📄 earth_science

A lightweight automatic workflow for expanding volcano-tectonic seismic catalogs: application to Villarrica Volcano, Chile

This paper presents a lightweight, fully automatic workflow that significantly expands volcano-tectonic seismic catalogs by detecting and locating previously missed events, as demonstrated by its successful application to Villarrica Volcano, Chile, which nearly doubled the existing record while maintaining accuracy comparable to expert manual analysis.

Original authors: Sergio Morales, Sarah Jaye Oliva, Andrés I. Ávila

Published 2026-08-05

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Sergio Morales, Sarah Jaye Oliva, Andrés I. Ávila

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

Imagine the Earth is like a giant, grumpy house that occasionally creaks, groans, and shifts its foundation. Sometimes, these creaks are just the house settling, but other times, they are warning signs that something big is about to happen, like a volcano waking up. Scientists who study these "house noises" are called seismologists, and they listen for specific types of rumbles called "volcano-tectonic" earthquakes. Think of these as the sharp, snapping sounds of rocks breaking deep inside the volcano's walls as magma (super-hot melted rock) pushes its way up. Catching these snaps is crucial because they tell us if the volcano is getting stressed and might erupt soon.

However, listening to a volcano is like trying to hear a single pin drop in a crowded, noisy stadium. The signals are often tiny, and the "noise" from wind or distant traffic can drown them out. Traditionally, scientists have relied on teams of human experts to sit in front of computer screens, listening to hours of audio recordings to find these tiny snaps. But humans get tired, they can't listen to everything at once, and during busy times when the volcano is really chattering, they inevitably miss the quietest, smallest snaps. This paper asks a simple question: Can we build a robot that listens better than a tired human, without needing a supercomputer the size of a city?

The researchers behind this study decided to build exactly that: a lightweight, automatic workflow that acts like a tireless, super-attentive detective for the Villarrica Volcano in Chile. They didn't just build a robot to listen; they built a whole team of digital specialists working together. First, they used a "deep learning" model (a type of AI trained on thousands of earthquake sounds) to act as the ears, scanning the data to spot the faintest P-waves and S-waves (the first and second rumbles of an earthquake). Next, they used a smart clustering algorithm called MeanShift to act as the brain, grouping those scattered rumbles together to figure out, "Hey, these three sounds happened at almost the same time; they must be from the same event!" Finally, they used a probabilistic location tool to act as the mapmaker, calculating exactly where underground that snap happened.

The results of this digital detective work are impressive. When they tested their system on two years of continuous data from Villarrica Volcano (from January 2023 to December 2024), they compared their robot's list against the official list created by human experts. The robot managed to find 72.5% of the events the humans had already found, proving it was just as good at catching the big, obvious ones. But the real magic happened with the ones the humans missed. The robot discovered 695 additional earthquakes that the human catalog didn't have. That's a 91.4% expansion of the known seismic activity!

These "extra" earthquakes weren't random noise; they were mostly small, shallow events right near the volcano's surface, with magnitudes between 1 and 2. These are the tiny, tricky snaps that humans often miss when they are overwhelmed by a high volume of activity. The robot didn't just find more; it found the ones that were hiding in plain sight. The timing of these events was also spot-on, with the robot's clock being only about 0.44 seconds off from the human experts' estimates—a difference so small it doesn't change the story of what the volcano is doing.

The paper makes it clear that this isn't a magic wand that solves every problem. The system is specifically tuned for "volcano-tectonic" earthquakes (the sharp rock-breaking kind) and might miss other types of volcanic sounds that don't have a clear second "S-wave" rumble. It also relies on the same maps of underground rock speeds that the humans use, so if those maps are imperfect, the robot's location guesses will be imperfect too. However, the study proves that this lightweight, automatic system can run on standard computer servers found at most observatories without needing expensive graphics cards or constant human babysitting. It suggests that by letting a robot handle the heavy lifting of listening, volcanologists can expand their catalogs significantly, catching the quiet whispers of the volcano that humans might otherwise miss, all while running on the same hardware they already have.

Technical Summary: A Lightweight Automatic Workflow for Expanding Volcano-Tectonic Seismic Catalogs

Problem Statement
Volcano-tectonic (VT) seismicity is a primary indicator of stress changes and fluid migration within a volcanic edifice, making accurate cataloging essential for hazard assessment. However, operational monitoring at volcanoes like Villarrica (Chile) faces two compounding limitations: the sparsity of monitoring networks combined with high background noise, and the labor-intensive nature of manual analysis. Manual workflows often fail to detect the smallest VT events (typically $M_L$ 1–2) during periods of high seismic activity due to overlapping waveforms and analyst time constraints. This results in incomplete catalogs that may mask precursory sequences and bias hazard assessments. While deep learning (DL) has shown promise in seismic detection, existing models often require site-specific training, lack transferability across different geological settings, and have not been rigorously validated against independent expert catalogs in continuous operational environments.

Methodology
The authors propose a fully automatic, modular, and lightweight workflow designed to run on standard observatory infrastructure without manual intervention. The system integrates three open-source components:

Phase Detection: The workflow utilizes PhaseNet, a deep learning model trained to identify P- and S-wave arrivals. To maximize the recovery of low-magnitude events, the detection threshold was set to 0.4 (below the default), with stricter quality control delegated to subsequent stages.
Phase Association: Instead of complex network-level rules, the system employs MeanShift clustering (a time-based clustering algorithm) to group P-phase arrivals from proximal stations into candidate events. The clustering bandwidth is a single, physically interpretable parameter derived from the network aperture and a representative P-wave velocity. This approach allows the system to adapt to any network geometry without retraining the model.
Probabilistic Location: Event location is performed using NonLinLoc, a probabilistic, non-linear earthquake location algorithm. The workflow follows a multi-stage procedure:
- Initial Location: Computed using P-phases from the proximal network.
- S-Phase Search: A targeted search for S-phases at proximal stations based on the initial hypocenter.
- Distal Integration: P- and S-phases from distal stations are incorporated only if they satisfy amplitude consistency criteria (ensuring they belong to the same event rather than a closer source).
- Iterative Refinement: Solutions are accepted only if they reduce the mean location uncertainty.

The workflow distinguishes between a "proximal" network (stations within 15 km of the summit) and a "distal" network to balance phase detectability with depth resolution. The entire pipeline was applied to two years of continuous data (January 2023–December 2024) from Villarrica Volcano, using a 16-station network.

Key Results
The automatic workflow was validated against the expert-reviewed catalog from the Observatorio Volcanológico De los Andes del Sur (OVDAS), restricted to events with location errors less than 5 km.

Catalog Expansion: The workflow recovered 72.5% (551 events) of the 760 well-located manual events. Crucially, it identified 695 additional events not present in the manual catalog, representing a 91.4% expansion of the well-located VT record.
Event Characteristics: The newly detected events are concentrated near the volcanic edifice at shallow depths and within a narrow magnitude range of $M_L$ 1–2. This population is typically underrepresented in routine manual processing.
Origin Time Accuracy: The automatic origin times showed a small, consistent negative bias of −0.44 s relative to manual picks. This is attributed to PhaseNet placing arrivals slightly after the analyst-picked onset, a trend that does not affect the interpretation of seismicity rates or spatial patterns.
Location Quality: For matched events, epicentral locations agreed closely between the two methods. Depth estimates showed greater scatter for shallow events (<5 km), consistent with known limitations of depth resolution in near-surface velocity models.
Computational Efficiency: The workflow processed two years of data on a standard virtual machine (4 CPU cores, 16 GB RAM, no GPU) with a throughput of approximately one hour of computation per day of continuous data.

Significance and Claims
The paper positions this workflow as a practical, scalable alternative for volcanic observatories seeking to expand their seismic catalogs without proportional increases in analyst workload or computational infrastructure.

Operational Viability: The system is designed to run on standard hardware with no manual intervention, making it suitable for near-real-time deployment.
Transferability: The workflow requires only a small set of physically interpretable parameters (e.g., station lists, clustering bandwidth, quality thresholds) to adapt to different volcanic networks, avoiding the need for site-specific model retraining.
Targeted Recovery: The system specifically targets the "missing" population of small VT events ( $M_L$ 1–2) that are often missed during high-activity periods, thereby improving the resolution of stress changes in the shallow crust.
Limitations and Scope: The authors explicitly note that the workflow is implicitly biased toward brittle-failure sources with impulsive S-wave onsets (VT events) and under-represents long-period or hybrid events. Furthermore, the magnitude of completeness was not formally quantified, and the 91.4% expansion is presented as a lower bound on the improvement in detection sensitivity. The study also acknowledges that the 1D velocity model used is a shared limitation with manual processing, particularly for distal stations.

In conclusion, the study demonstrates that a lightweight, open-source pipeline can significantly enhance the detection and location of volcano-tectonic seismicity, providing a robust tool for observatories to monitor volcanic unrest more comprehensively.

Technical Summary: A Lightweight Automatic Workflow for Expanding Volcano-Tectonic Seismic Catalogs

More like this