Neural Network-Based Automation of Seismic Wave Arrival Time Picking in Subsurface Imaging

• Grant number: 25/12920-4
• Support Opportunities: Scholarships abroad - Research Internship - Scientific Initiation
• Start date: November 01, 2025
• End date: February 28, 2026
• Field of knowledge: Physical Sciences and Mathematics - Geosciences - Geophysics
• Principal Investigator: Luigi Jovane
• Grantee: Otavio Tamio Fukushima Neto
• Supervisor: Cedric John
• Host Institution: Instituto Oceanográfico (IO). Universidade de São Paulo (USP). São Paulo , SP, Brazil
• Institution abroad: Queen Mary University of London, England
• Associated to the scholarship: 25/03205-0 - Integration of Processing and Inversion Methods for High-Resolution Multichannel Seismic Data with Sparker Source, BP.IC

Abstract

Seismic data is essential for the characterization of subsurface structures and has significant applications in exploration of oil & gas, and recently offshore wind farm projects.This data can produce high resolution imagery of the subsurface up to the third dimension, providing important insights to geoscientists and engineers. This imagery is obtained through data processing, a complex and time-consuming assignment. It consists of a fundamental sequence of steps, including demultiplexing, sorting, and applying corrections in the time domain. One of these corrections, known as static correction, is particularly demanding, as it must be applied to each seismic trace to adjust the wave arrival times. The determination of these arrivals, which will be used for the static correction, is an operation that holds potential for automation through artificial intelligence. To explore this potential, a neural network trained with processed seismic data from the database of CORE (Oceanographic center of stratigraphic registers IO - USP) will be studied. The main data considered are three seismic profiles annotated with its corresponding wave arrival picks. Each seismic trace serves as the input, while the picks act as the target values. To evaluate the model, metrics such as the Mean Absolute Error (MAE), Mean Squared Errors (MSE), Root Mean Squared Errors (RMSE) and R² Score will be analyzed. In order to train and optimize the model, this Research Internship Abroad (BEPE) project proposes a collaborative research at Queen Mary University of London to develop a functional neural network for automated seismic picking. Professor Cèdric John is the head of the Data Science for Environment and Sustainability group with extensive experience in machine learning applied to Earth Sciences.