AI4Seismo: Seismic AI Open Benchmark Community

AI4Seismo is an open community for AI-based seismic monitoring, focusing on open datasets, models, benchmarks, and deployment-level evaluation workflows for seismological AI.

We aim to support reproducible research and practical deployment of AI methods for seismic monitoring, especially under realistic continuous-waveform conditions.

Mission

AI4Seismo is built around a simple idea:

Seismic AI should be evaluated not only on event-centered datasets, but also under realistic continuous-waveform monitoring scenarios.

Most existing seismic AI datasets are organized around earthquake events. This is useful for model training and controlled evaluation, but real seismic monitoring systems operate on continuous waveform streams.

In practical deployment, AI models must deal with:

• Long continuous waveform data
• Background noise
• False alarms
• Missed detections
• Phase association errors
• Incomplete station coverage
• Cross-region generalization
• Event-type ambiguity
• Computational efficiency
• End-to-end monitoring performance

AI4Seismo aims to promote open benchmarks, tools, and discussions for these deployment-level challenges.

Focus Areas

AI4Seismo focuses on the following topics:

• Seismic phase picking
• Earthquake detection
• Seismic event classification
• Continuous-waveform benchmark datasets
• Deployment-level model evaluation
• End-to-end automatic seismic monitoring workflows
• Open-source seismic AI models
• Reproducible evaluation protocols
• AI for Earth science
• Scientific machine learning for seismology

Why Continuous-Waveform Benchmarks?

Many seismic AI models are trained and evaluated on event-centered waveform windows. These datasets are valuable, but they do not fully represent real monitoring conditions.

In real deployment, models need to answer harder questions:

• Can the model detect earthquakes from continuous data streams?
• How many false alarms does the model produce?
• Can the model maintain performance across different regions?
• How sensitive is the model to station coverage?
• Can phase picking, association, and event detection be evaluated as a complete workflow?
• How efficient is the model when applied to large-scale continuous waveform archives?

AI4Seismo encourages benchmark design that better reflects these practical questions.

Planned Resources

This community plans to organize and release resources in the following directions.

1. Open Benchmark Index

A curated index of open seismic AI benchmark tasks, including:

• Phase picking
• Earthquake detection
• Event classification
• Continuous-waveform monitoring
• Cross-domain generalization
• End-to-end workflow evaluation

2. Open Datasets

Datasets related to:

• Continuous seismic waveforms
• Earthquake catalogs
• Phase labels
• Event-centered waveform windows
• Non-earthquake seismic events
• Regional monitoring datasets
• Deployment-level evaluation datasets

3. Open Models

Baseline and community-contributed models for:

• P-phase and S-phase picking
• Earthquake detection
• Event classification
• Seismic signal representation learning
• Foundation models for seismology

4. Demo Applications

Interactive Spaces for:

• Waveform visualization
• Phase picking demonstration
• Earthquake detection demonstration
• Model comparison
• Benchmark leaderboard
• Dataset exploration

5. Reproducible Evaluation

Evaluation protocols for:

• Precision, recall, and F1 score
• Phase picking time residuals
• Event-level recall
• False alarm rate
• Continuous-stream detection performance
• Cross-region generalization
• Deployment-level efficiency

Planned Repositories

The following repositories may be developed under this organization:

• awesome-ai4seismo
  A curated list of papers, datasets, models, tools, and benchmarks for seismic AI.

• seismic-ai-benchmark-index
  A benchmark index for seismic phase picking, earthquake detection, event classification, and continuous-waveform evaluation.

• continuous-waveform-benchmark
  A benchmark resource for evaluating seismic AI models under realistic continuous monitoring conditions.

• phase-picking-benchmark
  Benchmark protocols and baseline models for P/S phase picking.

• earthquake-detection-benchmark
  Benchmark protocols for earthquake detection from continuous waveform streams.

• event-classification-benchmark
  Benchmark resources for seismic event type classification.

• waveform-viewer
  An interactive waveform visualization demo.

• model-leaderboard
  A leaderboard for open seismic AI model evaluation.

Community Topics

We welcome discussions on:

• How to evaluate phase-picking models fairly
• How to define earthquake detection performance in continuous data
• How to measure false alarms in deployment scenarios
• How to evaluate end-to-end seismic monitoring workflows
• How to build open and reusable seismic AI datasets
• How to compare models across regions and monitoring networks
• How to handle non-earthquake seismic events
• How to design benchmark tasks for real-world seismic monitoring
• How to make seismic AI research more reproducible

Contribution Ideas

You can contribute by:

• Sharing open seismic datasets
• Submitting trained models
• Reporting benchmark results
• Proposing evaluation protocols
• Improving documentation
• Opening discussions
• Building demo applications
• Contributing visualization tools
• Reproducing published papers
• Suggesting open research questions

Suggested Benchmark Tasks

Task 1: Phase Picking on Continuous Waveforms

Evaluate whether a model can accurately pick P and S phases from continuous waveform data.

Possible metrics:

• P-phase precision, recall, and F1
• S-phase precision, recall, and F1
• Picking time residual
• False picks per hour
• Performance under different noise levels

Task 2: Earthquake Detection from Continuous Streams

Evaluate whether a model can detect earthquake events from continuous waveform streams.

Possible metrics:

• Event-level recall
• Event-level precision
• False alarm rate
• Detection latency
• Missed event rate
• Performance across magnitude ranges

Task 3: End-to-End Monitoring Workflow Evaluation

Evaluate a complete workflow including phase picking, association, and event detection.

Possible metrics:

• Event detection recall
• Event false alarm rate
• Associated phase accuracy
• Catalog-level consistency
• Deployment efficiency

Task 4: Seismic Event Classification

Evaluate whether a model can classify different seismic event types.

Possible categories:

• Natural earthquakes
• Mining-induced events
• Blasts
• Quarry blasts
• Noise-like events
• Other local seismic events

Possible metrics:

• Accuracy
• Macro-F1
• Cross-region performance
• Cross-domain generalization

Getting Started

If you are interested in AI for seismology, you can start by:

1. Following this organization
2. Opening a discussion
3. Introducing your research interests
4. Sharing useful papers, datasets, or tools
5. Suggesting benchmark tasks
6. Contributing models or evaluation results

For Researchers

AI4Seismo welcomes researchers working on:

• Seismology
• Earthquake monitoring
• Machine learning
• Deep learning
• Scientific AI
• Signal processing
• Geophysical data science
• Earth science informatics

We especially encourage contributions that improve reproducibility, openness, and practical deployment of seismic AI methods.

For Students

Students are welcome to participate through:

• Paper reading
• Code reproduction
• Dataset exploration
• Model testing
• Benchmark discussions
• Visualization development
• Literature collection

This community can also serve as a place to discuss research ideas, open datasets, and practical challenges in seismic AI.

Keywords

AI for seismology · seismic AI · earthquake detection · seismic phase picking · seismic event classification · continuous waveform · benchmark · deployment-level evaluation · open datasets · open models · AI for Earth science · scientific machine learning

Contact and Community

This community is under active development.

More resources, datasets, models, demos, and benchmark protocols will be added gradually.

Feel free to open discussions and contribute ideas.