A machine-learning-based approach for predicting the geomagnetic secular variation

Abstract

We present a machine-learning-based approach for predicting the geomagnetic main field changes, known as secular variation (SV), in a 5-year range for use for the 14th generation of International Geomagnetic Reference Field (IGRF-14). The training and test datasets of the machine learning (ML) models are geomagnetic field snapshots derived from magnetic observatory hourly means, and CHAMP and Swarm-A satellite data (MCM Model; Ropp et al., 2020). The geomagnetic field data are not used as-is in the original time series but were differenced twice before training. Because SV is strongly influenced by the geodynamo process occurring in the Earth's outer core, challenges still persist despite efforts to model and forecast the realistic nonlinear behaviors (such as the geomagnetic jerks) of the geodynamo through data assimilation. We compare three physics-uninformed ML models, namely, the Autoregressive (AR) model, Vector Autoregressive (VAR) model, and Recurrent Neural Network (RNN) model, to represent the short-term temporal evolution of the geomagnetic main field on the Earth’s surface. The quality of 5-year predictions is tested by the hindcast results for the learning window from 2004.50 to 2014.25. These tests show that the forecast performance of our ML model is comparable with that of candidate models of IGRF-13 in terms of data misfits after the release epoch (Year 2014.75). It is found that all three ML models give 5-year prediction errors of less than 100nT, among which the RNN model shows a slightly better accuracy. They also suggest that Overfitting to the training data used is an undesirable machine learning behavior that occurs when the RNN model gives accurate reproduction of training data but not for forecasting targets.