Improving Lateral Continuity in Direct Petrophysical Inversion from Seismic Using Deep Learning

Estimating petrophysical properties directly from measured seismic using multi-realisation of 1D training synthetic database for deep learning training resulted in ‘jittery’ artifact. The 1D training datasets has random geological scenarios, where each realisation is independent and spatially uncorrelatable. A method is developed to generate realistic 2D training database that provides flexibility for the network in analysing neighbouring traces. The process of building realistic 2D training data involves the utilization of the estimated porosity, Vclay and hydrocarbon saturation (Shc) cubes from Lew et al. (2023) as the inputs. These properties were initially predicted based on network model (MLT2) that trained on 1D training dataset. 2D sections consist of 29,480 and 3,685 traces for training and testing data are selected respectively from the property cubes. The petrophysical properties are linked to elastic properties through Fast Xu-White model and fluid substitution using Gassmann's equation, then computation of angle dependent reflectivities (0°–55°) using the full Zoeppritz equations. These reflectivities are grouped according to the configuration of measured seismic, resulting in five groups of reflectivities. The reflectivities within each group are convolved with angle-dependent source wavelets to produce the 2D synthetic angle stacks. The computed synthetic seismics are added with estimated field noise. Modification is performed on the 2D training dataset by splicing it to contain 32 traces. Using the concept of stride, a step of one is used to slide over the 2D dataset one trace at a time, each slide contains 32 traces. This process allows the generation of mini-2D sections for the network to learn. UNet architecture, chosen for its ability to preserve spatial resolution, processes angle stack seismics as input and petrophysical properties as outputs. Mean-squared error and L1 norm are implemented as the loss functions during the training on synthetic dataset. The trained network model (ML2D) on 2D training data and the MLT2 are applied to the testing data which is unseen by the network. Using mean structural similarity index measure (mssim) as the metric, both models provide promising result (mssim>0.65) for the estimated properties when evaluated on the testing data. We then applied the models to the field dataset. The computed metric (Pearson correlation coefficient) from the blind test wells indicates that ML2D outperform the MLT2. The results from ML2D provide better lateral continuity for the estimated petrophysical properties compared to MLT2, as the 2D training data allows network to learn the neighbouring traces. A methodology to improve the lateral continuity of the estimated porosity, Vclay and Shc simultaneously from seismic has been tested. The method allows realistic generation of 2D training database, and improved the inputs and outputs shaping arrangement into the network architecture to allows flexibility in analysing neighbouring traces, potentially promoting lateral smoothness.