Speaker
Description
Space infrastructures increasingly need highly efficient monitoring systems that detect rare and energetic events. Deep learning pipelines are one of the most promising and innovative approaches towards the development of such systems. However, leveraging machine learning to detect or predict rare events requires careful data preparation, as the training process significantly impacts the performance of the neural network architecture.
To illustrate the intricacies of this approach, we will consider the example of forecasting geomagnetic storms using data that includes typical quantities collected by satellites like the China Seismo-Electromagnetic Satellite (CSES). The CSES has proven its ability to observe various space weather phenomena, including solar energetic particle events, solar flares, and geomagnetic storms. The CSES captures a wide range of data through a multi-payload approach, including energetic particles spanning three orders of magnitude, electromagnetic fields, and plasma density. Soon to expand into a constellation of two satellites, CSES will become a powerful resource for observational astrophysics.
We will show that the strategy proposed for creating datasets for training and validation plays a fundamental role in guaranteeing the capability of giving correct forecasting of stormy and disturbed geomagnetic periods.
