Machine learning for the prediction and Analysis of Bouguer Gravity anomaly of the Kiri uplift in Congo sedimentary Basin

Abstract

Bouguer gravity anomalies (BGA) play an important role in exploration of mineral resources. Allowing the delineation of large geological structures, BGA participate into discovery of the deposits. However, the Kiri uplift region where several oil seeps have been recognized faces sparse coverage of data due to the difficult conditions of data acquisition on the field. This situation increases the non-uniqueness and nonlinearity problems of the solution using inverse methods. Although, potentially good at quantifying uncertainties, inverse approaches involve enormous computational tasks. We used machine-learning algorithms to predict and analyze gravity data in Kiri uplift region. The algorithms learned to perform as a multiple regression. During training steps, each independent variable included X and Y coordinates, digital elevation model (DEM) and geology. BGA values calculated by experts were provided as the dependent variables. K-fold cross-validation has been used ensure the models are well fit. Since the well-trained algorithms should result in small losses and errors, we experimented several optimizers. By comparison, testing results showed that deep neural network-based algorithm (DNN) has proven to be the most efficient with 5.37 Mean Squared Error and 1.75 Mean Absolute Error as model metrics. DNN showed the most accurate prediction, which, together with the measured BGA reported strongest Pearson correlation coefficient of 0.996. In addition, analysis showed that DNN result is one that conforms perfectly to the regional geology information of the study area. Machine learning algorithms proved their effectiveness to predict and analyze BGA in the study area.ML algorithms proved their effectiveness to predict BGA where measurements lackedThe joint analysis of predicted, measured and regional lithology meets expectationsAs predictors are X, Y, DEM and, lithology, we can customize the acquisition grids