Authors: Lidiya Y. Ahmed1,2 , Michael L. Stevens1, Kristoff Paulson1, Anthony Case1 1 Smithsonian Astrophysical Observatory 2 Harvard University

Solar wind parameters such as density, temperature, and speed are typically estimated by peak fitting or performing approximate velocity moment sums on charged particle spectra. Vech et al (2021) showed that these traditional methods can be bypassed by a well-trained artificial neural network (ANN). Lacking ground truth, however, the Vech et al algorithm was trained to analyze simulated particle spectra for a perfectly calibrated instrument. In this study, we propose a neural network approach to estimate solar wind parameters from low level data without being converted to physical units. Among other advantages, this technique avoids assumptions such as perfect calibration or a particular form for the ion distribution function. Dynamic time warping (DTW) is used to prepare ground truth from measurements provided by a well-calibrated, similarly instrumented reference spacecraft in a similar orbit (Wind), and the ANN is trained to predict the Wind solar wind parameters from the uncalibrated DSCOVR PlasMag FC data. Applications for Parker Solar Probe, Helioswarm, and other missions as well as L1 monitoring are discussed.