Advances in Operational and Research Earth Orientation Parameters Prediction at BKG: Hybrid and Physics-Informed Approaches

Accurate and low-latency Earth Orientation Parameters (EOP) are essential for precise transformations between terrestrial and celestial reference frames, supporting satellite navigation, space missions, and geodetic and astronomical applications. Since official IERS EOP products are available with inherent delays of hours to days, robust short-term EOP prediction remains a critical operational requirement.

This contribution presents recent operational and research developments at the Federal Agency for Cartography and Geodesy (BKG) in cooperation with the University of Alicante (UA), focusing on machine-learning (ML) and deep-learning (DL) approaches for EOP prediction that exploit effective angular momentum (EAM) forecasts from GFZ as physically motivated input parameters. The prediction framework is driven by a comprehensive set of technique-specific (VLBI, GNSS, SLR) and multi-technique combined EOP products generated at BKG, complemented by the official IERS EOP reference series for training, validation, and benchmarking. The approach builds on BKG’s established hybrid prediction system, in which deterministic signals are modeled using Singular spectrum analysis and least squares, while stochastic variability is traditionally captured via autoregressive and Copula-based analysis models. In the proposed framework, ML/DL architectures, such as multi-task networks for polar motion and dUT1 prediction and convolutional models for short-term LOD forecasting are employed to replace or augment the stochastic component, without imposing explicit physical constraints within the learning process. Results demonstrate that combining EAM-based predictors with BKG’s technique-specific and multi-technique EOP products leads to systematic improvements in short-term (1–10 day) prediction accuracy compared to purely data-driven baselines. The EAM-based ML/DL framework has been under operational testing at BKG since early 2025 and represents a significant step toward an operational ML-supported EOP prediction service, with ongoing work addressing full EOP integration and impact assessment on VLBI analysis and satellite orbit determination.