Prestack simultaneous inversion of P-wave impedance and gas saturation using multi-task residual networks

It is challenging to realize the gas saturation (GS) estimation via the integration of seismic data and more complementary data (e.g., elastic attributes) in a traditional physical framework. Machine learning, especially multi-task learning (MTL), provides an alternative way for the fuse of multiple information and simultaneous inversion of two or multiple reservoir parameters without model-driven limitations and interactive operators. To improve the estimation accuracy of GS, we propose the prestack simultaneous inversion of P-wave impedance (PI) and GS using the multi-task residual network (MT-ResNet). The designed MT-ResNet consists of two task-related subnets. The first subnet establishes the nonlinear links among low-frequency PI, prestack seismic data, and well-log derived PI. Furthermore, seismic data and the inverted PI via the first subnet are jointly entered into the second subnet and evolved into the well-log interpreted GS. A model based on measured petrophysical parameters associated with the field deep tight dolomite reservoir is used to test the proposed method. Tests on the synthetic data example and the field example demonstrate that the MT-ResNet can simultaneously estimate PI and GS models with the highest reliability, in comparison with single-task residual network (ST-ResNet) and the conventional seismic inversion and rock-physics equations-based method. And the MT-ResNet inverted PI can be utilized as complementary information for improving the prediction accuracy of MT-ResNet inverted GS. Our proposed MT-ResNet has the potential to guide the design of the MTL-based multiple reservoir parameters prediction and practical application.