The transformer earthquake alerting model: A new versatile approach to earthquake early warning

Abstract: Earthquakes are major hazards to humans, buildings and infrastructure. Early warning methods aim to provide advance notice of incoming strong shaking to enable preventive action and mitigate seismic risk. Their usefulness depends on accuracy, the relation between true, missed and false alerts, and timeliness, the time between a warning and the arrival of strong shaking. Current approaches suffer from apparent aleatoric uncertainties due to simplified modelling or short warning times. Here we propose a novel early warning method, the deep-learning based transformer earthquake alerting model (TEAM), to mitigate these limitations. TEAM analyzes raw, strong motion waveforms of an arbitrary number of stations at arbitrary locations in real-time, making it easily adaptable to changing seismic networks and warning targets. We evaluate TEAM on two regions with high seismic hazard, Japan and Italy, that are complementary in their seismicity. On both datasets TEAM outperforms existing early warning methods considerably, offering accurate and timely warnings. Using domain adaptation, TEAM even provides reliable alerts for events larger than any in the training data, a property of highest importance as records from very large events are rare in many regions.

The Transformer Earthquake Alerting Model: A New Versatile Approach to Earthquake Early Warning

The paper presents the Transformer Earthquake Alerting Model (TEAM), a deep learning-based approach designed to enhance earthquake early warning (EEW) systems. This method addresses two fundamental requirements for effective EEWs: accuracy and timeliness. Traditional methodologies often experience limitations such as aleatoric uncertainties due to simplified models and short warning times. TEAM seeks to overcome these shortcomings by leveraging the capabilities of deep learning, specifically through the use of transformers, to provide real-time analysis of seismic data.

Overview of TEAM

TEAM processes raw strong motion waveforms from an arbitrary number of seismic stations in real-time, enhancing adaptability to changing seismic networks. The model consists of three core steps: event detection, Peak Ground Acceleration (PGA) estimation, and thresholding for issuing warnings. PGA estimation is particularly pivotal, utilizing a convolutional neural network for feature extraction, a transformer model for feature combination, and a mixture density network to predict PGA distributions. This end-to-end learning approach allows the model to dynamically adapt and predict complex shaking patterns, surpassing traditional EEW methods.

Performance Evaluation

TEAM was evaluated using datasets from seismically active regions in Japan and Italy. These datasets provided diverse testing scenarios due to their complementary seismic characteristics. Precision-recall analysis across various PGA thresholds demonstrated TEAM's superior performance over traditional EEW methods, such as those based on source estimation (e.g., estimated point source (EPS) methods) and propagation (e.g., PLUM-based methods). The improved capability of TEAM was particularly significant at high PGA thresholds, where traditional methods tend to underperform due to data scarcity of high-magnitude events in training sets.

Using domain adaptation, TEAM also shows enhanced capability in predicting rare large events, an important consideration for regions that historically experience significant seismic activity. The model exploits auxiliary datasets and fine-tuning to adapt to new regions effectively, mitigating the issue of limited local large-event data.

Implications for AI and Earthquake Early Warning

The implications of TEAM for both theoretical developments in AI and practical EEW applications are substantial. The application of transformers in real-time seismic data processing highlights the potential for deep learning models to transform how early warning systems are conceived and implemented, moving towards more adaptive and precise approaches. The model's ability to adapt to dynamically changing seismic inputs and provide probabilistic predictions with calibrated uncertainties reflects a significant stride towards integrating machine learning with geophysical applications.

Beyond practical insights, the study suggests several directions for future research. Developing enhanced domain adaptation strategies, potentially involving synthetic data, could tackle data scarcity issues related to exceedingly rare large events. Moreover, the study highlights the importance of ensemble models in improving the calibration of probabilistic predictions, a valuable line of inquiry for improving reliability in machine applications.

Conclusion

In conclusion, the development of TEAM represents a pertinent advancement in the field of earthquake early warning, providing a methodological foundation that integrates deep learning with seismic hazard assessment. With superior accuracy and adaptive capabilities, TEAM offers a promising improvement over traditional earthquake alerting models. Future research efforts in expanding training datasets and refining domain adaptation could further elevate the potential of machine learning in seismic risk mitigation. The use of deep learning in such critical applications underscores the evolving role of AI in enhancing societal resilience against natural disasters.