Year: 2025
Type: article
Abstract: Assessing the risk associated with drilling and wellbore stability studies requires the shear sonic log. These logs apart from distinguishing formation fluid from lithology are needed to obtain geo-mechanical rock parameters required for the safe design of rock fracturing. Although sonic logs are of great importance, they are usually not obtained due to the limiting cost of acquisition. Neural networks have been used to generate these logs to save cost, but these networks are prone to overfitting. The dropout rate has been proposed to tackle this problem, however selecting the optimum dropout probability rate can be challenging manually or expensive computation wise. This research therefore investigated concrete dropout, a dynamic technique for adapting the dropout rate of a neural network to the data. The concrete dropout technique was applied to an artificial neural network (ANN) and a convolutional neural network (CNN) model to predict the shear sonic log with Monte Carlo simulation. Comparison was also made with the deterministic ANN and CNN models which had no dropout layers added and a Bayesian-optimized multilayer perceptron (MLP) model. These models were trained and validated with four (4) wells from the Volve field, using features with the highest correlation. The Concrete dropout ANN was found to outperform both the deterministic versions and the MLP model with R2, RMSE, MSE and MAE scores of 0.9548, 3.6415, 2.4433 and 0.0179 respectively. The neural networks built in this study showed an enhanced predictive performance with concrete dropout addition over the networks with no dropout added, showing that the technique was able to adapt the dropout rate to fit the nature of data and improve performance, which finds great application in real time deployment. The findings of this study also proposed a cost-effective way of sampling and averaging multiple outputs from a single neural network model, leading to enhanced predictive performance as the addition of concrete dropout allowed the network output distributions rather than point predictions.
Authors: Ephraim Ojoajogu Enemali, Sulaiman Alhaji Dodo Ibrahim, Ibraheem Salaudeen, Mahmud Abubakar Abdulmalik
Institution: Abubakar Tafawa Balewa University
Language: English
Subfield: Ocean Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W4411655018