Estimating Petrophysical Properties Directly from Seismic: A Deep Learning Application to Carbonate Field for CO2 Storage Potential

Abstract Geological carbon capture and storage is vital for reducing carbon dioxide (CO2) emissions. Carbonate Field 1 in Luconia Province, offshore Sarawak is a potential CO2 storage site. Porosity and clay volume (Vclay) estimation from seismic provide valuable spatial and temporal information in characterizing reservoir distribution and overburden for assessing containment integrity and storage capacity. A deep learning inversion method for simultaneous estimation of porosity and Vclay was applied and tested in Carbonate Field 1. UNet architecture, chosen for its ability to preserve spatial resolution, processes post-stack seismic as input and petrophysical properties as outputs. Mean-squared error is implemented as the loss functions during the training on synthetic dataset. We use facies-based geostatistical simulation to generate 1D synthetic petrophysical logs. The petrophysical properties are linked to elastic properties through a carbonate rock physics model and fluid substitution using Gassmann's equation. The computed elastic properties are then used to calculate the angle dependent reflectivities (0°-30°) using the full Zoeppritz equations. The reflectivities are convolved with possible source wavelets derived from the seismic post-stack to generate the synthetic seismograms. These synthetic seismograms are then averaged to obtain a single noiseless synthetic seismic data, prior to addition of estimated field noise. 40,000 realizations of 1D synthetic datasets are generated for the training dataset. An equal distribution of 4000 realizations is allocated for both validation and testing datasets. Each 1D synthetic dataset consists of 128 samples, with a sampling rate of 4ms. The trained network model is applied to the testing data which has not been seen by the network during the training. Using Pearson correlation coefficient (cc) as the metric to evaluate the prediction performance, the model provides promising result (cc>0.80) for the estimated porosity and Vclay when evaluated on the testing data. Application on the field dataset of Carbonate Field 1 demonstrates satisfactory prediction performance, with cc value exceeding 0.65 and 0.85 for both the estimated porosity and Vclay respectively. Low (<0.2) and high (>0.5) estimated Vclay content is interpreted as carbonate and shale respectively. The findings allow interpreter to characterize the heterogeneity of carbonate depositional facies and quality by integrating the estimated porosity and Vclay for CO2 storage planning. The process of estimating the porosity and Vclay directly from seismic using the deep learning approach takes one week, which include training data preparation and implementation.