The geodynamo is a dynamic process, involving convective motion in the Earth's fluid outer core, which is responsible for generating and maintaining the main magnetic field of the planet. That magnetic field changes on timescales ranging from decades to millions of years. On shorter timescales, the detailed morphology of the field can vary, while on millennial timescales, dramatic changes, such as reversals in the polarity of the axial dipole, can occur. We develop and examine tools for understanding and predicting variations in the Earth's magnetic field on each of these timescales.

For shorter timescales, we explore the growing field of geomagnetic data assimilation (GDA), in which observations of the geomagnetic field are merged with numerical geodynamo models to estimate the dynamic state of the outer core and initialize forecasts of decadal-scale variations. We develop a proxy model, which is simpler and less computationally expensive than a numerical geodynamo, that allows us to systematically explore the challenges facing GDA through extensive numerical experiments. The outcome is a first of its kind testing environment for GDA and an accompanying first round of numerical experiments. The results lead us to propose an assimilation scheme for consideration in operational GDA systems and outline a path for future GDA development.

On longer timescales, we investigate predictions of major excursions and reversals of the magnetic field. This is done through considering a hierarchy of paleomagnetic reconstructions and numerical models ranging from simple scalar models to 3D geodynamos. We examine a set of prediction strategies using support-vector machines (SVMs), long short-term memory (LSTM) networks and a simple thresholding of the axial dipole field intensity. The results lead us to explicitly characterize differences in the way excursions and reversals occur among the hierarchy of reconstructions and models. This motivates the proposal of new criteria for identifying Earth-like models.