Improving the Predictive Ability of A Geomechanical Model Using Neural Networks (Deep Learning)

Abstract

Summary With the growth of hydrocarbon consumption and depletion of developed fields, the task of extracting hard recovering oil reserves becomes actual. To solve this problem, modern drilling methods are used, such as the drilling of directional and horizontal multi-lateral wells. Also methods for increasing oil recovery are used, such as multi-hydraulic fracturing. In the first case, it is important to avoid possible complications during drilling associated with absorption and collapse of the borehole walls. In the second one for successful hydraulic fracturing, it is necessary to understand stress field distribution in the formation, for selecting successful fracture initiation point and determining injection volumes. These problems could be solved with three dimensional geomechanical modeling. One of the main input for geomechanical model is well logs. In common, the required volume of surveys is partially absent in wells selected for modeling. In this paper, effective tool for log data recovery was considered - deep learning neural network. A number of mathematical approaches was observed: • One-hot coding Each layer has its own unique lithology, strength and elastic properties. One-hot-coding allows to convert categorical data “reservoir name” into a digital format that is convenient for neural network. It allows automatically find out a unique property correspondence for each individual layer. • Construction of additional features Using mathematical transformations of input data allows better processing of peak curve values, increasing model prediction accuracy. • Residual connections, regularization of weights and dropping These methods were used to solve the problem of damped gradient in a multilayer neural network model. • Hyper parameters automatically configuring and callbacks They allow for more sensitive tuning of network hyper parameters, and save weights of the model with the best validation result. • Optimizer observation Usually, the SGD optimizer is used - Stochastic gradient descent optimizer. In this neural network, more efficient optimizers such as ADAM and RMSPROP were tested and used. • Ensemble models It is another technique for increasing neural networks predictive ability. Ensemble models use an average value of neural networks results with different architectures, for example, LSTM and fully connected network. To sum up, listed approaches allowed us to obtain a high-precision tool for solving the regression problem. It significantly surpasses methods of multi-regression, which are usually used to restore data in geomechanics and allows to get a complete set of data for high-quality 3D geomechanical modeling.