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

This study presents a deep learning (DL)-based 4D inversion approach for estimating dynamic reservoir properties—pressure change, water saturation change and gas saturation change—from 4D seismic data. The workflow reduces reliance on reservoir simulation-driven static modelling stage in the sim2seis pipelines by using DL-derived 3D petrophysical volumes as static framework, combined with fluid flow consistent dynamic perturbations to build the 4D synthetic training datasets. A map-based UNet architecture is adopted to incorporate spatial context and promote lateral consistency in the estimated property maps. However, this spatial coupling can also propagate noise-related artefacts into low seismic confidence regions. To account for seismic noise, both Gaussian and more realistic field-specific simulated noise—derived from overburden seismic differences—were introduced into the training data. Although adding specific overburden-derived noise in the training data produced the lowest average error for the case study considered, this outcome is field-dependent and does not imply general superiority over Gaussian noise. Applied to 4D field data from the Malay Basin, the inversion produced physically consistent results—such as simultaneous increases in pressure and water saturation near injectors, and gas saturation increases associated with pressure depletion near producers. Model uncertainty was assessed using a deep ensemble approach, with the resulting standard deviation maps aligning with pre-defined confidence zones.