Application of Artificial Intelligence in Satellite Orbit and Clock Correction: A Review

Abstract

Aims: This review examines AI-assisted methods for satellite orbit determination and prediction, clock-bias forecasting, thermospheric-density calibration, atmospheric-delay modelling, and integrated GNSS correction. It identifies recurrently successful approaches, limitations in the current evidence, and priorities for future research. Study Design: A structured narrative synthesis of foundational and recent literature published through early 2026 was undertaken. Targeted keyword searches and backward citation checking were used to identify studies of classical machine learning, recurrent and attention-based networks, ensembles, graph neural networks, and physics-informed neural networks. Evidence was organised by task, data regime, baseline, forecast horizon, metric, and reported limitation; no pooled meta-analysis was performed. Results: Across the reviewed orbit studies, hybrid residual-learning schemes were the most recurrently successful design, although no pooled effect was estimated. In one thermospheric-density evaluation, AI calibration improved mean absolute percentage error by 61% against NRLMSISE-00 and 39% against JB-08. In one one-day GPS clock experiment, QP-assisted BPNN, WNN, LSTM, and GRU models improved on quadratic polynomial prediction by about 39%, 58%, 27%, and 29%, respectively. In a separate four-city dataset, a graph-based model reduced horizontal positioning error by 40-80% relative to the tested deep-learning baselines. These values are study-specific, not general effect estimates. Common barriers were limited cross-satellite generalisation, scarce precise LEO data, computational constraints, weak uncertainty calibration, and separate orbit and clock pipelines. Conclusion: AI can improve specific parts of the orbit and clock correction chain when models are evaluated against suitable physical or statistical baselines. The evidence favours transferable, physics-informed, uncertainty-aware, and computationally efficient joint models, supported by common benchmarks and independent validation.