Applied machine learning-based models for predicting the geomechanical parameters using logging data

Abstract Geomechanical (GM) parameters play a significant role in geomechanical studies. The calculation of GM parameters by analyzing finite rock samples is very limited. The GM parameters show a nonlinear trend; thus, applying empirical relationships is unreliable to predict their quantities. Machine learning (ML) methods are generally used to improve the estimation of such parameters. Recent researches show that ML methods can be useful for estimating GM parameters, but it still requires analyzing different datasets, especially complex geological datasets, to emphasize the correctness of these methods. Therefore, the aim of this study is to provide a robust recombinant model of the ML methods, including genetic algorithm (GA)–multilayer perceptron (MLP) and genetic algorithm (GA)–radial basis function (RBF), to estimate GM parameters from a complex dataset. To build ML models, 48,370 data points from six wells in the complicated Norwegian Volve oil field are used to train GA–MLP and GA–RBF methods. Moreover, 20,730 independent data points from another three wells are used to verify the GM parameters. GA–MLP predicts GM parameters with the root-mean-squared error (RMSE) of 0.0032–00079 and coefficient determination (R 2) of 0.996–0.999. It shows similar prediction accuracy when used to an unseen dataset. Comparing the results indicates that the GA–MLP model has better accuracy than the GA–RBF model. The results illustrate that both GA–MLP and GA–RBF methods perform better at estimating GM parameters compared to empirical relationships. Concerns about the integrity of the methods are indicated by assessing them on another three wells.