Moscow University Physics Bulletin最新文献

Marine radars are vital for safe navigation at sea, detecting vessels and obstacles. Sea clutter, caused by Bragg scattering, is usually filtered out as noise. It becomes detectable in unfiltered radar images, acquired using SeaVision hardware package, when wind speed and wave height exceed certain thresholds. The parameters of wind-induced ocean waves can be determined using these images; however, traditional spectral methods for obtaining wave characteristics face limitations in improving accuracy. Deep learning techniques offer advantages in image processing tasks, being more robust and able to handle noisier data, yet delivering the results without Fourier transformations and not necessarily requiring long series of radar imagery. In our study, we present the method exploiting convolutional neural networks (CNNs) for estimating wave characteristics from shipborne radar data captured using SeaVision package. In particular, we train our CNN to infer significant wave height using estimates provided by the Spotter buoy as ground truth. Our CNN-based method has an advantage over the classical methods due to the low requirements for radar image data since we process just one SeaVision snapshot, whereas classical method requires more than 20 min of radar images.

Marine X-band radar is an important navigational tool that records signals reflected from the sea surface. Theoretical studies show that the initial unfiltered signal contains information about the sea surface state, including wind wave parameters. Physical laws describing the intensity of the signal reflected from the rough surface are the basis of the classical approaches for significant wave height (SWH) estimation. Nevertheless, the latest research claims the possibility of SWH approximation using machine learning models. Both classical and AI-based approaches require in situ data collected during expensive sea expeditions or with wave monitoring systems. An alternative to real data is generation of synthetic radar images with certain wind wave parameters. This Fourier-based approach is capable of modelling the sea clutter images for wind waves of any given height. Assuming a fully-developed sea, we generate synthetic images from the Pierson–Moskowitz wave spectrum. After that, we apply an unsupervised learning using synthetic radar images to train the convolutional part of the neural network as the encoding part of the autoencoder. In this study, we demonstrate how the accuracy of SWH estimation based on radar images changes when the neural network is pretrained on synthetic data.

Imaging atmospheric Cherenkov telescopes are used to record images of extensive area showers caused by high-energy particles colliding with the upper atmosphere. The images are analyzed to determine events’ physical parameters, such as the type and the energy of the primary particles. The distributions of some of the physical parameters can be used as well, for example, to determine the properties of a gamma ray source. The key problem of any experiment is the calibration of experimental data. For this purpose, Monte Carlo simulated data with known values of the physical parameters are used. The main disadvantage of this method is its extremely high requirements for computing resources and the large amount of time spent on modelling. In this paper, we use an alternative approach: Cherenkov telescope images are simulated with conditional variational autoencoders. We compare the characteristics of both the individual images and their Hillas parameter distributions with those of the images generated by the Monte Carlo method.

Exploration geophysics requires solving specific inverse problems — reconstructing the spatial distribution of the medium properties in the thickness of the earth from the geophysical fields measured on its surface. We consider inverse problems of gravimetry, magnetometry, magnetotelluric sounding, and their integration, which means simultaneous use of various geophysical fields to reconstruct the desired distribution. Integration requires the determined parameters for all the methods to be the same. This may be achieved by the spatial statement of the problem, in which the task is to determine the boundaries of geophysical objects. In our previous studies, we considered the parameterization scheme where the inverse problem was to determine the lower boundary of several geological layers. Each layer was characterized by variable values of the depth of the lower boundary along the section, and by fixed values of density, magnetization, and resistivity, both for the layer and over the entire dataset. It was demonstrated that the integration of geophysical methods provides significantly better results than the use of each of the methods separately. The present study considers an extended and more realistic model of data—a parameterization scheme with variable properties of the medium, both along each layer and over the dataset.

Fluxes of greenhouse gases (GHG) may be assessed in situ using the eddy covariance method through processing high-frequency measurements of gas concentration and wind speed acquired at certain sites, e.g., carbon measurement test areas of the pilot project of the Ministry of Education and Science of Russia. The measurements commonly come with noise, anomalies, and gaps of various natures. These anomalies result in biased GHG flux estimates. There are a number of empirical and heuristic approaches for filtering noise and anomalies, as well as for gap-filling. These approaches are characterized by many tuning parameters that are commonly adjusted by an expert, which is a limiting factor for large-scale deployment of GHG monitoring stations. In this study, we propose an alternative approach for anomaly detection in high-frequency measurements of GHG concentration and wind speed. Our approach is based on machine learning techniques. This approach is characterized by a lower number of tuning parameters. The goal of our study is to develop a fully automated data preprocessing routine based on machine learning algorithms. We collected the dataset of high-frequency GHG concentration and wind speed measurements from one of the carbon measurement test areas. In order to compare anomaly detection algorithms, we labeled anomalies in a subset of this dataset. We present two approaches for anomaly detection, namely: (a) identification of outliers based on the error magnitude in time series statistical forecasts performed by a machine learning (ML) algorithm; and (b) classification of anomalies using an ML model trained on the labeled dataset of outliers we mentioned above. We compared the approaches and algorithms based on the F1-score metric assessed with respect to an expert-labeled subset of anomalies in GHG concentration and wind speed time series. Within the forecast-error based approach, we trained several ML models: the ARIMA autoregression method, the CatBoost model for autoregression, the CatBoost model for forecasting employing additional features, and the LSTM artificial neural network. Within the supervised classification approach, we tested the CatBoost classification model. We demonstrate that ML models for forecasting deliver a high quality of time series prediction within the autoregression approach. We also show that the anomaly identification method based on the autoregression approach delivers the best quality with the F1-score reaching (0.812).

We report on the development of neural networks for generating readings from Telescope Array’s surface detectors with the largest registered integral signal. To achieve this goal, we implemented generative Wasserstein adversarial networks with the gradient penalty. The data used to train the model was generated using the Monte Carlo method. We obtained visually similar data which are consistent with the physics of the underlying processes. The anomaly search method can be employed to identify discrepancies between real and simulated data, as well as to introduce a quantitative measure of similarity between the real detector readings and those generated by the neural network’s readings.

Baikal-GVD is a large ((sim)1 km({}^{3})) underwater neutrino telescope located in Lake Baikal, Russia. In this report, we present two machine learning techniques developed for its data analysis. First, we introduce a neural network for an efficient rejection of noise hits, emerging due to natural water luminescence. Second, we develop a neural network for distinguishing muon- and neutrino-induced events. By choosing an appropriate classification threshold, we preserve (90%) of neutrino-induced events, while muon-induced events are suppressed by a factor of (10^{-6}). Both of the developed neural networks employ the causal structure of events and surpass the precision of standard algorithmic approaches.

The digital twin of a technological process is a complex of mathematical models that allow the determination of qualitative and quantitative dependencies between process parameters. It can predict the values of controlled and observed variables dynamically, depending on the process state and control actions. Additionally, it can identify hidden dependencies, states, and factors affecting the technological process and implement the selection of optimal control actions based on goals, technological limitations, or financial constraints. When building such models using data-driven methods, the input data consist of multidimensional time series from the production system’s sensor readings. The aim of this work is to develop, implement, and evaluate a subset of the digital twin functionality as applied to the oil cracking process. The key components of the proposed methods include data preprocessing, which encompasses the challenge of selecting stable operational periods for the plant, and feature selection, represented by a gradient boosting approach. We also focus on the construction of differentiable predictive models, which use modern deep learning methods to predict controlled parameter values based on dynamic system states and control. Moreover, we apply differentiable neural network models as constraints, objective functions, and state equations to solve the optimal control problem using a classical optimal control approach.

One of the methods for the analysis of complex spectral bands (especially for spectra of liquid objects) is their decomposition into a limited number of spectral curves with physically reasonable shapes (Gaussian, Lorentzian, Voigt, etc.). Subsequent analysis of the dependences of the parameters of these contours on some external conditions in which the spectra are obtained may reveal some regularities that bear information about the physical processes taking place in the object. The problem with the required decomposition is that such a decomposition in the presence of noise in spectra is an incorrect inverse problem. Therefore, this problem is often solved by advanced optimization methods that are less likely to become stuck in local minima, such as genetic algorithms (GA). In the conventional version of GA, all individuals are similar regarding the probabilities and implementation of the main genetic operators (crossover and mutation) and the procedure of selection. In their preceding studies, the authors tested the gender GA (GGA), where the individuals of the two genders differ in terms of the mutation probability (higher for males) and the selection procedures for crossover (with the number of crossovers limited for females). In this study, we introduce additional differences between the genders in the procedures of selection and mutation. The improved modification of GGA is tested by comparing the efficiency of the conventional GA, GGA, and three versions of GGA with and without subsequent gradient descent in solving the problems of decomposition of the Raman valence band of liquid water into Gaussian contours.

Sea surface salinity (SSS) is a key physicochemical characteristic of the ocean that plays a significant role in describing the climate. Routine SSS retrieval algorithms exploiting remote sensing data have been developed and validated with high precision for typical regions of the World Ocean. Their effectiveness is worse in the Arctic though. To address this limitation, in this study, we employ machine learning (ML) techniques to enhance the quality of standard algorithms. We evaluate a few ML models, ranging from classical methods that process vector features, provided by standard Soil Moisture Active Passive (SMAP) satellite salinity algorithms, to deep artificial neural networks that combine vector features with two-dimensional fields extracted from the ERA5 reanalysis. We validate these models using in situ the data collected by the Shirshov Institute of Oceanology RAS during the expeditions to the Barents, Kara, Laptev, and East Siberian seas from 2015 to 2021. The results of the study indicate that the SMAP sea surface salinity standard product is improved in these regions. The ML models developed in this study make it possible to further study the Arctic region using enhanced sea surface salinity maps.