PS-VSP: Deep Learning for qP and qS Arrival Picking in Vertical Seismic Profiles

This repository includes the raw dataset and manual picks for qP- and qS-waves. Our machine learning approach aims to support the inversion of elastic anisotropy in metamorphic bedrock.

Data Generation:

Geological Model: Synthetic data is produced using a two-phase model from the Deep Fault Drilling Project (PFPD-2b) on the Alpine Fault, New Zealand, featuring an upper sediment layer and a lower metamorphic bedrock layer.

Modeling: Forward modeling was performed with Devito on a rotated staggered grid by varying the elastic stiffness of the metamorphic bedrock.

Manual Picking: The corresponding manual picks were completed using SLB VISTA desktop seismic data processing software.

Machine Learning Workflow:

A Python script demonstrates the training process using PyTorch with a three-layer U-Net.

Input: 3-component velocity images (vx, vy, vz) from a continuous borehole.

Output: Probability maps for qP- and qS-waves.