Reduced Order Probabilistic Emulator of RAM‐SCB: Toward Non‐Linearity With Autoencoders

Abstract An accurate understanding of the Earth's ring current dynamics is integral to predicting the impacts of geomagnetic storms. We use physics‐based models to simulate the ring current such as the Ring‐current Atmosphere interactions Model with Self‐Consistent Magnetic Field (RAM‐SCB). However, physics‐based models are computationally expensive. Therefore, we employ reduced order models (ROMs) to speed up computation. In this study, we present a ROM of RAM‐SCB using autoencoder neural networks (AE) as well as orthogonal autoencoders (OAE) and compare it to a previous study which used Principal Component Analysis (PCA). We also use this ROM to produce a Reduced Order Probabilistic Emulator (ROPE), where we use an ensemble Long Short‐Term Memory (LSTM) in combination with ResNets (LSTM + ResNet) neural networks to emulate RAM‐SCB in the reduced state. We show a significant improvement in using AEs over PCA and present a ROPE that can forecast the ring current particle fluxes 1 hour ahead.