Improving the accuracy of rapid Earth Orientation Parameters with the "ResLearner" machine learning method

Determination of Earth Orientation Parameters (EOP) with utmost accuracy requires the combination of various data sources from different space geodetic techniques, some of which requiring long processing time. This results in a latency of up to several weeks by which the so-called final EOP are released. Since some of the important applications, including satellite navigation and orientation of deep space telescopes, require instantaneous EOP information, the so-called rapid determination and also prediction of EOP are needed. International Earth Rotation and Reference Systems Service (IERS) provides rapid EOP by using the most recent Global Positioning System (GPS) and Very Long Baseline Interferometry (VLBI) 24-hour and intensive sessions data. However, there are some discrepancies between these rapid data and the final, most accurate EOP. In order to reduce these discrepancies and achieve more accurate rapid EOP, we focus on applying machine learning algorithms for polar motion components (xp, yp) and dUT1=UT1-UTC. We focus on a window of 63 days with 31 day predictions to the past and 31 day predictions to the future.

We devise a new algorithm called ResLearner, which is a machine learning method based on multilayer perceptrons trying to learn the differences between rapid and final EOP data. We use informative features such as Effective Angular Momentum (EAM) data (both the observations provided by GFZ and the 14-day forecasts provided by ETH Zurich), tides, Liouville equation for (xp, yp), and a linear relation between dUT1 and the axial components of EAM.
We use ResLearner in the context of Deep Ensembles in order to derive the uncertainty in the estimations. We also address the so-called unmixing and self-calibration problems. The former enables us to unravel the causes behind the discrepancies between rapid and final EOP as provided by IERS, while the latter could help us reduce these erroneous effects. 

We train the algorithms on both the IERS and Jet Propulsion Laboratory (JPL) final EOP data. Our ResLearner method can consistently reduce the discrepancies between rapid and final EOP across all days. The improvement in accuracy is up to 55%. We observe some unexpected behaviour related to day 0 of prediction, in which the accuracy of the IERS is significantly better than the immediately preceding or following values. Our unmixing algorithm shows that this behaviour is probably related to erroneous, non-linear effects of EAM at day 0, and also semi-diurnal, diurnal, long-period retrograde and prograde, and zonal tides. Using our self-calibration algorithm for EAM, we can slightly improve the prediction performance by up to 14%.

Finally, we provide the improved rapid EOP data publicly, operationally, and on a daily basis, on the ETH Zurich prediction center website at https://gpc.ethz.ch/EOP/Rapid/