Abstract A machine learning emulator has been developed to efficiently reproduce electron densities from a three‐dimensional, data‐driven, regional simulation of postsunset ionospheric irregularities associated with equatorial spread F (ESF). The simulation has previously been shown to produce accurate ESF conditions observed at the Jicamarca Radio Observatory (JRO). Its high computational cost limits its ability to run in real‐time or serve as an operational forecast of ESF. Using data collected from JRO's newly implemented medium‐power ISR mode to drive the simulation, a large collection of simulation results has been developed for use as training data in a machine learning emulator. The emulator combines dimensionality reduction techniques and time series forecasting with long‐short‐term memory neural networks. Specifically, principal component analysis and a convolutional autoencoder are used to transform electron density results to a 20‐dimensional latent space. A new “noisy training” method is used to improve dynamic time series forecasting within the latent space. The resulting forecasts show both qualitatively and quantitatively accurate density reconstructions over 30–60 min periods. The significant speedup provided by the emulator over the simulation motivates future developments toward a real‐time forecast of ESF.
Kirchman et al. (Mon,) studied this question.