The radiation belts of the Earth, filled with energetic electrons, comprise complex and dynamic systems that pose a significant threat to a variety of satellite systems. While various models of the relativistic electron flux have been developed for geostationary orbit (GEO), the behaviour of the medium energy (120-600 keV) electrons below GEO remains poorly quantified. In this paper we present a Medium Energy electRon flux In Earth’s outer radiatioN belt (MERLIN) model based on the Light Gradient Boosting (LightGBM) machine learning algorithm. The MERLIN model takes as input the satellite position, a combination of geomagnetic indices and solar wind parameters including the time history of velocity, and does not use persistence. MERLIN is trained and validated on $>$15 years of the GPS electron flux data, and tested on more than $1.5$ years of measurements. 10-fold cross validation (CV) yields that the model predicts the MEO radiation environment well, both in terms of dynamics and amplitudes of flux. Evaluation on the test set yields high correlation between the predicted and observed electron flux (0.8) and low values of absolute error. The MERLIN model can have wide Space Weather applications, providing information for the scientific community in the form of radiation belts reconstructions, as well as industry for satellite mission design, nowcast of the MEO environment and surface charging analysis.

In recent years, feedforward neural networks (NNs) have been successfully applied to reconstruct global plasmasphere dynamics in the equatorial plane. These neural network-based models capture the large-scale dynamics of the plasmasphere, such as plume formation and erosion of the plasmasphere on the nightside. However, their performance depends strongly on the availability of training data. When the data coverage is limited or non-existent, as occurs during geomagnetic storms, the performance of NNs significantly decreases, as networks inherently cannot learn from the limited number of examples. This limitation can be overcome by employing physics-based modeling during strong geomagnetic storms. Physics-based models show a stable performance during periods of disturbed geomagnetic activity, if they are correctly initialized and configured. In this study, we illustrate how to combine the neural network- and physics-based models of the plasmasphere in an optimal way by using the data assimilation Kalman filtering. The proposed approach utilizes advantages of both neural network- and physics-based modeling and produces global plasma density reconstructions for both quiet and disturbed geomagnetic activity, including extreme geomagnetic storms. We validate the models quantitatively by comparing their output to the in-situ density measurements from RBSP-A for an 18-month out-of-sample period from 30 June 2016 to 01 January 2018, and computing performance metrics. To validate the global density reconstructions qualitatively, we compare them to the IMAGE EUV images of the He+ particle distribution in the Earth’s plasmasphere for a number of events in the past, including the Halloween storm in 2003.

Understanding the dynamic evolution of relativistic electrons in the Earth’s radiation belts during both storm and non-storm times is a challenging task. The U.S. National Science Foundation’s Geospace Environment Modeling (GEM) focus group “Quantitative Assessment of Radiation Belt Modeling” (QARBM) has selected two storm time and two non-storm time events that occurred during the second year of the Van Allen Probes mission for in-depth study. Here, we perform simulations for these GEM challenge events using the 3-Dimensional Versatile Electron Radiation Belt (VERB-3D) code. We set up the outer $L^*$ boundary using data from Geostationary Operational Environmental Satellites (GOES) and validate the simulation results against satellite observations from both the GOES and Van Allen Probe missions for 0.9 MeV electrons. Our results show that the position of the plasmapause plays a significant role in the dynamic evolution of relativistic electrons. The magnetopause shadowing effect is included by using last closed drift shell (LCDS), and it is shown to significantly contribute to the dropouts of relativistic electrons at high $L^*$.

The magnetosphere, ionosphere and thermosphere (MIT) act as a coherently integrated system (geospace), driven in part by solar influences and characterized by variability and complexity. Among the most important and yet uncertain aspects of the geospace system is energy and momentum coupling between regions, which is, in part, accomplished by the transfer of charged particles from the magnetosphere to the ionosphere in a process known as particle precipitation, and in the opposite direction by ion outflow. Both processes are inherently multiscale and manifest the variabilities and complexities of the geospace system. Despite the importance of the transfer of particles, existing models are increasingly ill-equipped to provide the specification necessary for the growing demand for geospace now- and forecasts. Due to recent trends in the availability of data, we now face an exciting opportunity to progress particle transfer in geospace through the intersection of traditional approaches and state-of-the-art data-driven sciences. We reveal novel particle transfer models utilizing machine learning (ML), present results from the models, and provide an evaluation of their capabilities including comparisons with observations and the current ’state-of-the-art’ models (e.g., OVATION Prime for particle precipitation and the Gamera-Ionosphere Polar Wind Model for ion outflow). We detail the data wrangling required to utilize the available geospace observations to make progress on the long-standing challenge of particle transfer and place specific emphasis on the discovery possible when ML models are appropriate and robustly interrogated in the context of physical understanding. Our presentation helps illustrate the trends in the application of data science in space science.

We present a comparison of magnetospheric plasma mass/electron density observations during an 11-day interval which includes the geomagnetic storm of 22 June 2015. For this study we used: equatorial plasma mass density derived from geomagnetic field line resonances (FLRs) detected by Van Allen Probes and at the ground-based magnetometer networks EMMA and CARISMA; in situ electron density inferred by the Neural-network-based Upper hybrid Resonance Determination algorithm applied to plasma wave Van Allen Probes measurements. The combined observations at L ~ 4, MLT ~ 16 of the two longitudinally-separated magnetometer networks show a temporal pattern very similar to that of the in situ observations: a density decrease by an order of magnitude about 1 day after the Dst minimum, a partial recovery a few hours later, and a new strong decrease soon after. The observations are consistent with the position of the measurement points with respect to the plasmasphere boundary as derived by a plasmapause test particle simulation. A comparison between plasma mass densities derived from ground and in situ FLR observations during favourable conjunctions shows a good agreement. We find however, for L < ~3, the spacecraft measurements to be higher than the corresponding ground observations with increasing deviation with decreasing L, which might be related to the rapid outbound spacecraft motion in that region. A statistical analysis of the average ion mass using simultaneous spacecraft measurements of mass and electron density indicates values close to 1 amu in plasmasphere and higher values (~ 2-3 amu) in plasmatrough.

The Neural Decision Tree (NDT) is a hybrid supervised machine-learning algorithm that combines the self-limiting property of a decision tree (CART) algorithm with the artificial neural network (ANN). We demonstrate the use of NDT for a regression problem of building a prediction model for the plasmasphere electron density with solar and geomagnetic measurements as inputs. Our work replicates the work by Zhelavskaya et al. reported in their 2017 article to show that NDT makes available sophisticated network layout for building a predictive model, thus taking advantage of the deep-learning potential of the neural network. We also demonstrate that with the ability to automatically select an appropriate network layout, as well as, effective initialization, the NDT algorithm allows research scientists in space weather to focus more of their attention on physically and statistically relevant aspects of using machine-learning techniques. In fact, our example highlights the fact that the basic assumptions of standard supervise machine-learning problems are often unsatisfied in real-world space weather applications. Greater attention to these fundamental issues may create significantly different solutions to space weather forecast problems.