Rock Permeability Forecasts Using Machine Learning and Monte Carlo Committee Machines

We developed new concepts of extended Monte Carlo cross validation and Monte Carlo committee machines. We subsequently used those concepts to predict permeability by linear regression and machine learning methods such as Neural Networks, Support Vector machines, and Regression Tree. Among the parameters we calculated using extended Monte Carlo cross validation are: root-mean squared error of individual forecasts, forecast bias, correlation between forecast and actual permeability, and forecast instability as a measure of sensitivity to perturbations of the training set. Output of Monte Carlo committee machines is constructed as the average of machine learning outputs generated from multiple versions of perturbed training sets. We observed that Monte Carlo committee machines produced high stability forecasts, while individual machine learning forecasts (e.g. a single ANN) were characterized by lower stability. Higher accuracy forecasts were achieved when we applied machine learning methods and linear regression using permeability models that included quantitative and categorical predictors and second-order interactions among the predictors.