Reconstructing Historical Solar Indices for Predicting Past Space Weather Events

Modeling and forecasting the near-Earth space environment, specifically the thermosphere, is particularly important because it affects the motion of low-Earth orbit objects through atmospheric drag. Solar indices such as F10.7, S10.7, M10.7, and Y10.7 are commonly used as inputs to ionospheric and thermospheric density models. Availability of such indices during extreme space weather events is particularly important; however, due to the limited extent of these indices and the relative rarity of extreme events, these indices do not capture many strong events. Unfortunately, this challenges evaluating model performance for historical solar events in terms of atmospheric drag effects resulting from extreme solar forcing. In this study, we introduce a method to reconstruct historical solar indices back to 1826 using a Long Short-Term Memory based machine learning approach. The model is based on three long-standing solar activity indicators—sunspot number, sunspot area, and solar plage area. Training was conducted using overlapping periods of modern observations, and the model's performance was the highest when the three solar indicators were taken simultaneously to train the model. The minimum root mean square error (RMSE) while predicting F10.7, S10.7, M10.7, and Y10.7 are 13.29, 13.96, 14.4, and 9.65 solar flux units, respectively, for the 1997–2001 interval. This extended historical record will enable testing and validating of empirical and physics-based models under a wide range of solar conditions, including extreme events, contributing to improved operational forecasting and space situational awareness.