Radio astronomy: an XGBoost classifier detects calibration anomalies in ALMA data more effectively than the pipeline rules

Trained on a vast dataset of verified calibrations, the model significantly reduces detection errors and could be used in next-generation radio telescopes.

To produce usable images, a radio interferometer like ALMA in Chile must correct the response of its instruments over the entire frequency range observed: this is called bandwidth calibration. Some of these calibration solutions contain anomalies that, if they go unnoticed, degrade the scientific products delivered to astronomers. Observatory pipelines today rely on simple thresholds applied to isolated indicators, a method that is losing precision as observation configurations multiply.

A team of eleven researchers proposes in a paper filed on arXiv to treat detection as a supervised classification task. An XGBoost model combines several features, including kernel scan statistics that localize anomalous spectral intervals, to learn from nonlinear decision boundaries. Trained on a large set of hand-controlled ALMA calibrations, it makes significantly fewer classification errors than current amplitude heuristics.

The authors present it as an evolutionary complement to the observatory's operations, extendable to other characteristics and transposable to future radio installations.