Presentation Information
[S06-05]Convolutional Neural Networks for Seismic Velocity Model Building and Application to Field Data
*Fan Yu1, Ehsan Jamali Hondori2, Hamzeh Mohammadigheymasi1, Jin-Oh Park1 (1. Atmosphere and Ocean Research Institute, The University of Tokyo, 2. Geoscience Enterprise Inc.)
Velocity model building is a key step in seismic exploration. Accurate interval velocity models are essential for reliable depth migration of multichannel seismic reflection (MCS) data. The main outcome of velocity model building is an initial subsurface model that is subsequently used in seismic imaging and interpretation workflows. Conventional approaches to velocity model construction, such as traveltime tomography and full waveform inversion (FWI), are generally operator dependent, time-consuming, and often require a reasonably accurate initial model to converge to a stable solution. To address these challenges, we propose a method that leverages Convolutional Neural Networks (CNNs) to estimate interval velocity models more efficiently and accurately.Recent studies have demonstrated that CNN-based velocity model building tools can be effective in relatively simple geological settings (Araya-Polo et al., 2018; Simon et al., 2023). However, in more complex geologies, CNNs have a gap compared to traditional model building methods. Moreover, due to the limited availability of training data, CNNs typically perform poorly when applied to field datasets.To address this issue, we synthesized 500,000 1D and 3,000 2D velocity models based on the Marmousi2 model (Martin, G. S., 2004) to pre-train our CNN. Our workflow first trains the CNN with 1D velocity models and then transfers the learned features to 2D velocity estimation. We employ a U-Net architecture to extract features from seismic gathers, which are subsequently mapped to interval velocity profiles in both the time and depth domains. The CNN takes common midpoint (CMP) gathers as input, with corresponding interval velocity profiles (time and depth domain) serving as labels. The network was trained on synthetic datasets and validated against the Marmousi2 benchmark model. Finally, we tested the trained CNN on field data line K1 from the survey KH24E1 in the Noto Peninsula, showing that the trained CNN generalizes well from synthetic to field data.