Method Research
Extensive research experience in loss functions, training strategies, and data engineering raises the performance ceiling of the model.
Core technology · Seismic AI
Deploying research methods
as production systems
SeisBlue combines seismic-AI research experience with engineering capability. It builds training pipelines and integrates real-time data streams so that research methods run reliably in production. The full workflow and case studies follow below.
Offline
Data input
Earthquake catalogs paired with continuous waveforms become labeled training data.
Offline
Preprocessing
Cleaning, windowing, and arrival-time alignment turn raw waveforms into learnable samples.
Offline
Model training
Reproducible training pipelines manage experiment records, hyperparameters, and training strategies together.
Offline
Evaluation
P- and S-wave detection performance is validated on standard test sets, and a model is released only if it outperforms the current version.
Online
Deployment
Models are packaged as services and connected to real-time data streams, running reliably 7×24.
Online
Monitoring
Latency, performance, and data drift are monitored, and new events detected in production are collected as training data.
↺ New events return as the next round of training data, steadily improving model accuracy
Core capabilities
Three pipeline stages
and the corresponding core capabilities
Training Infrastructure
Reproducible, scalable training infrastructure, including data preprocessing, model architecture, and training strategy.
Production Deployment
Models are integrated into real-time data streams (e.g. Earthworm) so that services run reliably 7×24.
Method Research
BlueDisc · S-wave detection method
BlueDisc originates from Chun-Ming Huang's years of research on phase picking. Targeting the long-standing S-wave amplitude suppression problem in deep-learning phase pickers, the work dissects the role of the loss function in training and why the data fails to converge, and validates the findings through a shape-then-align training strategy with a conditional GAN proof of concept, raising effective S-phase detections by 64% on standard benchmarks. Source at github.com/SeisBlue/BlueDisc. For a non-technical explanation, see the article Why Does AI Often Miss S-Waves? →
Huang, C.-M., Chang, L.-H., Chang, I.-H., Lee, A.-S., & Kuo-Chen, H. (2025). Recovering Sub-threshold S-wave Arrivals in Deep Learning Phase Pickers via Shape-Aware Loss. arXiv:2511.06731. arxiv.org/abs/2511.06731
Training Infrastructure
SeisBlue Platform · seismic data framework
SeisBlue is a deep-learning seismic data processing framework developed and open-sourced by Chun-Ming Huang. The platform packages the full pipeline, from waveform ingestion through phase picking to catalog output, into a service that runs reliably in production, adopted by multiple peer-reviewed studies as a core processing tool. Source at github.com/SeisBlue/SeisBlue.
Huang, C.-M., Chang, L.-H., Kuo-Chen, H., & Zhuang, Y. (2023). SeisBlue: a deep-learning data processing platform for seismology. EGU General Assembly 2023, EGU23-13927. doi.org/10.5194/egusphere-egu23-13927
Follow-up research adopting the SeisBlue Platform (4)
- Sun, W.-F., Pan, S.-Y., Huang, C.-M., Guan, Z.-K., Yen, I.-C., Ho, C.-W., Chi, T.-C., Ku, C.-S., Huang, B.-S., Fu, C.-C., & Kuo-Chen, H. (2024). Deep learning-based earthquake catalog reveals the seismogenic structures of the 2022 MW 6.9 Chihshang earthquake sequence. Terrestrial, Atmospheric and Oceanic Sciences. doi.org/10.1007/s44195-024-00063-9
- Huang, B.-S., Ku, C.-S., Lin, C.-J., Lee, S.-J., Chen, Y.-L. E., Jiang, J.-S., & Sun, W.-F. (2024). The first 30 min hidden aftershocks of the 2022 September 17, ML 6.4, Guanshan, Taiwan earthquake and its seismological implications. Terrestrial, Atmospheric and Oceanic Sciences. doi.org/10.1007/s44195-023-00059-x
- Sun, W.-F., Pan, S.-Y., Liu, Y.-H., Kuo-Chen, H., Ku, C.-S., Lin, C.-M., & Fu, C.-C. (2025). A Deep-Learning-Based Real-Time Microearthquake Monitoring System (RT-MEMS) for Taiwan. Sensors, 25(11), 3353. doi.org/10.3390/s25113353
- Kuo-Chen, H., Sun, W.-F., Kan, L.-Y., Pan, S.-Y., Yen, I.-C., Liang, S.-H., Guan, Z.-K., Liu, Y.-H., Chen, W.-S., & Brown, D. (2025). Real-Time Earthquake Monitoring with Deep Learning: A Case Study of the 2025 ML 6.4 Dapu Earthquake and Its Fault System in Southwestern Taiwan. The Seismic Record, 5(3), 320–329. geoscienceworld.org
Production Deployment
TT-SAM · real-time earthquake early warning
TT-SAM (Taiwan Transformer Shaking Alert Model) is an earthquake early warning system. Its architecture references the TEAM framework by Münchmeyer et al. (2020); Chun-Ming Huang led the AI engineering core, including training environment setup, pipeline design, and production deployment. The real-time pipeline runs on the Earthworm real-time data stream.
Chen, Y.-H., Chan, C.-H., Chang, C.-C., & Ma, K.-F. (2026). A Deep Learning Framework for Peak Ground Velocity Prediction Using Multi-Station Velocity Waveforms: The Taiwan Transformer Shaking Alert Model (TT-SAM). Journal of Geophysical Research: Machine Learning and Computation. doi.org/10.1029/2025JH001005