Machine learning workflow to predict multi-target subsurface signals for the exploration of hydrocarbon and water

NMR T1 distribution is the multitarget subsurface signal used in this study. NMR logs are acquired as depth-wise measurements of T1 and/or T2 distributions, usually every 0.5 ft, along a wellbore. NMR T1/T2 distributions contain valuable information about the in-situ permeability, viscosity and movable fluid volumes in the near-wellbore region. However, NMR logs are not readily available due to operational and financial constraints. A robust machine learning workflow is developed to process conventional "easy-to-acquire" well logs (e.g. resistivity, neutron, density, sonic and spectral gamma ray) for the depth-wise multitarget synthesis of NMR T1 distribution along the length of an entire well drilled into a hydrocarbon-bearing geological formation. Random forest model performs the best with average R2 score of 0.84, MMAPE of 0.14, and RMSE of 0.4. The random forest model is then enhanced using quantile regression forest for computing the "confidence index" associated with multitarget synthesis of NMR T1 distribution at each depth. The confidence index quantifies the certainty of multitarget synthesis on new, unseen data. The proposed data-driven workflow can be used for robust prediction/synthesis of any multi-target subsurface signal and provide a measure for evaluating the accuracy and certainty of the predicted signals.