Application of Neural Networks for the Assessment of linking Damage Zones of Geological Faults

Resumo

The definition of damage zones in geological faults is important for oil and gas exploration due to the potential occurrence of preferential flow paths through such zones. In practice, some damage zone thicknesses in isolated faults are evaluated using exponential laws based on statistical relationships with fault displacement. However, in the field, linking damage zones resultant from fault interactions are observed, compromising the production process in reservoirs. One way to estimate the geometry of linking damage zones is through numerical modeling using the Finite Element Method and elastoplastic constitutive models. Those models impose displacements or throws over fault surfaces, triggering plastic zones obtained in Gauss Points around the faults which are related to the damage zones. However, model building, simulation execution, and result analyses can be time-consuming, hindering decision-making in field operations. As an alternative, the present study proposes the construction of a neural network to predict damage zones resulting from the interaction of two geological zones. The dataset used was built from numerical results of several analyses using a set of geometrical parameters such as overlap, separation, and angle between two faults. This strategy is interesting because it reduces the dimensionality of the number of parameters to be sampled while maintaining the key information to reconstruct the data in global coordinates. The fault setting as well as the results obtained from the simulations at each Gauss point were used as input parameters to train the neural network in order to identify the damage zones given by plastic deformations. A regression strategy is used to preserve the magnitude of the results. The Adaptive Moment Estimation (ADAM) and hyperbolic tangent (Tanh) is being used as optimizer and activation function, respectively. 80% of the dataset is being used for training and 20% for testing, resulting in approximately 9 million data points for adjustment and 2 million for validation. The preprocessing used for network inference consists of generating a grid of points considering a fault arrangement. After inference, the network predicts the equivalent plastic strain field. In this way, it is possible to interpret the results obtained using contour lines. Therefore, it is expected to predict the occurrence of linking damage zones more quickly while considering the geometry of two faults.