General machine learning-based algorithms have been developed to predict important matrix properties and an estimate of their uncertainties. Grain density, neutron porosities, cross sections for thermal neutron absorption (Sigma) and fast neutron elastic scattering (FNXS) are the considered properties. The models are trained in a proprietary core database of more than 2000 samples from a global distribution of sedimentary rocks. The performance of the matrix property predictions is demonstrated in independent core data and case studies from the field. The machine-learning models provide a significant improvement in accuracy over linear regression models, while eliminating the need for user inputs. The models help to account for the presence of unmeasured thermal absorbers (e.g., boron) through geochemical associations among the available log elements. The nonlinear models also provide better precision despite accounting for statistical noise on the input elements from logging measurements. We use field datasets to illustrate the improved performance in wells from the North Sea penetrating complex lithologies.


Every petrophysical log responds to some combination of the solid rock matrix and its fluid-filled pores. Knowledge of the matrix properties is usually required for accurate interpretation of fundamental quantities like porosity, saturations, and permeability. The matrix properties that affect log measurements span all the domains of formation evaluation, including nuclear, electromagnetic, and acoustic properties. Particularly in the case of neutron-based measurements, significant variation in the matrix response can be caused by trace amounts of thermal absorbers.

In traditional petrophysics, matrix properties may be fixed for an assumed lithology, they may be assigned with endpoints for a limited set of minerals, or they may be estimated via empirical relationships with elements from geochemical spectroscopy logs (Herron and Herron, 2000; Herron et al., 2002). All these classical approaches require manual input, leading to a degree of subjectivity in formation evaluation.

Nonlinear mapping techniques have been proposed in the past to predict mineralogy from geochemical elements (Freedman at al., 2015). However, these methods did not attempt to quantify uncertainty on the predictions, nor did they account for statistical noise in the spectroscopy measurement. More recent work on mineralogy (Craddock et al., 2021) applies a machinelearning architecture with explicit estimates of uncertainty and with a model trained for robustness in the presence of noise.

This content is only available via PDF.
You can access this article if you purchase or spend a download.