Year: 2026
Type: article
Abstract: Artificial Intelligence and machine learning have transformed the field of seismology over the past decade. Fueled by advances in high-performance computing, the availability of large seismic and labeled datasets, and breakthroughs in deep neural network architectures, the seismological community has developed many machine learning models for seismic phase detection and association, enabling the construction of high-resolution earthquake catalogs around the globe. However, this rapid methodological growth has also encountered several potential issues: redundant model development, mundane application of default tools, neglect of data quality and preprocessing, and an overemphasis on methods at the expense of scientific discovery. This article argues that we must pivot from method-centric to science-centric research. Four future directions are identified as particularly promising: (1) mining high-resolution catalogs through interpretable point-process modeling, change-point detection, and generative simulation for new physical insights into earthquake nucleation, fault interaction, and aftershock forecasting; (2) extending machine-learning detection capabilities to tectonic tremors, low-frequency earthquakes (LFEs) and non-earthquake signals, which remain poorly characterized globally; (3) applying machine-learning techniques to detect events in challenging periods (e.g., overlapping arrivals); and (4) developing large seismic foundation models that can be fine-tuned across a wide range of downstream tasks. Ultimately, unique datasets, innovative techniques, and outstanding scientific questions must remain in the three pillars of research innovations in source seismology.
Source: Big Data and Earth System
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid:
OpenAlex ID: W7225317092
