Machine Learning (ML) applications are conquering new areas taking advantage of the availability of large amounts of data. Considering space probes such as SOHO or ACE that are providing information about the solar wind at the L1 point for more than 25 years, it is feasible that ML approaches can be applied to the Space Weather (SW) [2-3] context. Examples of ML techniques have been published recently, aiming to forecast the solar events that can affect the Earth’s magnetosphere [6-12]. Particularly, assessing the forecasting of geomagnetic indices (that quantify the disturbance of the Earth's magnetosphere) such as the Dst or the higher resolution ones, such as the SYM-H or ASY-H indices. The usage of ML approaches has gained importance, as seen in recent surveys and guidelines [4, 5]. Using these techniques to drive operational geomagnetic indices forecasting systems can minimize the damages caused by geomagnetic storms, providing an alert ahead enough to take containment measures. The SW prediction can be seen as a time series forecasting problem; thus, applying the state of the art in Artificial Neural Networks (ANN) seems to be the next-generation algorithms for SW forecasting. However, the current ML models are limited to a prediction at a single time horizon, specifically, at the next hour, and mainly focused on SYM-H. We propose research on: - The development of novel ANN architectures (i.e. encoder-decoder) to forecast SYM-H and ASY-H at multiple time-steps, extending both the current maximum prediction (two hours or up) and its accuracy. - Models that can be used in real-time operation since most current models are not optimized to handle it. The solutions should be resistant to missing data, allowing forecasting in less-than-ideal situations, when one or more variables are missing due to the instrument's saturation. - Forecasting regional indices, such as mid-latitude indices already included in the ESA Portfolio, which has not been addressed yet.
