Attempting to predict the time until a reversal using machine learning techniques.
Data is previously computed in MATLAB. Data is an integration of an SDE of the form
where the drift term is
the noise term is
and the noise process is an Ornstein-Uhlenbeck process with characteristic time
.
These values are fitted through comparison to the paleomagnetic model PADM2M (Ziegler et al., 2011). Drift and noise functions are displayed below.
Observed quantity is the amplitude of the axial dipole field. The quantity to be predicted from this is the "time until next polarity reversal".
Statistical features are calculated using a moving window over the ADM.
A random forest method similar to the methodology of Rouet‐Leduc et al., 2017 is used. Hyperparameters have not been explored or tuned yet.
Preliminary results are not great. I believe this is because there is insufficient information in the SDE integration to accurately predict reversals. Physically richer models may provide the depth of information needed for predictions.
