INTERGEO Agenda - english

Accurate regional gravity field modeling is essential for the advancement of modern geodetic infrastructure. In particular, a high-resolution geoid determination is fundamental for the establishment of the International Height Reference Frame (IHRF), ensuring consistent physical heights on global scales and enabling global applications such as sea-level monitoring and mass variation observation. While classical methods, such as integral formulas or harmonic analysis, reach their limits regarding the efficient combination of large datasets and heterogeneous boundary conditions, machine learning methods offer new opportunities. This contribution investigates the potential of Physics-Informed Neural Networks (PINNs) for modeling the regional geoid. The focus of the study lies on the development of a framework for the topographically challenging Colorado test area. In contrast to purely data-driven approaches, PINNs integrate the Laplace equation and boundary conditions directly into the neural network's loss function, enabling a consistent coupling of the model and geodetic observations. The validation of the model is initially based on synthetic data. These data consist of disturbance potential and gravity anomalies derived from the GGM EIGEN-6C4 up to a maximum degree and order of 2190. Particular emphasis is placed on robustness analysis, specifically examining the sensitivity of the PINN model to artificially noisy input data. This aims to test the hypothesis that the implicit physical regularization of PINNs provides a filtering effect compared to conventional methods, thereby enhancing the reconstruction quality of the potential. To quantify the added value of physical information, a comprehensive benchmark analysis is conducted, comparing the PINN model to a classical, purely data-driven Multilayer Perceptron (MLP). Finally, we discuss the transferability of these findings to dense terrestrial and satellite-based datasets, positioning deep learning as a robust, complementary tool for physical geodesy.