AI, Cloud Computing, and Software Engineering Concepts in Seismology

The current state of earthquake monitoring systems utilize standard algorithms that process seismic data in realtime, and run software on local servers. Earthquake products such as focal mechanisms, moment tensors, ShakeMaps and waveforms are generated. While this system has proven effective and robust for twenty-plus years, it is a dated system in need of modernization. The Southern California Seismic Network is exploring ways to generate and distribute earthquake products with modern software technologies such as cloud-native services in Amazon Web Services (AWS), serverless computing, containers, and infrastructure as code. AI-powered models are implemented within this monitoring system, as a substitute for standard algorithms.

The Southern California Seismic Network has developed a system known as Quakes2AWS. It is a versatile system that uses AWS cloud components and runs machine learning models on both real-time and historical time-series data. In addition, the network has developed a terrestrial, API-based post-processing system to run AI models on historical time-series data. A discussion on both system architectures will touch upon topics including streaming data, latency, Docker containers, databases, and CI/CD. 

Also discussed are the artificial intelligence models used in seismology, for phase picking and for earthquake event association. Many of these models are open-source. The goal is to have these AI models augment humans in analyzing earthquakes, such as finding earthquakes that timers or standard algorithms might miss. From an archival and academic perspective, an alternate catalog can be generated that other researchers might to download.

The results of the monitoring system pipeline will be discussed, as it pertains to the AI models’ findings against the earthquake waveform data. The results have shown to be promising, particularly in utilizing the Phasenet neural network and GaMMa associator to spot smaller earthquake events that timers may have missed. The modular design of the pipeline also allows for swapping out models seamlessly. Finally, there will be a discussion on future directions and implications.