Multiphysics property prediction from hyperspectral drill core data

Abstract

Hyperspectral data provide rich information on both the mineralogical and fine-scale textural properties of rocks, which also control their petrophysical characteristics. We propose that some physical rock properties can be predicted directly from hyperspectral data, improving petrophysical characterisation and reducing the need for often laborious measurements. In this contribution we explore correlations between hyperspectral and petrophysical data using a deep convolutional neural network. Our model learns relevant features from high-dimensioned hyperspectral data to predict slowness, density, and gamma-ray values using training and testing data from Spremberg, Germany. Our results show that, with careful preprocessing and thorough data cleaning, differences in resolution can be overcome to learn the relationship between hyperspectral data and petrophysics. Using a test dataset from a spatially independent borehole, we generated a pixel-resolution (≈1 mm2) model of the petrophysical properties and resampled it to match the measured logs. This test indicated substantial accuracy, with R2 scores and root-mean-squared errors (RMSEs) of 0.7 and 16.55 μsm-1, 0.86 and 0.06 gcm-3, and 0.90 and 15.29 API for the slowness, density, and gamma-ray predictions respectively. We also analysed the Shapley values of our model to gain deeper insights into its predictions. These findings lay the groundwork for building deep learning models that predict physical and mechanical rock properties from hyperspectral data. Such models could provide the high-resolution but large-extent data needed to bridge the different scales of mechanical and petrophysical characterisation.