Authors: Kimberly Moreland (University of Texas at San Antonio/Southwest Research Institute), Maher Dayeh (Southwest Research Institute/University of Texas at San Antonio), Subhamoy Chatterjee (Southwest Research Institute), Andres Munoz-Jaramillo (Southwest Research Institute), Hazel Bain (CIRES, NOAA/SWPC)

We present a new approach to forecasting solar energetic particle (SEP) events and their subsequent properties using a multivariate ensemble of Convolutional Neural Networks (CNNs) to estimate the true probability of the SEP occurrence. We focus on providing a probabilistic outcome instead of a binary classification for the event occurrence. Furthermore, rather than imposing a detection threshold that optimizes the classification, we provide flexibility to the user of our forecast to determine their acceptable level of risk. We turn the class imbalance of negative and positive events into an advantage by training an ensemble of models, this provides a clear and robust measure of uncertainty in our forecast. Each model ingests a series of remote imaging and in situ data from a dataset we specifically assembled for machine learning applications. After training the classification branch, we freeze it and train the regression branch to predict SEP physical properties (e.g., peak, spectral profile). This allows the classification branch’s outcome to penalize the regression branch’s loss function. Preliminary analysis shows that the combination of remote and in situ data improves SEP occurrence validation accuracy by 2% and reasonably predicts SEP event properties.