DECA: A Novel Deep Learning Framework for Seismic Phase Association Using Deep Embedded Clustering

Seismological monitoring systems face persistent challenges in accurately associating seismic phases, a critical step for robust automated event detection. Traditional techniques, predominantly reliant on arrival times, often struggle with misassociations, especially in tectonically complex or densely instrumented regions. These limitations highlight the urgent need for innovative solutions that leverage recent advances in artificial intelligence to enhance both detection and association performance. In this study, we explore Deep Embedded Clustering, an unsupervised learning technique that simultaneously performs feature extraction and clustering directly from raw waveform data. This framework addresses long-standing challenges in seismology related to signal complexity and scalability in large datasets. The model exhibited strong performance, achieving a Silhouette score above 0.9 and a 96.55% match rate with previously cataloged events. At its core, the system employs a convolutional autoencoder to derive latent representations from multi-channel waveform inputs, capturing critical seismic features. Training combines reconstruction loss with clustering refinement loss, resulting in more coherent and meaningful representations. Our results underscore the model’s ability to detect subtle waveform variations without labeled data. Future work will extend validation to larger, noisier datasets and diverse tectonic settings, while exploring advanced deep learning models to further solidify AI’s role in next-generation seismic phase association.