Neural Network for Solar Irradiance Modeling (NN-SIM)

Abstract

An understanding of solar variability over a broad range of wavelengths and timescales is needed by scientists studying Earth’s climate. While the current understanding of solar irradiance from measurements and models is maturing, there remain notable areas of discrepancy that highlight a lack of understanding of the variability in solar spectral irradiance (SSI) on 27-day solar-rotational timescales and longer, and in total solar irradiance (TSI) at solar-cycle timescales and longer. The sources of instrumental noise and instability suspected behind differences in independent measurement records are actively debated. Furthermore, estimates from solar-irradiance empirical-proxy models and semi-empirical models also differ from each other and from the observations by varying degrees. To investigate whether models and observations can be brought into closer agreement we developed a novel, data-driven, solar-irradiance model using an ensemble of feed-forward artificial neural networks. Key features of our model architecture include a non-linear relationship between solar-activity proxy and irradiance with a high degree of freedom that comes from the incorporation of a greater number of solar-activity proxies than previous proxy models. Furthermore, we utilize a recent re-analysis of solar spectral irradiance (SSI) observations, stemming from a new degradation-correction methodology, to develop our model. Our approach, the Neural Network for Solar Irradiance Modeling (NN-SIM), reconstructs total solar irradiance and SSI from 205 nm to 2300 nm and from 1979 to the present day. We find close agreement between NN-SIM and various observational records as well as independent models. NN-SIM is available at .