From Global to Local: Fine-Tuning of Earthquake AI Detection Methods with Costa Rican Seismic Data for Improved Phase Detection Performance

Seismology has advanced significantly with the integration of artificial intelligence. In seismic phase detection, a key step in constructing earthquake catalogs, modern tools can now identify picks with high accuracy and reliability. EQTransformer is a deep learning model originally trained on STEAD, a global seismic dataset. However, its performance may be limited in regions underrepresented in the training data. This study explores the fine-tuning of EQTransformer using Costa Rican seismic waveforms, with an extended dataset of 7,426 event picks detected using manual methods, recorded nationwide in February 2025. The model was adapted by freezing the final layers and retraining the previous ones. Evaluation in the Tilarán region demonstrates improved detection of local seismic phases, leading to more reliable catalogs and underscoring the potential of transfer learning for regional earthquake monitoring.