In this work we conduct a thorough examination of utilizing machine learning and computer
vision techniques for classifying meteors based on their characteristics. The focus of the re-
search is the analysis of light curves emitted by meteors as they pass through the Earth’s atmo-
sphere, including aspects such as luminosity, duration, and shape. Through extracting features
from these light curves and comparing them to established meteors orbits, valuable informa-
tion about the meteor’s origin and chemical composition is sought to be obtained. A significant
contribution of the thesis is the development of methods for classifying meteors by extracting
features from the light curve shape through the usage of unsupervised classification algorithms.
This approach allows for the automatic classification of meteors into various groups based on
their properties. Data for the research is collected by a three-camera setup at the Armagh observatory,
comprising one medium-angle camera and
two wide-angle cameras. This setup enables the capturing of detailed images of meteor light
curves, as well as various other observations such as coordinate and angular data. The research
also involves the use of machine learning algorithms for data reduction and classification tasks.
By applying these techniques to the data collected from the camera setup, the identification of
parent objects based on chemical composition and meteor path is facilitated, along with the
acquisition of other valuable information about the meteors.

How to cite: Mandl, L., Christou, A., and Windisch, A.: A machine learning approach to meteor light curve analysis, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-21463, https://doi.org/10.5194/egusphere-egu24-21463, 2024.

Solar eruptive events are complex phenomena, which most often include coronal mass ejections (CME), flares, compressive/shock waves, and filament eruptions. CMEs are large eruptions of magnetized plasma from the Sun’s outer atmosphere or corona, that propagate outward into the interplanetary space. Solar Energetic Particles (SEP) are produced through particle acceleration in flares or CME-driven shocks. Exact mechanisms behind SEP production are yet to be understood, but it is thought that most of their acceleration occurs in shocks starting in the low corona. Over the last several decades a large amount of remote solar eruption observations have become available from ground-based and space-borne instruments. This has required the development of software approaches for automated characterization of eruptive features. Most solar feature detection and tracking algorithms currently in use have restricted applicability and complicated processing chains, while the complexities in engineering machine learning (ML) training sets limit the use of data-driven approaches for tracking or solar eruptive related phenomena. Recently, we introduced a hybrid algorithmic—data driven approach for characterization and tracking of solar eruptive features with the improved wavelet-based, multi-instrument Wavetrack package (Stepanyuk et.al, J. Space Weather Space Clim. (2024)), which was used to produce training datasets for data driven image segmentation with convolutional neural networks (CNN). Its perfomance was shown on a limited set of SDO AIA 193A instrument data perfoming segmentation of EUV and shock waves. Here we extend this approach and present an ensemble of more general CNN models for data-driven segmentation of various eruptive phenomena for the set of ground-based and remote instruments data. We discuss our approach to engineering training sets and data augmentation, CNN topology and training techniques. 

How to cite: Stepanyuk, O. and Kozarev, K.: Segmentation and Tracking of Solar Eruptive Phenomena with Convolutional Neural Networks (CNN), EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-19248, https://doi.org/10.5194/egusphere-egu24-19248, 2024.

In our previous publication (Arnaut et al. 2023), we demonstrated the application of the Random Forest (RF) algorithm for classifying disturbances associated with solar flares (SF), erroneous signals, and measurement errors in VLF amplitude data i.e., anomaly detection in VLF amplitude data. The RF algorithm is widely regarded as a preferred option for conducting research in novel domains. Its advantages, such as its ability to avoid overfitting data and its simplicity, make it particularly valuable in these situations. Nevertheless, it is imperative to conduct thorough testing and evaluation of alternative algorithms and methods to ascertain their potential advantages and enhance the overall efficiency of the method. This brief communication demonstrates the application of the XGBoost (XGB) method on the exact dataset previously used for the RF algorithm, along with a comparative analysis between the two algorithms. Given that the problem is framed as a machine learning (ML) problem with a focus on the minority class, the comparative analysis is exclusively conducted using the minority (anomalous) data class. The data pre-processing methodology can be found in Arnaut et al. (2023). The XGB tuning process involved using a grid search method to optimize the hyperparameters of the model. The number of estimators (trees) was varied from 25 to 500 in increments of 25, and the learning rate was varied from 0.02 to 0.4 in increments of 0.02. The F1-Score for the anomalous data class is similar for both models, with a value of 0.508 for the RF model and 0.51 for the XGB model. These scores were calculated using the entire test dataset, which consists of 19 transmitter-receiver pairs. Upon closer examination, it becomes evident that the RF model exhibits a higher precision metric (0.488) than the XGB model (0.37), while the XGB model demonstrates a higher recall metric (0.84) compared to the RF model (0.53). Upon examining each individual transmitter-receiver pair, it was found that XGB outperformed RF in terms of F1-Scores in 10 out of 19 cases. The most significant disparities are observed in cases where the XGB model outperformed by a margin of 0.15 in terms of F1-Score, but conversely performed worse by approximately -0.16 in another instance for the anomalous data class. The XGB models outperformed the RF model by approximately 6.72% in terms of the F1-score for the anomalous data class when averaging all the 19 transmitter-receiver pairs. When utilizing a point-based evaluation metric that assigns rewards or penalties for each entry in the confusion matrix, the RF model demonstrates an overall improvement of approximately 5% compared to the XGB model. Overall, the comparison between the RF and XGB models is ambiguous. Both models have instances where one is superior to the other. Further research is necessary to fully optimize the method, which has benefits in automatically classifying VLF amplitude anomalous signals caused by SF effects, erroneous measurements, and other factors.

References:

Arnaut, F., Kolarski, A. and Srećković, V.A., 2023. Random Forest Classification and Ionospheric Response to Solar Flares: Analysis and Validation. Universe, 9(10), p.436.

How to cite: Arnaut, F., Kolarski, A., and Srećković, V.: Comparative Analysis of Random Forest and XGBoost in Classifying Ionospheric Signal Disturbances During Solar Flares, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-2046, https://doi.org/10.5194/egusphere-egu24-2046, 2024.

The solar activity has various direct and indirect impacts on human activities. During periods of high solar activity, the harmful effects triggered by solar variability are maximized. On a decadal to multidecadal time scale, solar variability exhibits a main cycle of around 11 years known as the Schwabe solar cycle, leading to a solar maximum approximately every 11 years. The most commonly used variable for measuring solar activity is the sunspot number. Over the last few decades, numerous techniques have been employed to predict the time evolution of the solar cycle for subsequent years. Recently, there has been a growing number of studies utilizing machine learning methods to predict solar cycles. One such method is the Gaussian process, which is well-suited for working with small amounts of data and can also provide an uncertainty measure for predictions. In this study, the Gaussian process technique is employed to predict the sunspot number between 2024 and 2050. The dataset used to train and validate the model comprises monthly averages of sunspots relative to the period 1700-2023. According to the results, the current solar cycle, currently at its maximum, is anticipated to last until 2030. The subsequent solar maximum is projected to occur around the end of 2033, with an estimated maximum sunspot number of approximately 150. If this prediction holds true, the next solar cycle's maximum will resemble that observed in the current one.

How to cite: Frigo, E. and Gonçalves, I.: Prediction of sunspot number using Gaussian processes, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-4181, https://doi.org/10.5194/egusphere-egu24-4181, 2024.

This study aims to identify sets of in-situ measurements of the solar wind which sample the same volume of plasma at different times and locations as it travels through the heliosphere using ensemble machine learning methods. Multiple observations of a single volume of plasma by different spacecraft - referred to here as conjunctions - are becoming more frequent in the current “golden age of heliophysics research” and are key to characterizing the expansion of the solar wind. Specifically, identifying these related observations will enable us to test the current understanding of solar wind acceleration from the corona to the inner heliosphere with a more comprehensive set of measurements than has been used in previous analyses.

Using in-situ measurements of the background solar wind from Solar Orbiter, Parker Solar Probe, STEREO-A, Wind and BepiColombo, we identify a set of criteria based on features of magnetic field, velocity, density and temperature timeseries of known conjunctions and search for other instances for which the criteria are satisfied, to find previously unknown conjunctions. We use an ensemble of models, including random forests and recurrent neural networks with long short-term memory trained on synthetic observations obtained from magnetohydrodynamic simulations, to identify candidate conjunctions solely from kinetic properties of the solar wind. Initial results show a previously unidentified set of conjunctions between the spacecraft considered in this study. While this analysis has thus far only been performed on observations obtained since 2021 (start of Solar Orbiter science operations), the methods used here can be applied to other datasets to increase the potential for scientific return of existing and future heliophysics missions.

The modular scientific software built over the course of this research includes methods for the retrieval, processing, visualisation, and analysis of observational and synthetic timeseries of solar wind properties. It also includes methods for feature engineering and integration with widely used machine learning libraries. The software is available as an open-source Python package to ensure results can be easily reproduced and to facilitate further investigation of coordinated in-situ data in heliophysics.

How to cite: Faes, Z., Hayes, L., Müller, D., and Walsh, A.: Finding Hidden Conjunctions in the Solar Wind, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-12885, https://doi.org/10.5194/egusphere-egu24-12885, 2024.

Magnetohydrodynamic (MHD) models provide a quantitative 3D distribution of the solar corona parameters (density, radial velocity, and temperature). However, this process is expensive and time-consuming. For this, we apply deep learning models to reproduce the 3D distribution of solar coronal parameters from 2D synoptic photospheric magnetic fields. We consider synoptic photospheric magnetic fields as an input to obtain 3D solar coronal parameters simulated by the MHD Algorithm outside a Sphere (MAS) from June 2010 to January 2023. Each parameter is individually trained using 150 deep learning models, corresponding to 150 solar radial distances ranging from 1 to 30 solar radii. Our study yields significant findings. Firstly, our model accurately reproduces 3D coronal parameter structures across the 1 to 30 solar radii range, demonstrating an average correlation coefficient value of approximately 0.96. Secondly, the 150 deep-learning models exhibit a remarkably shorter runtime (about 16 seconds for each parameter), with an NVIDIA Titan XP GPU, in comparison to the conventional MAS simulation time. As the MAS simulation is a regularization model, we may significantly reduce the simulation time by using our results as an initial magnetic configuration to obtain an equilibrium condition. In the future, we hope that the generated solar coronal parameters can be used for near real-time forecasting of heliospheric propagation of solar eruptions.

How to cite: Rahman, S., Jeong, H.-J., Siddique, A., and Moon, Y.-J.: Near real-time construction of Solar Coronal Parameters based on MAS simulation by Deep Learning , EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-14186, https://doi.org/10.5194/egusphere-egu24-14186, 2024.

When deriving fluid equations from the Vlasov equation for collisionless plasmas, one runs into the so-called closure problem: each equation for the temporal evolution of one particle moment (density, current, pressure, heat flux, …) includes terms depending on the next order moment. Therefore, when choosing to truncate the description at the nth order, one must approximate the terms related to the (n+1)th order moment included in the evolution equation for the nth order moment. The order at which the hierarchy is closed and the assumption behind the approximations used determine how accurately a fluid description can reproduce kinetic processes.

In this work, we aim at reconstructing specific particle moments from kinetic simulations, using as input the electric and magnetic field and the lower moments. We use fully kinetic Particle In Cell simulations, where all physical information is available, as the ground truth. The approach we present here uses supervised machine learning to enable a neural network to learn how to reconstruct higher moments from lower moments and fields.

Starting from the work of Laperre et al., 2022 we built a framework which makes it possible to train feedforward multilayer perceptrons on kinetic simulations to learn to predict the higher moments of the Vlasov equation from the lower moments, which would also be available in fluid simulations. We train on simulations of magnetic reconnection in a double Harris current sheet with varying background guide field obtained with the semi-implicit Particle-in-Cell code iPiC3D (Markidis et al, 2010). We test the influence of data preprocessing techniques, of (hyper-)parameter variations and of different architectures of the neural networks on the quality of the predictions that are produced. Furthermore, we investigate which metrics are most useful to evaluate the quality of the outcome.

How to cite: Köhne, S., Laperre, B., Amaya, J., Jamal, S., Lautenbach, S., Grauer, R., Lapenta, G., and Innocenti, M. E.: Addressing the closure problem using supervised Machine Learning, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-15981, https://doi.org/10.5194/egusphere-egu24-15981, 2024.

we show the evolutions of the separated strands within the apparent single coronal loops observed in Atmospheric Imaging Assembly (AIA) images. The loop strands are detected on  the upsampled AIA 193 images, which are   generated using a super-resolution convolutional neural  network, respectively. The architecture of the network is designed to map the AIA images to unprecedentedly high spatial resolution coronal images taken by  High-resolution Coronal Imager (Hi-C) during its brief flight. At some times, pairs of individual strands appeared to braid with each other and subsequently evolved to become pairs of almost parallel ones with their segments having exchanged totally.  These evolutions provide  morphological evidence supporting occurrences of magnetic reconnections between the braiding strands, which are further confirmed by  the occurrences of the transient hot emissions (>5 MK)  located at the footpoints of  the braiding structures. 

How to cite: Bi, Y.: The coronal braiding structures detected in the machine-learning upscaled SDO/AIA images, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-1494, https://doi.org/10.5194/egusphere-egu24-1494, 2024.

The detection of asteroids involves the processing of sequences of astronomical images. The main challenges arise from the huge volume of data that should be processed in a reasonable amount of time. To address this, we developed the NEARBY platform [1], [2] for efficiently automatic detection of asteroids in sequence of astronomical images. This platform encompasses multidimensional data processing capabilities, human-verified visual analysis, and cloud-based adaptability. This paper outlines the enhancements we have made to this automated asteroid detection system by integrating a machine learning-based classifier known as the CERES module. The integration of the CERES module [3] into the NEARBY platform substantially enhances its performance by automatically reducing the number of false positive detections. Consequently, this leads to a more reliable and efficient system for asteroid identification, while also reducing the time and effort required by human experts to validate detected candidates (asteroids). The experiments highlight these improvements and their significance in advancing the field of asteroid tracking. Additionally, we explore the applicability of the asteroid classification model, initially trained using images from a specific telescope, across different telescopes.

Acknowledgment:

• This work was supported by a grant of the Romanian Ministry of Education and Research, CCCDI - UEFISCDI, project number PN-III-P2-2.1-PED-2019-0796, within PNCDI III. (the development of the dataset and CNN models)
• This research was partially supported by the project 38 PFE in the frame of the programme PDI-PFE-CDI 2021.

References:

• Bacu, V., Sabou, A., Stefanut, T., Gorgan, D., Vaduvescu, O., NEARBY platform for detecting asteroids in astronomical images using cloud-based containerized applications, 2018 IEEE 14th International Conference on Intelligent Computer Communication and Processing (ICCP), pp. 371-376
• Stefanut, T., Bacu, V., Nandra, C., Balasz, D., Gorgan, D., Vaduvescu, O., NEARBY Platform: Algorithm for automated asteroids detection in astronomical images, 2018 IEEE 14th International Conference on Intelligent Computer Communication and Processing (ICCP), pp. 365-369
• Bacu, V.; Nandra, C.; Sabou, A.; Stefanut, T.; Gorgan, D. Assessment of Asteroid Classification Using Deep Convolutional Neural Networks. Aerospace 2023, 10, 752. https://doi.org/10.3390/aerospace10090752

How to cite: Bacu, V.: Enhancement of the NEARBY automated asteroid detection platform with a machine learning-based classifier, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-8018, https://doi.org/10.5194/egusphere-egu24-8018, 2024.

Physics-informed neural networks (PINNs) provide a novel approach for data-driven numerical simulations, tackling challenges of discretization and enabling seamless integration of noisy data and physical models (e.g., partial differential equations). In this presentation, we discuss the results of our recent studies where we apply PINNs for coronal magnetic field extrapolations of the solar atmosphere, which are essential to understand the genesis and initiation of solar eruptions and to predict the occurrence of high-energy events from our Sun.
We utilize our PINN to estimate the 3D coronal magnetic fields based on photospheric vector magnetograms and the force-free physical model. This approach provides state-of-the-art coronal magnetic field extrapolations in quasi real-time. We simulate the evolution of Active Region NOAA 11158 over 5 continuous days, where the derived time profile of the free magnetic energy unambiguously relates to the observed flare activity.
We extend this approach by utilizing multi-height magnetic field measurements and combine them in a single magnetic field model. Our evaluation shows that the additional chromospheric field information leads to a more realistic approximation of the solar coronal magnetic field. In addition, our method intrinsically provides an estimate of the height corrugation of the observed magnetograms.
We provide an outlook on our ongoing work where we use PINNs for global force-free magnetic field extrapolations. This approach enables a novel understanding of the global magnetic topology with a realistic treatment of current carrying fields.
In summary, PINNs have the potential to greatly advance the field of numerical simulations, accelerate scientific research, and enable advanced space weather monitoring.

How to cite: Jarolim, R., Tremblay, B., Rempel, M., Thalmann, J., Veronig, A., Molnar, M., and Podladchikova, T.: Physics-informed neural networks for advanced solar magnetic field extrapolations, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-12961, https://doi.org/10.5194/egusphere-egu24-12961, 2024.