[Paper Review] Deep learning for magnitude prediction in earthquake early warning

Yanwei Wang, Zifa Wang+2 more|arXiv (Cornell University)|Dec 2, 2019

Seismology and Earthquake Studies4 citations

TL;DR

This paper proposes EEWNet, a deep learning model that predicts earthquake magnitude using raw vertical-component accelerograms without feature engineering. By adapting its architecture to input length and training on 10,000+ records, EEWNet achieves superior magnitude prediction accuracy compared to traditional parameter-based methods like τlog, τc, and Pd.

ABSTRACT

Fast and accurate magnitude prediction is the key to the success of earthquake early warning. We have proposed a new approach based on deep learning for P-wave magnitude prediction (EEWNet), which takes time series data as input instead of feature parameters. The architecture of EEWNet is adaptively adjusted according to the length of the input, thus eliminates the need of complicated tuning of hyperparameters for deep learning. Only the unfiltered accelerograms of vertical components are used. EEWNet is trained on a moderate number of data set (10,000s of records), but it achieves excellent results in magnitude prediction compared with approaches using parameters τlog, τc and Pd.

Motivation & Objective

• To improve the accuracy and speed of earthquake magnitude prediction in early warning systems.
• To eliminate reliance on handcrafted features such as τlog, τc, and Pd in magnitude estimation.
• To develop a deep learning model that dynamically adapts to varying input lengths without hyperparameter tuning.
• To train a robust model using only unfiltered vertical-component accelerograms from a moderate-sized dataset.
• To enable faster and more reliable early warning by leveraging end-to-end learning from raw seismic time series.

Proposed method

• EEWNet processes raw, unfiltered vertical-component accelerogram time series as input, bypassing feature extraction.
• The model employs an architecture that adaptively adjusts based on input length, reducing the need for manual hyperparameter tuning.
• It uses a deep neural network with convolutional and fully connected layers to learn temporal patterns directly from seismic signals.
• The network is trained end-to-end on a dataset of approximately 10,000 earthquake records to predict magnitude.
• No preprocessing or feature engineering (e.g., no τlog, τc, or Pd) is applied; only raw accelerograms are used.
• The model is optimized for fast inference, suitable for real-time earthquake early warning applications.

Experimental results

Research questions

• RQ1 Can a deep learning model trained directly on raw seismic time series outperform traditional magnitude prediction methods?
• RQ2 To what extent can adaptive architecture design reduce the need for hyperparameter tuning in earthquake magnitude prediction?
• RQ3 How does the performance of EEWNet compare to established methods like τlog, τc, and Pd when using only vertical-component accelerograms?
• RQ4 Can a model trained on a moderate-sized dataset (10,000s of records) achieve high accuracy in real-time magnitude estimation?
• RQ5 Does end-to-end learning from raw data improve prediction speed and accuracy in earthquake early warning systems?

Key findings

• EEWNet achieves superior magnitude prediction accuracy compared to traditional methods such as τlog, τc, and Pd.
• The model eliminates the need for feature engineering by directly processing raw accelerograms without filtering or parameter extraction.
• Adaptive architecture design allows EEWNet to handle variable-length inputs without requiring manual hyperparameter adjustments.
• Despite being trained on a moderate dataset of 10,000+ records, EEWNet delivers excellent performance in magnitude prediction.
• The approach enables faster and more reliable early warning by leveraging deep learning on raw seismic signals.
• The results demonstrate the feasibility and effectiveness of end-to-end deep learning for real-time earthquake magnitude estimation.