4D Data-driven Deep Waveform Learning: a Case Study at CCUS Site, San Juan NM
Abstract
Time-lapse full waveform inversion (FWI) is a powerful geophysical technique that uses seismic records to reconstruct high-resolution subsurface velocity maps at different timestamps, which is critical for monitoring CO2 storage. Though proven to be successful in meeting most industrial demands, FWI involves numerically solving partial differential equations twice in one iteration to calculate the gradient. Usually, hundreds if not thousands of iterations are necessary until a satisfied velocity map is reconstructed. Thus, a successful FWI requires tremendous computational cost in real-data implementations, and such cost will dramatically increase when dealing with large-scale time-lapse datasets. Recently, with the rapid development of computational resources, state-of-art deep learning (DL) methods have been leveraged to replace the heavy computational framework in FWI. DL requires training a network model to predict the subsurface structures efficiently from the recorded data. Such a training process demands a massive amount of labeled data, and the performance strongly depends on the generalization from the training dataset to the inference samples. Once the inference sample distribution is well covered by the training sample distribution, such in-domain prediction will produce convincing results. Thus, the quality of the training dataset is essential to the DL-based time-lapse FWI. In this work, we introduce two novel approaches to generate the training samples that would yield excellent representativeness of the inference set: one is based on a style-transfer network that creates seismic velocity maps from natural images (Feng et al., 2021); the other one is based on synthetically building targeted training samples from a baseline velocity and well-log information (Liu et al., 2021). We implement both approaches and test them on the data collected at the San Juan CO2 reservoir (located in New Mexico, USA) via the CarbonSAFE project funded by U.S. the Department of Energy. Both methods show promising results (figure) and great potential in industrial applications.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFM.U45B0524W