A deep CNN-LSTM model for predicting interface depth from gravity data over thrust and fold belts of North East India

Abstract

Geological interface depth modeling from the gravity field data is crucial for the exploration of oil and gas, mapping of sediment-basement interfaces and many other geological modeling studies. Two-dimensional radially averaged power spectrum analyses yield ensemble mean depth from the gravity anomaly which lacks information away from the focus of the depth point modeled. This paper aims to introduce adeep learning techniquecombining a convolutional neural networkand long short-term memory network (CNN-LSTM) with variogram modeling to model interface depths using geospatial coordinates, Bouguer gravity anomaly and altitude variations data over the thrust and fold belts of North-East India. Prior to actual data analysis, the method is tested on a synthetic model. After the optimal network selection with the ‘Relu’ transfer function and ‘Adam’ optimization, the robustness of the CNN-LSTM model is examined and the test results show that the model is robust up to the data assorted with moderate level (∼20%) of correlated noise. The CNN-LSTM model detects important geological structures, subsurface faults, and thrust zones namely, Kohima Synclinorium, Naga and Disang thrust, and Eastern boundary thrust. Comparative results suggest that the CNN-LSTM is skillful based on the mean-squared error (MSE) between the observation and the prediction (MSECNN-LSTM ∼ 0.014) relative to the other machine learning models (Adaptive-Neuro-Fuzzy-Inference System (MSEANFIS ∼ 0.020), Random Forest with Extra Tree(MSERF ∼ 0.075), Support Vector Regression (MSESVR ∼ 0.025), Bayesian Neural Network(MSEBNN ∼ 0.116), Convolutional Neural Network(MSECNN ∼ 0.022), Long Short-Term Memory Network(MSELSTM ∼ 0.023), and Gaussian Process(MSEGP ∼ 0.095)) used in the study. We, therefore, conclude that the proposed approach is robust enough to model the various geological interface depths precisely with appropriate input data of geospatial coordinates, Bouguer gravity anomaly and altitude variation. The approach used here has potential and could be explored further in other complex geological provinces around the world.