Random Forest with Spectral Extrapolation

Machine learning models can be used to predict subsurface properties in 3D, with the advantages of increased resolution and independence from flawed physical assumptions. Spectral Geosolutions has developed a robust Random Forest machine learning solution for 3D prediction of log, acoustic, elastic and petrophysical properties in the subsurface. Missing logs can also be estimated from measured logs using this technique.

Random Forest is an ensemble machine learning technique that uses bagged decision trees with sample replacement in order to predict target variables from observations (attributes). The average set of values from a suite of decision trees is taken as the solution, which decreases the variance of the model without increasing the bias, thereby boosting immunity to over-training.

While the Random Forest algorithm (and other machine learning algorithms) naturally boosts the frequency content at the location of the training well(s), the accuracy of the high frequencies decays as rock properties vary laterally with distance from the well(s). By using enhanced seismic data from our Spectral Extrapolation as input to Random Forest, the reliability of the high frequency content is significantly improved, thus maximizing vertical AND lateral resolution for subsurface characterization.
***For more details, please see Publications section - Puryear et al, 2021.***