Estimating total organic carbon (TOC) of shale rocks from their mineral composition using stacking generalization approach of machine learning

A fundamental parameter in the exploration and development of unconventional shale reservoirs is total organic carbon (TOC). To achieve reliable TOC values, it requires a labour intensive and time-consuming laboratory experiment. On the other hand, models have been proposed using geophysical well logs as input variables with little attention paid to the contribution of mineralogical parameters in the evaluation of TOC. In this paper, a novel stacking machine learning technique is examined to generate accurate TOC predictions from the mineral content of the shale rock in the Sichuan Basin. The stacking machine learning model involves first-level models of multivariate adaptive regression spline (MARS), random forest (RF) and gradient boosted machine (GBM) known as base learners, while MARS was further used in the next step as the meta learner model. The research result indicated that the stacking TOC model outperformed the single applied models of MARS, GBM and RF. The proposed stacking TOC model generated estimates having the least error statistics of 0.29, 0.54 and 0.54 for MSE, RMSE and MAPE respectively while producing the highest correlation of 0.86 during the model validation stage. Therefore, stacking machine learning approach permits an improved estimation of TOC from the mineralogy of the rock.