Developing GAN-boosted Artificial Neural Networks to model the rate of drilling bit penetration

Abstract

The goal of achieving a single model for estimating the rate of drilling bit penetration (ROP) with high accuracy has been the subject of many efforts. Analytical methods and, later, data-based techniques were utilized for this purpose. However, despite their partial effectiveness, these methods were inadequate for establishing models with sufficient generality. Based on deep learning (DL) concepts, this study has developed an innovative approach that produces general and boosted models capable of more accurately estimating ROPs compared to other techniques. A vital component of this approach is using a deep Artificial Neural Network (ANN) structure known as the Generative Adversarial Network (GAN). The GAN structure is combined with regressor (predictive) ANNs, to boost their performance when estimating the target parameter. The predictive ANNs of this study include Multi-Layer Perceptron Neural Network (MLP-NN) and 1-Dimensional Convolution Neural Network (1D-CNN) structures. More specifically, the key idea of our approach is to utilize GAN’s capability to produce fake samples comparable to true samples in predictive ANNs. Therefore, the proposed approach introduces a two-step predictive model development procedure. As the first step, the GAN structure is trained with the target of the problem as the input feature. GAN’s generator part, which can produce fake ROP samples similar to the true ones after training, is frozen and then replaces the output layer’s neuron of the predictive ANNs. In the second step, final predictive ANNs carrying frozen trained-generator called GAN-Boosted Neural Networks (GB-NNs) are trained to make predictions. Because this approach reduces the computational load of the predictive model training process and increases its quality, the performance of predictive ANN models is improved. An additional innovation of this research is using the residual structure during 1D-CNN network training, which improved the performance of the 1D-CNN by combining the input data with those features extracted from the inputs. This study revealed that the GB-Res 1D-CNN model, a GAN-Boosted 1-Dimensional Convolutional Neural Network with a Residual structure, results in the most accurate prediction. The validity of the GB-Res 1D-CNN model is confirmed by its successful implementation in blind well. As the final step of this study, we conducted a sensitivity analysis to identify the effect of different parameters on the predicted ROP. As expected, the DS and DT parameters significantly affect the model-estimated ROP.