A Deep Learning‐Based Approach for Modeling the Dynamics of AMPERE Birkeland Currents

Abstract

The existence of Birkeland magnetic field‐aligned current (FAC) system was proposed more than a century ago, and it has been of immense interest for investigating the nature of solar wind‐magnetosphere‐ionosphere coupling ever since. In this paper, we present the first application of deep learning architecture for modeling the Birkeland currents using data from the Active Magnetosphere and Planetary Electrodynamics Response Experiment (AMPERE). The model uses a 1‐hr time history of several different parameters such as interplanetary magnetic field (IMF), solar wind, and geomagnetic and solar indices as inputs to determine the global distribution of Birkeland currents in the Northern Hemisphere. We present a comparison between our model and bin‐averaged statistical patterns under steady IMF conditions and also when the IMF is variable. Our deep learning model shows good agreement with the bin‐averaged patterns, capturing several prominent large‐scale features such as the Regions 1 and 2 FACs, the NBZ current system, and the cusp currents along with their seasonal variations. However, when IMF and solar wind conditions are not stable, our model provides a more accurate view of the time‐dependent evolution of Birkeland currents. The reconfiguration of the FACs following an abrupt change in IMF orientation can be traced in its details. The magnitude of FACs is found to evolve with e‐folding times that vary with season and MLT. When IMF Bz turns southward after a prolonged northward orientation, NBZ currents decay exponentially with an e‐folding time of ∼25 min, whereas Region 1 currents grow with an e‐folding time of 6–20 min depending on the MLT.

Forecasting the Probability of Large Rates of Change of the Geomagnetic Field in the UK: Timescales, Horizons, and Thresholds

“…We also wish to provide the models with the time history of the solar wind. In principle, this can be either done by providing the time series data to the models explicitly and letting the model learn the most important information (e.g., Kunduri et al., 2020), or by using features that describe the variability in a time window (e.g., Camporeale et al., 2020). In this work we provide the time history explicitly, testing the ability of several different neural network architectures to extract the important and necessary information from the rich input data.…”

“…For space weather forecasting in particular, models can be structured to have a “memory” of the preceding solar wind conditions (Bhaskar & Vichare, 2019; Kugblenu et al., 1999). This has typically been done through the use of recurrent layers (Gruet et al., 2018; Keesee et al., 2020; Liu et al., 2020; Tan et al., 2018; Wu & Lundstedt, 1996), or by using filters on the historical data to extract important information, convolution layers for example (Kunduri et al., 2020; Siciliano et al., 2020).…”

“…Often this has taken the form of forecasting a geomagnetic index (Liemohn et al, 2018), for example the Sym-H (Siciliano et al, 2020;Bhaskar & Vichare, 2019), Dst/Est (Chandorkar et al, 2017;Kugblenu et al, 1999;Lethy et al, 2018;Lundstedt et al, 2002;Wu & Lundstedt, 1996;Gruet et al, 2018;Wintoft & Wik, 2018;Tasistro-Hart et al, 2021) or Kp indices (Zhelavskaya et al, 2019;Ji et al, 2013;Tan et al, 2018;Wing et al, 2005;Wintoft et al, 2017). Models have also been produced that aim to predict phenomena such as ionospheric current systems (Kunduri et al, 2020), geomagnetic storms (Chakraborty & Morley, 2020), substorms (Maimaiti et al, 2019) or SSCs (Smith, Rae, Forsyth, Oliveira, et al, 2020).…”

“…In principle, this can be either done by providing the time series data to the models explicitly and letting the model learn the most important information (e.g. Kunduri et al, 2020), or by using features that describe the variability in a time window (e.g. .…”

“…Often this has taken the form of forecasting a geomagnetic index (Liemohn et al., 2018), for example, the Sym‐H (Bhaskar & Vichare, 2019; Siciliano et al., 2020), Dst/Est (Chandorkar et al., 2017; Gruet et al., 2018; Kugblenu et al., 1999; Lethy et al., 2018; Lundstedt et al., 2002; Tasistro‐Hart et al., 2021; Wintoft & Wik, 2018; Wu & Lundstedt, 1996) or Kp indices (Ji et al., 2013; Tan et al., 2018; Wing et al., 2005; Wintoft et al., 2017; Zhelavskaya et al., 2019). Models have also been produced that aim to predict phenomena such as ionospheric current systems (Kunduri et al., 2020), geomagnetic storms (Chakraborty & Morley, 2020), substorms (Maimaiti et al., 2019) or SSCs (Smith, Rae, Forsyth, Oliveira, et al., 2020). On a local level, studies have also looked at forecasting the geomagnetic perturbations at magnetometer stations (Camporeale et al., 2020; Keesee et al., 2020; Wintoft et al., 2015).…”