[S-SS06] New trends in data acquisition, analysis and interpretation of seismicity

[SSS06-P05] Integrating Advanced Phase Pickers and Graph Neural Networks for Seismic Phase Association in Türkiye

*Nurcan Meral Ozel¹、Cagri Diner¹、Erdem Ata²、Fatih Turhan¹、Yavuz Gunes¹、Dogan Aksari¹、Mehmet Yilmazer¹、Mehmet Efe Akca³、Alperen Sahin⁴、Batuhan Kalem³ (1.Bogazici University, Kandilli Observatory and Earthquake Research Institute、2.TensorBundle,Bogazici University Kandilli Teknopark、3.Bogazici University, Department of Mathematics、4.Bogazici University, Department of Physics)

キーワード：Deep learning, Sismology, Real-time Monitoring, GENIE

Accurate phase association remains a significant challenge in real-time seismic monitoring, particularly in seismically active and complex regions like Türkiye. Traditional association methods often struggle with the large volume of picks generated by modern deep-learning-based phase pickers and the ambiguities arising from overlapping arrivals. To address this, we integrate multiple state-of-the-art phase picking algorithms, including PhaseNet, EQTransformer, EQCCT, and PhaseNO, into KOERI's seismic monitoring system. While these pickers enhance sensitivity and detection capabilities, they considerably increase the number of picks, complicating their association with the correct seismic event. To overcome this challenge, we introduce GENIE (Graph Earthquake Neural Interpretation Engine), a graph-based neural network designed for phase association and spatiotemporal event localization. GENIE leverages the relational structure of seismic phases by learning adjacency relationships between picks, ensuring globally consistent event classification. The model effectively resolves conflicts caused by overlapping sources and enhances association accuracy by incorporating; spatial coherence among stations detecting the same seismic event and temporal resolution of multiple sources recorded at the same station.Unlike conventional approaches, GENIE is tailored to handle challenging scenarios, including high-seismicity environments, near-simultaneous arrivals, and noisy or incomplete data. Initial results demonstrate that combining deep-learning-based phase pickers with a graph neural network significantly improves phase association accuracy and event detection in Türkiye’s dense seismic networks.