Model description

Methods

The Model is based on machine learning. The IDLmlSupportVectorMachineClassification (SVM) is used for data processing, and the IDLmlFeedForwardNeuralNetwork (FFNN) is used for model derivation.

Data

Only MAVEN-measured data were used in this model. Every data point consisting of Θ, P_sw, F and B_surf was classified as one of the three regions (solar wind, magnetosheath, or magnetosphere) according to simple conditions (see Němec et al. 2020). The unclassified data points were subsequently classified using SVM.

To create the FFNN model, the data points were modified according to the following conditions: The used data points are within 0.5 R_M from the expected boundary location (Němec et al. 2020). The data with Θ > 90^o are not considered for bow shock, and the data with Θ > 105^o are not considered for magnetic pile-up. The area was divided into 5^o sectors within Θ. The number of data points in front of and behind the considered boundary is equalized using the random over-sampling method within individual sectors. Finally, the data points are normalized using IDLmlVarianceNormalizer, applying a logarithm to both P and B.

Model

This FFNN model predicts whether a given point (R_M, Θ, P_sw, F, and B_surf) is in front or behind the boundary. The FFNN structure consists of three hidden layers: each one with seven neurons. The four activation functions consist of three IDLmlaftanh functions, and the last one is the IDLmlafSoftmax function.