Blockly Earthquake Transformer: A Python Toolbox for Customizing Seismic Phase Pickers

The state-of-the-art deep learning model Earthquake Transformer (EqT) outperforms most or all other AI techniques in seismic signal detection and phase picking tasks. However, applying EqT to a specific research target is still challenging for those researchers with little knowledge of deep learning programming, e.g., how to format the required input data, how to re-train the model and how to improve the performance with new data. To overcome these obstacles, we explore the potential of building a block-based interactive interface for EqT. We extend EqT to broaden its accessibility and reproducibility with blockly earthquake transformer (BET). BET is a no-code, deep learning approach using the EqT model. BET presents to users an interactive form enabling them to upload their data and customize the model and arguments to create their workflows. Once the form is filled out, BET executes the corresponding phase picking task without requiring the user to interact directly with the code. This tool is designed to simplify EqT applications to various fields, such as local events, teleseismic earthquakes and microseismicity. Here, users are able to detect events using the trained EqT model, train and evaluate new models with the original EqT architecture. In addition, transfer learning and fine-tuning functions are implemented in BET. In the transfer learning module, BET extends the phase picking range from P and S phase to additional phase types, e.g., Pn, Pg, Sn, Sg (based on the labeled training data), etc. In the fine-tuning module, detailed model architecture can be customized by users to build new models that may achieve better performance on specific projects than currently published models. We also present results from a small-scale transfer learning to accomplish Pg and Sg picking tasks in BET. The contribution of this work is in showing the potential for fast deployment of reusable workflows, building customized models, visualizing training processes and producing publishable figures in a lightweight, interactive, open-source Python toolbox.