in the SPHERE detector series requires a good trigger system for accurate separation of events from the background produced by starlight and airglow photons reflected from the snow. Here, we present the results of convolutional networks application for the classification of images obtained from Monte Carlo simulation of the detector. The simulated detector response includes photon tracing through the optical system, silicon photomultiplier operation, and the electronics response and digitization process. The results are compared to the SPHERE-2 trigger system performance.

We develop loss functionals for training physics–informed neural networks using variational principles for nonpotential operators. Generally, a quasiclassical variational functional is bounded from above or below, contains derivatives of lower order compared to the order of derivatives in partial differential equation and some boundary conditions are integrated into the functional, which results in lower computational costs when evaluating the functional via Monte Carlo integration. Quasiclassical loss functional of boundary value problem for hyperbolic equation is obtained using the symmetrizing operator by V.M. Shalov. We demonstrate convergence of the neural network training and advantages of quasiclassical loss functional over conventional residual loss functional of boundary value problems for hyperbolic equation.

Over the past decades, brain–computer interfaces (BCIs) have been rapidly evolving. A BCI is a system that records brain activity signals using electrophysiological methods and then processes these signals to generate control commands. The most challenging aspect of BCIs is the nonstationary nature of brain signals, which makes it difficult to achieve stable and accurate decoding. Therefore, developing robust methods for processing and classifying EEG signals to extract control commands is a critical research area. A related challenge is the low signal-to-noise ratio in EEG data, especially when target patterns are weak or the data is labeled inaccurately. This paper presents the results of an evaluation of an approach combining feature extraction and data augmentation techniques to address the aforementioned challenges applied to the classification of premotor potentials. The approach is based on the application of linear discriminant analysis (LDA) to sequentially extract informative components in the frequency and time domains For the first time, the applicability of this algorithm to EEG containing premotor patterns of real movements is demonstrated. Features of different nature (spectral power, Hjorth parameters, interchannel correlations) were tested and compared with each other and a traditional approach based on common spatial patterns and a linear classifier. It is shown that transformations in the frequency domain alone improve accuracy from 63.9(%) in the traditional approach to 77.5(%) on a dataset of 16 experiments on different subjects. With additional transformation in the time domain, accuracy increases to 98.8(%). On average, across different model configurations, a segment length of 500 ms is the most optimal. Two approaches were developed and tested to achieve algorithm universality across subjects: universal transformations in frequency domain trained on data from all subjects and without this step at all. It is shown that accuracies of up to 98.3(%) can be achieved with such approaches. A discussion of optimal frequency bands, segment lengths, and features is provided. Thus, data from different subjects can be effectively classified by a common model, which is rare in global research and is usually accompanied by a number of assumptions, cumbersome models, and inferior accuracy. Thus, in addition to the achieved accuracy enhancement, the proposed algorithm exhibits robustness to transient noise and artifacts through signal segmentation into short epochs. It also effectively addresses the critical task of extracting informative signal components in scenarios with potentially imprecise expert annotations. Finally, it can be adapted to mitigate the need for subject-specific calibration. These attributes render the proposed algorithm suitable for real-time applications, including closed-loop BCIs for addressing the pressing challenge of neurorehabilitation.