[S-CG50] Driving Solid Earth Science through Machine Learning

[SCG50-P12] Application of Convolutional Neural Networks for Seismic Velocity Model Building

*YU FAN¹, Ehsan Jamali Hondori², Jin-Oh Park¹ (1.Atmosphere and Ocean Research Institute, The University of Tokyo, 2.Geoscience Enterprise Inc.)

Keywords:Multi-channel seismic, Machine learning, Convolutional neural network

Velocity model building is a key step in seismic exploration. Accurate interval velocity models are required for reliable depth migration of multichannel seismic reflection data. The main product of velocity model building is a model of the subsurface that is subsequently used in seismic imaging and interpretation workflows. Conventional processes to develop a seismic velocity model (e.g. traveltime tomography or full waveform inversion) are generally operator dependent, time-consuming, and most of the time require a reasonably accurate initial velocity model to converge to a stable solution. In order to solve this problem, we propose a method to estimate the interval velocity models faster and more accurately by using Convolutional Neural Networks (CNN). The velocity model building with convolutional neural networks faces the lack of training data for various realistic problems, and one of the solutions is to use synthetic data in CNN model training. In the recently developed velocity model building tools, CNN have proven to be effective for simple geological settings (Araya-Polo et al., 2018; Simon et al., 2023). However, in more complex geologies such as subduction systems, CNN have a gap compared to traditional model building methods. In addition to training the CNN model with the simple layered geological settings, we implemented a more complex synthetic geological settings into the training process to solve this issue. The synthetic training dataset is generated by altering the “realistic” GO 3D OBS model (Górszczyk and Operto, 2021). Our CNN model uses the residual move-out semblance panel as input data for training and the labels for each input are from the corresponding interval velocity profile at the same location. In this study, our CNN model is trained by synthetic data and then tested on real seismic data after transfer learning. Although the CNN does require computational resources during the training step, the trained CNN can estimate reliable velocity models from semblance data within minutes, thus save processing time.


References
1. Mauricio Araya-Polo, Joseph Jennings, Amir Adler, and Taylor Dahlke, (2018), "Deep-learning tomography," The Leading Edge 37: 58–66.

2. Jérome Simon, Gabriel Fabien-Ouellet, Erwan Gloaguen, and Ishan Khurjekar, (2023), "Hierarchical transfer learning for deep learning velocity model building," GEOPHYSICS 88: R79-R93.

3. Górszczyk, A. and Operto, S., "GO_3D_OBS: the multi-parameter benchmark geomodel for seismic imaging method assessment and next-generation 3D survey design (version 1.0), " Geosci. Model Dev., 14, 1773–1799.