4D Seismic Inversion to Changes in Pressure and Saturation Using Deep Learning

An improved method for directly inverting to changes in pressure and saturations from the 4D seismic has been developed. Instead of using sim2seis modelling as part of the forward operator in generating the synthetic datasets for the network to train, we proposed an alternative method by using the products from the approach by Lew et al. (2023) as the geological frame to generate the synthetic datasets. Besides that, the network has been redesigned to support for map-based rather than pixel-by pixel training which helps to enhance lateral smoothness in the estimation of dynamic properties. Apart from introducing random noise, we have incorporated realistic noise models extracted from field data into the training synthetic dataset. Four network models have been trained on the training datasets of varying noise amplitudes. These trained models are subsequently applied to a 4D field dataset to estimate changes in pressure and saturations.