Thickness estimation of CO2 transition layer using a deep learning

After CO2 injection into a reservoir, the behavior of CO2 depends on permeability, porosity, cap rock, reservoir fluids, CO2 characteristics, pressure gradient, and buoyancy effects. Therefore, the thickness of the reservoir is an essential parameter for CO2 monitoring. In the case of reservoir thickness prediction, it is practical to consider a geological reservoir as a transition zone in which the physical properties linearly change. In the transition zone, the seismic reflections in the stack section are the normal incident reflection coefficient with continuously changing velocity. Since this is composed of a function of the velocity ratio of the upper and lower layers, frequency, and transition zone thickness, the seismic signals apply to predict the thickness of the reservoir layer. In this study, we use the frequency characteristics with time-varying to estimate the thickness of the transition zone. First, we prepare the time-frequency spectrum with various thicknesses and then analyze it through deep learning to determine an optimum reservoir thickness. We use a convolution neural network (CNN) for predicting the transition zone thickness, which has two more hidden layers in the feature extractions. Unlike the fully connected layer, CNN is composed of a convolutional layer and a pooling layer and requires many data to prevent overfitting. Since CNN can efficiently process nonlinear data, it is applied to image classification and argumentation. For the numerical modeling experiment, we prepared a geological model in which the velocity of the shale layer (3000 m/s), cap rock, is greater than the lower sandstone layer (2200 m/s). We verify variation of phase and amplitude according to various transition zone thicknesses. For example, when the thickness is 10 m, it shows the phase changes at 65 Hz, and the amplitude decrease with increasing frequency. For the thickness of 50 m, the phase changes at the cut-off frequency of 13 Hz, and the amplitudes decrease until 25 Hz, increasing and decreasing repeatedly. We suggest that CNN is one of the methods to predict the thicknesses of CO2- injected reservoir using a time-frequency spectrum with various thicknesses.