New publications by Kiani Shahvandi et al. (2024)

Predicting Earth Orientation Parameters (EOPs) is important for a host of geoscientific applications, including satellite and space navigation, as well as orientation of deep space telescopes. Nutation components, which represent the orientation of the Earth with respect to the space (inertial) frame, are important EOPs that are typically modelled with accurate theories in celestial dynamics and dynamical astronomy. Observed nutations, on the other hand, differ slightly from these models. These differences are referred to as Celestial Pole Offsets (CPO), with amplitudes of less than 1 milliseconds of arc. Notwithstanding the small amplitude, the prediction of CPO is important for the utmost accuracy in space-geodetic applications. Yet, the problem of CPO prediction remains challenging due to the noise in the observed CPO values and the non-stationarity nature of the Free Core Nutation (FCN) signal that dominates the CPO observations.

To address this problem, Kiani Shahvandi et al. (2024a,b) designed two different approaches based on the artificial intelligence concepts. Kiani Shahvandi et al. (2024a) used the dynamical relations of CPO evolution based on the connection to high-frequency Atmospheric and Oceanic Angular Momentum functions (AAM and OAM, respectively). This resulted in a physics-informed neural networks framework that accurately modelled and predicted CPO time series. An important discovery of this study was that FCN is most probably excited by a combination of AAM and OAM, and that the variable period of FCN can be explained to some very good extent with the annual variation in AAM and OAM. On the other hand, Kiani Shahvandi et al. (2024b) used the concept of neural additive models to predict the CPO time series. The motivation behind this study was to enhance the clarity and explainability of predictions based on the so-called interpretable machine learning algorithms. An important insight from this study was that the input CPO series has a significant role in the prediction accuracy, such that with Jet Propulsion Laboratory data the best prediction accuracy can be achieved.

Overall, the studies of Kiani Shahvandi, et al. (2024a,b) represent the power of machine learning in the prediction of EOPs, and more specifically CPO time series. This has also been confirmed in Wińska et al. (2024), which motivated the group of space geodesy at ETH Zurich to implement and operationally provide the CPO predictions on its Geodetic Prediction Center website hosted at https://gpc.ethz.ch/EOP/.