Time-lapse seismic data has shown great promise in accurate monitoring of CO2 injection sites. There are many sources of uncertainty in derived rock porosity and CO2 saturation from time lapse seismic data. Variability in noise during data acquisition and noise inherent in seismic data can degrade signal quality and contribute to uncertainty in derived saturation estimates. The process of deriving saturation estimates involves solving an ill-posed, non-unique and highly non-linear seismic petrophysical inversion. Deep learning, particularly utilizing convolutional neural networks (CNNs), has demonstrated potential in addressing such complex and nonlinear seismic inversion challenges. Neural networks frequently find it challenging to offer reliable uncertainty estimates similar to those achieved with Markov Chain Monte Carlo (MCMC) techniques which are widely recognized for their statistical rigor in solving inverse problems. However, MCMC techniques are computationally expensive due to the need for repeated forward model evaluations to adequately sample the posterior distribution. To address this issue, we investigate the use of Invertible Neural Networks (INNs) to predict the full posterior distribution of porosity and CO2 saturation directly from time lapse data and capture the related uncertainty. INNs provide bijective mapping between data (input) and models (output) and uses a latent vector sampled from a Gaussian distribution to model the uncertainty. Our proposed approach is validated using two seismic vintages and well-logs from the Cranfield reservoir.
扫码关注我们
求助内容:
应助结果提醒方式:
