Machine Learning-Based Ionospheric Corrections for Single-Frequency VLBI Observations

The International Celestial Reference Frame (ICRF) has traditionally relied on dual-frequency Very Long Baseline Interferometry (VLBI) observations. With the ICRF3, also a CRF solution based on single-frequency VLBI observations at K-band (24 GHz) was included for the first time. While K-band observations experience smaller dispersive effects, they have to be corrected using accurate models of the ionospheric electron content for highest precision. Traditionally, global ionospheric maps (GIMs) based on Global Navigation Satellite Systems (GNSS) observations have been used for this purpose.

The new project K-band VLBI Observations with Improved Scheduling and Ionospheric Correction (KOSMIC), conducted by a consortium of ETH Zurich, HartRAO, and TU Wien, aims to develop advanced ionospheric corrections for K-band VLBI using machine learning. KOSMIC will create new GIMs that rely on neural networks to model the ionospheric vertical total electron content from GNSS as a function of space and time. The ability of neural networks to capture non-linear and irregular relationships is expected to enhance the representation of complex ionospheric processes, particularly during periods of increased solar activity.

In this contribution, we will present initial results from the development of machine learning-based GIMs and their application to correct K-band VLBI data. We will evaluate the performance by processing the corrected VLBI data and computing metrics such as baseline length repeatability, alongside comparisons of Earth orientation parameters with external reference time series. This will allow us to gauge the suitability of machine learning for improving K-band VLBI observations and further advancing the celestial reference frame.