Deep-learning-based surrogate model for forward and inverse problems of wave propagation in porous media saturated with two fluids

The inversion of parameters characterizing reservoir property based on rock physics model is an important content in seismic exploration, while the strong nonlinearity between basic parameters and seismic attributes makes it challenging in practice or computationally expensive. This study focuses on the issue of wave propagation in porous media saturated with two immiscible fluids and proposes to utilize deep neural network (DNNs)-based surrogate model to conduct forward analysis and parameter inversion. First, three classical rock physics models were employed to generate the dataset for training DNNs. Among them, five basic parameters (the porosity, bulk modulus and shear modulus of rock matrix, the solid density, and the saturation) were sampled within each reasonable range, and then, the corresponding phase velocity and inverse quality factor of seismic wave were calculated based on the three models. Next, DNNs were trained as surrogate models of the three rock physics models for sensitivity analysis. Then, such approach was applied to parameter inversion and different scenarios with different provided information and unknown quantities were designed to investigate the feasibility. The results showed the sensitivity of different models to basic parameters determines the corresponding inversion accuracy. When providing the information on the compressional wave, the porosity can be well inferred in most cases, while the accuracies of inverting saturation based on different models vary. Furthermore, it was found that adding the information of the shear wave can improve the performance of parameter inversions to a high level. From the obtained results, it was concluded that the DNNs-based surrogate models can serve as efficient tools for forward and inverse problems.