Authors: Samuel Fordin (University of Delaware), Michael A. Shay (University of Delaware), Lynn B. Wilson III (NASA Goddard Space Flight Center), Gael Lucero (University of Delaware)

The roles that waves play at the ion gyroscale and into the turbulent dissipation range have been a major point of interest in the study of solar wind heating and dissipation for decades. However, manual methods to identify electromagnetic wave structures in time-series spacecraft data are prohibitively time-consuming. Machine learning is a natural solution to generalizing such methods to larger datasets, such as the almost 20-year-long Wind magnetic field dataset. We utilize a 1D convolutional neural network trained to identify waves in highpass-filtered Wind magnetic field data, covering the frequency range from 0.2-5.5 Hz, which includes the typical ion gyroscales at 1AU and the largest scales of the dissipation range. Approximately 1.6 million circularly-polarized wave intervals were identified from the dataset of ~90 million data intervals, covering the years 2005-2022. The number of waves per day contained underlying periodicity, corresponding to harmonics of the solar rotation period and likely incidence with the heliospheric current sheet. Based on the total sunspot number, the intervals of solar wind with sunspot numbers far from solar maximum and minimum had almost 50% more waves than the extremes of the sunspot number. Statistics of the estimated wavevector directions with respect to the directions of the mean magnetic field and solar wind velocity vectors support the argument that the majority of the identified wave intervals are Alfven/ion cyclotron waves together with a smaller population of compressive modes. The dependence on solar cycle of the number of waves per day indicates that the dynamics of the solar surface may play a role in the generation of waves.