Reservoir Prediction Based On Machine Learning And Rock Physical Modeling

With the booming development of artificial intelligence and cross-application of multiple disciplines,machine learning has been applied to various aspects of seismic exploration,including pre-stack inversion,velocity modeling,reservoir prediction,fault identification,etc.,all with good results.However,machine learning is a data-driven process whose ultimate goal is to obtain knowledge directly from the data,and the models built are statistical models of the data,which do not have physical interpretation capability.Also in seismic exploration although there is a large amount of data,there are often few high quality labels,which leads to often insufficient samples for machine learning,increasing the empirical risk and making the statistical model inadequate in generalization.In this paper,we introduce theoretical rock physical models into the construction of label libraries for machine learning algorithms to improve the physical interpretability of machine learning algorithms by using theoretical rock physical models to supplement the dataset with forward simulation and incorporate geological and geophysical knowledge into the dataset through a forward process.First,we established three types of theoretical rock physical models for different types of reservoirs,including conventional sandstone reservoirs,tight sandstone reservoirs,and fractured limestone reservoirs,and then combined the theoretical rock physical models with machine learning algorithms to establish more noise-resistant rock physical models,and verified the feasibility of these three types of rock physical models in actual production using well data from actual workings.Through the random combination of modeling parameters,the three types of rock physical models are used to traverse the rock physical elastic parameter space of the whole reservoir and construct a rich full-sample label library.Based on the full-sample label library,four parameters(compression wave velocity,clay content,porosity,and density)that are highly correlated with shear wave velocity are screened out and combined with deep learning algorithms to construct a deep shear wave velocity prediction network with good generalization,strong noise resistance,and non-repetitive training.The error between the predicted shear wave velocity and the measured data in laboratory and the logging data in actual work area is very small,and the prediction effect is better than that of the empirical formulas.Based on the idea of two-wheel drive of data and model,we also constructed a deep lithology identification network which achieved a lithology identification accuracy of more than 93% on synthetic data and actual well data,and 69% on seismic data.Based on the established rock physical model and random forest algorithm to filter the fluid sensitivity factor of conventional sandstone reservoir,different machine learning algorithms were used to identify the fluid in the actual working area,and the identification accuracy reached more than 84%,among which the back propagation neural network had the highest identification accuracy of 89%,the support vector machine algorithm was the second and the random forest algorithm had the lowest identification accuracy.The reservoir prediction method based on machine learning and rock physical modeling solves the problems of lack of high-quality labels,poor interpretability of machine learning algorithms,and poor generality of models in actual production,and provides a new idea of two-wheel drive based on data and model for reservoir prediction.