Abstract
Machine learning approaches have shown potential to automate the creation of seismic catalogues, but their abilities and performance are still questioned and being tested. We here present the Petrinja Earthquake Sequence Machine Learning Catalogue (PESMLC), generated using the EQTransformer seismic phase picker trained on the INSTANCE dataset, PyOcto for phase association, and NonLinLoc for event relocation. Our study focuses on the 2020 MW 6.4 Petrinja earthquake sequence in Croatia, covering the period from 28 December 2020 to 30 November 2022, during which we detected 41,989 seismic events. PESMLC achieves high matching rates — 96% for the first six months, and 85% over the complete study duration — when compared with detailed manual catalogues. The seismic network densification in January 2021 significantly enhanced event detection reliability, highlighting the importance of network configuration in efficient seismic monitoring. While our machine learning approach successfully identified previously undetected microseismicity and dramatically reduced processing time, we observed systematic limitations, including magnitude estimation biases and difficulties in accurately distinguishing closely spaced events in time. Optimisation of associator parameters remains essential, and should ideally be performed specifically for each network configuration, particularly for fully automated seismic monitoring systems. We further demonstrate that raw machine learning catalogues require quality filtering prior to seismological analysis: stricter magnitude of completeness estimators result in artificially high thresholds for unfiltered machine learning catalogues, and Omori p exponents are biased downward by low-quality detections. Both effects are resolved when analysis is restricted to a quality-filtered subset, which recovers parameters consistent with high-quality manual catalogues. Nevertheless, PESMLC overall provides a detailed, comprehensive seismic dataset aligned closely with manual catalogues, demonstrating the considerable potential of machine learning to support routine seismic monitoring, particularly in resource-limited observatories and/or for long seismic sequences and large datasets.