第98回統計地震学セミナー / The 98th Statistical Seismology Seminar

【Date＆Time】: 1 October 2024 14:30-17:00
Admission Free, No Booking Necessary
【Place】: online (with zoom) and seminar room D312B (ISM); Zoom meeting:
https://us06web.zoom.us/j/86800402605?pwd=hafOAYTWQyxjkuZoHUA0jQhoHDhcjd.1
Meeting ID: 868 0040 2605
Passcode: 498844
【Speaker1】: Stephen Wu (ISM)
【Title】: Potential of LLMs in seismology
【Abstract】: The application of Large Language Models (LLMs) in seismology holds significant promise, but direct usage of LLMs as standalone tools often falls short in addressing the complexity of seismic data and tasks. In this presentation, I will explore LLM orchestration as a system-building challenge, demonstrating how integrating multiple LLMs within a structured framework enhances the potential for accurate and efficient problem-solving in seismology. This approach allows for specialized task management and better adaptation to domain-specific needs. To further refine these ideas, we organized a local hackathon in Japan, inviting seismologists to provide feedback and insights on the utility of orchestrated LLM systems for seismic analysis. The outcomes of this collaborative effort will help shape future applications of LLMs in the field.
【Speaker2】: Chengxiang Zhan
【Title】: Neural Point Process: A Modulated Renewal Model for Temporal Event Modeling
【Abstract】: Recent research has demonstrated that Neural Point Processes (NPPs) are effective in modeling time series of discrete events. However, the underlying rationale for this approach remains insufficiently explored. Two key questions arise: First, how can NPPs successfully model event sequences using only a limited amount of historical data? Second, does the amount of required historical data vary depending on the type of point process being modeled? Here, we aim to address both questions. We propose that NPPs can be interpreted as modulated renewal models. Specifically, one neural network learns the renewal process using the current waiting time as input, while another neural network captures the modulation effect based on intervals between past events. Our findings further reveal that the amount of historical data needed for accurate modeling varies across different types of point processes.
【Speaker3】: Jiancang Zhuang (ISM)
【Title】: Quantification of earthquake predictability
【Abstract】: In earthquake forecasting, there is a significant gap between complete randomness and complete deterministicity. This presentation begins by discussing how to quantify predictability and outlining the current state of earthquake predictability from an information-theoretic perspective.
【Speaker4】: Dr. Timothy Clements
(Postdoctoral fellow at the U.S. Geological Survey’s Earthquake Science Center in Moffett Field, CA.)
【Title】: A Ground Motion-based Approach to Earthquake Early Warning and Earthquake Statistics
【Abstract】: Real-time earthquake early warning and forecasting systems have focused on using an earthquake's location, origin time, and magnitude to forecast either ground motion or the rate of earthquakes. Here, we suggest using continuous seismic ground motion to forecast ground motion in the near (seconds) to intermediate (days) future. I will introduce the GRAph Prediction of Earthquake Shaking (GRAPES) algorithm (Clements et al. 2024, GRL), a deep learning model trained to characterize and propagate earthquake shaking across a seismic network, for use in earthquake early warning (EEW). I will show that, when applied to earthquake seismology, deep learning is not a black box; GRAPES’ internal activations, which I call “seismic vectors”, correspond to the arrival of distinct seismic phases. While trained on earthquakes recorded in Japan, I will show that GRAPES, without modification, outperforms the USGS ShakeAlert earthquake early warning system on the 2019 M7.1 Ridgecrest, CA earthquake. I will then apply earthquake forecasting techniques to extract the Gutenberg-Richter b-value and Omori-Utsu parameters from the distribution of recent ground motion recorded at a single nearby seismometer during an aftershock sequence using a maximum likelihood approach. We apply our ground motion-based approach to the two weeks following the 2019 M7.1 Ridgecrest, CA earthquake sequence.