Development of a High-Latitude Convection Model by Application of Machine Learning to SuperDARN Observations

A new model of northern hemisphere high-latitude convection derived using machine learning (ML) is presented. The ML algorithm random forests regression was applied to a database of velocities derived from the Super Dual Auroral Radar Network (SuperDARN) observations processed with the potential mapping technique, Map-Potential (Ruohoniemi & Baker, 1998, https://doi.org/10.1029/98ja01288). The features used to train the model were the interplanetary magnetic field components B_x, B_y, and B_z; the solar wind velocity, v_sw; the auroral indices, A_u and A_l; and the geomagnetic index, SYM-H. The SuperDARN velocities were separated into north-south, and east-west components and sorted into a magnetic local time - magnetic latitude grid that ran from 55° to the magnetic pole with a bin size of 2° in latitude, and 1-hr in MLT. Separate models were created for each velocity component in each bin of the grid. It is found that even though the models in each bin are independent of one another a coherent convection pattern is formed when the models are viewed in aggregate. The resulting convection pattern responds to changes in the auroral indices by expanding and contracting in a way that is consistent with expectations for a substorm cycle. Further it is found that the mean-squared difference between predictions of the model and observed values of the velocity are substantially lower than the same quantity calculated for an existing climatology that was not formed with ML techniques.