Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Time Series Foundation Models and Deep Learning Architectures for Earthquake Temporal and Spatial Nowcasting

Alireza Jafari, Geoffrey Fox, John B. Rundle, Andrea Donnellan, Lisa Grant Ludwig

TL;DR

The paper addresses real-time earthquake nowcasting by evaluating a broad set of deep learning architectures, including pre-trained foundation models, for a 14-day horizon across 0.1-degree spatial bins in Southern California from 1986 to 2024. It introduces two innovations, MultiFoundationQuake and GNNCoder, and demonstrates that combining diverse foundation-model outputs with spatial graph learning yields superior predictive performance relative to single-domain or purely temporal models. The results show that pre-training data strongly influence foundation-model performance, while pattern models and GNNs effectively capture temporal-spatial dependencies; feature engineering further boosts accuracy. The work has practical implications for disaster risk reduction by improving nowcasting accuracy and highlighting pathways for future hybrid models and graph-construction improvements.

Abstract

Advancing the capabilities of earthquake nowcasting, the real-time forecasting of seismic activities remains a crucial and enduring objective aimed at reducing casualties. This multifaceted challenge has recently gained attention within the deep learning domain, facilitated by the availability of extensive, long-term earthquake datasets. Despite significant advancements, existing literature on earthquake nowcasting lacks comprehensive evaluations of pre-trained foundation models and modern deep learning architectures. These architectures, such as transformers or graph neural networks, uniquely focus on different aspects of data, including spatial relationships, temporal patterns, and multi-scale dependencies. This paper addresses the mentioned gap by analyzing different architectures and introducing two innovation approaches called MultiFoundationQuake and GNNCoder. We formulate earthquake nowcasting as a time series forecasting problem for the next 14 days within 0.1-degree spatial bins in Southern California, spanning from 1986 to 2024. Earthquake time series is forecasted as a function of logarithm energy released by quakes. Our comprehensive evaluation employs several key performance metrics, notably Nash-Sutcliffe Efficiency and Mean Squared Error, over time in each spatial region. The results demonstrate that our introduced models outperform other custom architectures by effectively capturing temporal-spatial relationships inherent in seismic data. The performance of existing foundation models varies significantly based on the pre-training datasets, emphasizing the need for careful dataset selection. However, we introduce a new general approach termed MultiFoundationPattern that combines a bespoke pattern with foundation model results handled as auxiliary streams. In the earthquake case, the resultant MultiFoundationQuake model achieves the best overall performance.

Time Series Foundation Models and Deep Learning Architectures for Earthquake Temporal and Spatial Nowcasting

TL;DR

The paper addresses real-time earthquake nowcasting by evaluating a broad set of deep learning architectures, including pre-trained foundation models, for a 14-day horizon across 0.1-degree spatial bins in Southern California from 1986 to 2024. It introduces two innovations, MultiFoundationQuake and GNNCoder, and demonstrates that combining diverse foundation-model outputs with spatial graph learning yields superior predictive performance relative to single-domain or purely temporal models. The results show that pre-training data strongly influence foundation-model performance, while pattern models and GNNs effectively capture temporal-spatial dependencies; feature engineering further boosts accuracy. The work has practical implications for disaster risk reduction by improving nowcasting accuracy and highlighting pathways for future hybrid models and graph-construction improvements.

Abstract

Advancing the capabilities of earthquake nowcasting, the real-time forecasting of seismic activities remains a crucial and enduring objective aimed at reducing casualties. This multifaceted challenge has recently gained attention within the deep learning domain, facilitated by the availability of extensive, long-term earthquake datasets. Despite significant advancements, existing literature on earthquake nowcasting lacks comprehensive evaluations of pre-trained foundation models and modern deep learning architectures. These architectures, such as transformers or graph neural networks, uniquely focus on different aspects of data, including spatial relationships, temporal patterns, and multi-scale dependencies. This paper addresses the mentioned gap by analyzing different architectures and introducing two innovation approaches called MultiFoundationQuake and GNNCoder. We formulate earthquake nowcasting as a time series forecasting problem for the next 14 days within 0.1-degree spatial bins in Southern California, spanning from 1986 to 2024. Earthquake time series is forecasted as a function of logarithm energy released by quakes. Our comprehensive evaluation employs several key performance metrics, notably Nash-Sutcliffe Efficiency and Mean Squared Error, over time in each spatial region. The results demonstrate that our introduced models outperform other custom architectures by effectively capturing temporal-spatial relationships inherent in seismic data. The performance of existing foundation models varies significantly based on the pre-training datasets, emphasizing the need for careful dataset selection. However, we introduce a new general approach termed MultiFoundationPattern that combines a bespoke pattern with foundation model results handled as auxiliary streams. In the earthquake case, the resultant MultiFoundationQuake model achieves the best overall performance.
Paper Structure (21 sections, 3 equations, 7 figures, 2 tables)

This paper contains 21 sections, 3 equations, 7 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Data
  3. Graph structure for GNNCoder
  4. Pre-training Datasets for Transformer Models
  5. TrafficL Dataset
  6. Weather Dataset
  7. M4 Dataset
  8. Models Description
  9. Pre-trained Transformer Models
  10. Graph Neural Networks models
  11. Memory-Based Models
  12. Convolutional and MLP-Based Models
  13. MultiFoundationQuake
  14. Model Training and Implementation
  15. Models Evaluation and Comparison
  16. ...and 6 more sections

Figures (7)

  • Figure 1: Illustration of the construction of a nowcast model for California. The nowcast is a 2-parameter filter on the small earthquake seismicity Rundle2024QuakeGPT2022EA002343. a) Seismicity in the Los Angeles region since 1960, M>3.29. b) Monthly rate of small earthquakes as cyan vertical bars. The blue curve is the 36-month exponential moving average (EMA). c) Mean rate of small earthquakes since 1970. d) Nowcast curve that is the result of applying the optimized EMA and corrections for time varying small earthquake rate to the small earthquake seismicity. e) Optimized Receiver Operating Characteristic (ROC) curve (red line) used in the machine learning algorithm. Skill is the area under the ROC curve and is used in the optimization. Skill tradeoff diagram shows the range of models used in the optimization.
  • Figure 2: Distribution of earthquake epicenters in Southern California (32°N to 36°N, -120° to -114°) from USGS data (1986-2024). The scatter plot shows the spatial density of seismic events used to analyze and optimize spatial bins for earthquake nowcasting. There is no magnitude cut, with data including all USGS recorded seismic events starting from magnitude 0.
  • Figure 3: The 500 most active and vulnerable spatial bins, marked in blue, selected for analysis out of the total 2400, based on the frequency of earthquakes from 1986 to 2024. This selection focuses on high-risk areas.
  • Figure 4: Six time series from randomly selected spatial bins, highlighting earthquakes of magnitude greater than 5.
  • Figure 5: The final graph structure representing the 500 most active bins, created using an epsilon of 0.15 degrees. Initially forming a multi-component graph, components are linked to ensure full connectivity.
  • ...and 2 more figures