Discovering patterns, imbalanced classification & boundary surfaces in Heliophysics with artificial neural networks

The Sun is constantly emitting a stream of charged particles called the solar wind. As these particles approach the Earth, they interact with its geomagnetic field and cause a fast collisionless shock to form, the Earth's bow shock. Downstream of it the magnetosheath region emerges which consists of a highly turbulent environment where several transient phenomena occur such as magnetosheath jets, mirror-mode waves, local magnetic reconnection, and current sheets.

An implicit separation of the magnetosheath environment is dictated by how the upstream magnetic field is oriented with respect to the normal vector of the shock surface. Typically, when the angle between the interplanetary magnetic field (IMF) and the shock normal (theta_Bn) is less than 45 degrees, the shock and the associated magnetosheath is named quasi-parallel, while for degrees of more than 45 degrees it is called quasi-perpendicular. When the angle is small (quasi-parallel regime) the particles that get reflected from the bow shock travel further upstream along the magnetic field lines and interact with the incoming solar wind.

In this work, we present our recent results on the use of neural networks in classifying high-speed jets that are found in both quasi-parallel and quasi-perpendicular magnetosheath. The classification is done in a supervised learning manner where the input corresponds to upstream solar wind measurements (OMNIweb) and the output is the jets' class based on in-situ downstream observations of NASA's Magnetosphere Multiscale (MMS). We compare the results that are obtained against more traditional methods, and we investigate whether neural networks can outperform them.

Moving on, we continue our efforts by exploring the characterization of the whole magnetosheath environment by addressing a supervised regression problem, in which we utilize data from ESA's Cluster mission. The spacecrafts of the Cluster mission have a large enough spatial separation where we have identified several instances where one satellite is found downstream while another is upstream of the bow shock. This is particularly useful since we can introduce these measurements to an artificial neural network and attempt to characterize the whole magnetosheath environment. This process already showed preliminary that it can establish a clear connection between the variations observed upstream of the shock (i.e., foreshock or its absence) with these observed downstream.