Comparison of Innovative Deep-Learning Model, Multilayer Perceptron, and Extreme Gradient Boosting for Prediction of Elastic Properties

Summary This work presents an innovative methodology, called Log FM, a wellbore log-related foundation model that combines deep learning, self-supervision, and transfer learning to predict elastic properties using commonly available wellbore logs. We test this method on a subset of wells from the Williston Basin, an important play for hydrocarbon production in the USA; the model predicts compressional slowness (DTC), shear slowness (DTS), bulk modulus (K), shear modulus (G), and Young’s modulus (E) using inputs such as gamma ray (GR), neutron porosity (NEUT), bulk density (RHOB), photoelectric factor (PEF), caliper (CALI), depth, and deep resistivity (RES_DP). Log FM is trained as a foundation model using a self-supervised learning strategy that captures interlog correlations and sequential dependencies. Once pretrained, the model is fine-tuned for specific tasks, such as predicting elastic properties. We benchmark Log FM against two widely used machine learning (ML) methods, extreme gradient boosting (XGBoost) and multilayer perceptron (MLP), using multiple wells with varying formation types and data availability. Compared with these baselines, Log FM achieves superior accuracy in both complete and incomplete data scenarios, resulting in lower mean squared error (MSE), mean absolute error (MAE), and a stronger correlation with ground-truth measurements. Notably, Log FM handles missing input logs and predicts multiple outputs simultaneously, capabilities that are often lacking in traditional ML methods. The model also generalizes effectively from limited training data, making it particularly suitable for real-world applications such as automated 1D geomechanical modeling and velocity model construction for seismic workflows. These results highlight the potential of Log FM to advance log analysis by improving accuracy, reducing manual effort, and enabling scalable, data-driven characterization of subsurface formations.