All-in-one proxy to replace 4D seismic forward modeling with machine learning algorithms

Abstract

Time-lapse seismic data provides valuable information to assist reservoir monitoring, model-based reservoir management, and development activities. 4D seismic information can be used quantitatively to estimate saturation-pressure changes or update reservoir simulation models through a data assimilation process, which traditionally requires a 4D seismic forward model that includes two steps: (1) petro-elastic model to relate rock and fluid properties to elastic attributes and (2) seismic model to depict wave propagation in the reservoir. The traditional forward model brings some complications in quantitative applications, such as time-lapse seismic history matching. Its multidisciplinary nature is a significant bottleneck to update simulation models, once observed 4D seismic data is used with production data in the assimilation process. Furthermore, it is time-consuming, especially in three-dimensional simulation models within data assimilation with several iterations. Here, the traditional forward model needs to be run hundreds or even thousands of times. Aside from being time-consuming, it requires substantial computer memory for complex geology, where more sophisticated seismic models are needed. In 4D seismic quantitative applications, the timely use of the information and its full benefits fade away due to the computational cost and time of the 4D seismic forward model. This research proposes a proxy (we call it, S4D-Proxy) to replace the traditional approach. The S4D-Proxy leverages machine learning models to detect hidden patterns and learn relations between input features (e.g., porosity and saturation-pressure changes) and the target (time-lapse difference of seismic amplitude). Training data are prepared and fed to the machine learning algorithms to relate the inputs and the desired output. The performances of the machine learning models are evaluated based on a numerical measure (coefficient of determination) and visual comparison of the results with the traditional forward model. Our results show that the average coefficient of determination for the test dataset is in an acceptable range. Moreover, the visual comparison of the proxy predictions with those from the traditional approach shows high similarity. Most hardening and softening 4D signals are reproduced with the proxy models. The main advantage of our approach is that it lowers the computational cost and time. The application of the proxy in data assimilation process could offer a significant speed advantage with faster 4D seismic forward modeling for data assimilation iterations. Unlike the traditional forward model, which is one step (petro-elastic model) then another (seismic model), our approach performs the forward modeling at once (all-in-one proxy model). It also offers an alternative 4D seismic forward model for time-lapse seismic feasibility study. All these advantages make the S4D-Proxy valuable and practical in permanent reservoir monitoring (PRM), where the reduced turnaround time in the forward model indicates the timely use of the 4D seismic information.