📄 earth science

VLPNet: A unified deep learning model for progressive single-station magnitude estimation from variable-duration P-waves in earthquake early warning

The paper introduces VLPNet, a unified deep learning model that enables progressive single-station magnitude estimation from variable-duration P-wave observations (0.5–6.0 s) without retraining, significantly improving accuracy and balancing warning timeliness with magnitude precision for earthquake early warning systems.

Original authors: Chunde Lu, Aihemaitijiang Tuerhong, Xiaopeng Gu, Yongtao Ma, Xiaohong Long

Published 2026-10-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Chunde Lu, Aihemaitijiang Tuerhong, Xiaopeng Gu, Yongtao Ma, Xiaohong Long

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

When the earth begins to shake, the first signal to arrive is not the violent, rolling motion that causes damage, but a faster, quieter pulse known as a P-wave. These waves travel through the planet's interior at high speeds, reaching seismic sensors seconds before the destructive shaking arrives. In the world of earthquake early warning, those few seconds are a precious commodity. They are enough time for a train to slow down, a gas line to shut off, or a person to drop to the ground. The critical challenge for scientists is to determine the size of the earthquake—the magnitude—using only this tiny, initial sliver of data. The longer they wait to gather more information, the more accurate the estimate becomes, but the less time remains to issue a warning. For decades, researchers have tried to find a balance, using mathematical formulas to guess the final size of a quake based on the first few seconds of its P-wave signal, but these methods often struggle to adapt as the signal grows and changes.

A team of researchers at Huazhong University of Science and Technology has developed a new approach to solve this problem, creating a system called VLPNet. Instead of relying on fixed formulas that look for specific, pre-defined numbers in the wave, this system uses a type of artificial intelligence designed to learn directly from the raw shape of the seismic signal. The researchers trained this digital model on a vast collection of real earthquake records from Japan, teaching it to look at the very first moment a P-wave hits a sensor and then watch how the signal evolves over time. The goal was to build a single, unified tool that could accept a P-wave signal of any length, from half a second up to six seconds, and instantly refine its estimate of the earthquake's magnitude as more data arrived. Unlike previous methods that might require a different calculator for a one-second window versus a three-second window, this new model handles all durations with the same brain, updating its prediction continuously as the wave unfolds.

To train this system, the researchers gathered over 13,000 records from 550 different seismic stations across Japan, covering earthquakes of various sizes and distances. They taught the model to look at the vertical shaking recorded deep underground, which arrives first, and predict the maximum horizontal shaking that would eventually occur at the surface. This is a crucial step because the surface shaking is what determines the damage and the magnitude. The model learned to connect the dots between the early, quiet P-wave and the eventual, violent ground motion without being told to look for specific peaks or frequencies. It simply learned the patterns. Once trained, the system was tested on new earthquakes it had never seen before. The results showed that as the model was allowed to listen to longer segments of the P-wave, its predictions became significantly more accurate. In the first half-second, the estimates had a certain level of uncertainty, but by the three-second mark, the error dropped substantially, and by six seconds, the predictions were highly reliable.

The study compared this new learning-based system against traditional methods that rely on manually picking out specific features, such as the highest point of the wave or its dominant rhythm. The new model consistently outperformed these older techniques. At the three-second mark, a point often used for issuing warnings, the new system reduced the average error by nearly half compared to the best traditional method. It also proved to be robust across different types of ground conditions. Whether the sensor was sitting on hard rock, soft soil, or something in between, the model maintained its accuracy, correctly estimating the magnitude within one unit for more than 93% of the cases. This consistency is vital because real-world sensors are placed on all kinds of terrain, and a warning system cannot afford to fail just because the ground is soft or hard.

The researchers found that the most dramatic improvement in accuracy happened during the first three seconds of the signal. After that, the gains continued but became more gradual. This suggests a practical trade-off for emergency managers: waiting longer yields a more precise number, but the warning must be issued early enough to be useful. The new model allows for a continuous update, meaning a warning can be issued early with a slightly wider margin of error and then refined instantly as more data comes in, rather than waiting for a fixed window to close. The study concludes that this unified approach offers a powerful way to handle the variable nature of earthquake signals. While the system is not a perfect crystal ball and still relies on the quality of the initial data, it represents a significant step forward in turning the fleeting moments of a P-wave into a reliable, life-saving alert.

Technical Summary: VLPNet for Progressive Single-Station Magnitude Estimation

Problem Statement
Earthquake Early Warning (EEW) systems rely on rapid magnitude estimation from initial P-wave observations to issue alerts before damaging S-waves arrive. A critical challenge in EEW is that P-wave observations are continuously accumulating; the available warning time and the duration of the observed waveform evolve in real-time. Traditional single-station magnitude estimation methods typically rely on empirical relationships using manually defined scalar features (e.g., peak displacement $P_d$ or maximum predominant period $\tau_p^{max}$ ) extracted from fixed-duration windows. These approaches face two main limitations:

Inflexibility: Models trained for specific input durations cannot be directly applied to arbitrary, evolving observation windows without retraining or zero-padding, which may introduce artifacts.
Information Loss: Reducing complex waveforms to a few scalar features discards rich amplitude, frequency, and temporal-evolution information contained in the raw signal.
Furthermore, existing deep learning models often struggle to maintain consistent performance across varying site conditions (e.g., hard rock vs. soft soil) and variable observation durations using a single unified framework.

Methodology
The authors propose VLPNet, a unified deep learning model designed to process variable-duration P-wave records without requiring separate models for different warning times.

Dataset: The study utilizes 13,245 station-event records from 1,916 earthquakes (JMA magnitude > 4.0, hypocentral distance 30–200 km) recorded at 550 KiK-net stations in Japan. The data was split chronologically into training (1997–2011), validation (2011–2015), and test (2015–2019) sets to prevent information leakage.
Input and Target:
- Input: The borehole vertical acceleration component (UD2) starting from an automatically identified P-wave arrival. The input duration is variable, ranging from 0.5 s to 6.0 s.
- Target: Instead of predicting magnitude directly, VLPNet predicts the logarithm of the complete-record peak horizontal displacement ( $\log_{10}A$ ). This value is later combined with hypocentral distance using an empirical JMA magnitude relation to estimate the final magnitude. The target is derived from the full surface horizontal acceleration records (NS1, EW1).
Network Architecture:
- Feature Extraction: A Convolutional Neural Network (CNN) module with two 1D convolutional blocks extracts local features from the raw acceleration sequence. No temporal pooling is applied to preserve temporal resolution.
- Variable-Length Handling: To handle mini-batches of different lengths, the model employs a packed-sequence strategy (pack_padded_sequence). Input sequences are zero-padded to a fixed length (600 samples) for convolution, but the valid lengths are retained. Before entering the recurrent layer, padded steps are excluded from the computation, ensuring that only valid data influences the temporal modeling.
- Temporal Modeling: A two-layer Long Short-Term Memory (LSTM) network with 64 hidden units captures temporal dependencies. The final hidden state corresponds to the actual end of the observation window, not the padded length.
- Regression: A fully connected layer maps the LSTM output to the predicted $\log_{10}A$ .
Training: The model was trained using the Adam optimizer to minimize Mean Squared Error (MSE) between predicted and target $\log_{10}A$ . During training, records were randomly truncated to simulate variable observation durations (0.5–6.0 s).

Key Contributions

Unified Framework: Development of a single model capable of processing P-wave records of any duration between 0.5 s and 6.0 s, enabling continuous magnitude updates as new data arrives without retraining.
Variable-Length Processing: Implementation of a packed-sequence strategy that allows LSTM layers to process sequences of varying effective lengths within a single mini-batch, avoiding the information loss associated with standard zero-padding in recurrent networks.
Systematic Evaluation: Comprehensive comparison against conventional methods ( $P_d$ , $\tau_p^{max}$ , and their combination) and rigorous assessment of model robustness across three distinct site classes (hard rock, intermediate, soft soil).

Results

Progressive Accuracy: As the observation duration increased from 0.5 s to 6.0 s, the Mean Absolute Error (MAE) for magnitude estimation decreased from $0.461 \pm 0.002$ to $0.317 \pm 0.010$ . The proportion of estimates within $\pm 1$ magnitude unit increased from 89.2% to 96.5%.
Performance Gain: At a 3.0 s observation window, VLPNet reduced the MAE by 28.0% to 50.9% compared to conventional methods ( $P_d$ , $\tau_p^{max}$ , and $P_d + \tau_p^{max}$ ). Specifically, VLPNet achieved an MAE of 0.345, compared to 0.479 for the best-performing conventional combination.
Site Robustness: The model maintained stable performance across hard-rock, intermediate, and soft-soil sites. At 3.0 s, more than 93% of estimates for all three site classes were within $\pm 1$ magnitude unit. While hard-rock sites showed slightly higher MAE and a tendency toward underestimation, the overall dispersion of errors remained consistent across site types.
Error Distribution: The model exhibited a systematic underestimation for larger magnitude events, particularly at short observation durations (0.5 s), which improved as more waveform data became available.

Significance and Claims
The paper claims that VLPNet successfully addresses the trade-off between warning timeliness and magnitude accuracy by utilizing a unified deep learning architecture. By directly learning from raw, variable-length waveforms rather than relying on predefined scalar features, the model captures richer amplitude and temporal-evolution information. The study demonstrates that a single model can effectively quantify magnitude estimates progressively as P-wave information accumulates, offering a more flexible and accurate alternative to traditional empirical methods for single-station earthquake early warning. The authors note that while the results are promising on the KiK-net dataset, further evaluation is needed regarding transferability to other seismic regions and robustness to real-time uncertainties in P-wave picking and distance estimation.

Technical Summary: VLPNet for Progressive Single-Station Magnitude Estimation

More like this