Imaging geomechanical properties of shales with infrared light

Hyperspectral images are data cubes that associate with each pixel an infrared (IR) reflectance spectrum. We hypothesize there is an implicit but discernable link between IR spectra of rocks and their mechanical response that can be learned by a machine learning (ML) algorithm. Our rationale is that IR spectra are sensitive to both rock composition and microstructure, two attributes that determine all hydro-thermo-mechanical properties of the rock. To test this hypothesis, we confine ourselves to shale specimens from the Green River formation, Utah, US, and aim to map acoustic velocity (V_p), acoustic attenuation coefficient (A_p), and X-ray attenuation coefficient (µ) from hyperspectral images acquired in the near-, shortwave-, and mid-IR. We propose an experimental workflow that generates large quantities of labelled data needed to train supervised ML algorithms. The trained algorithms are shown to accurately predict V_p, A_p, and µ directly from pixelwise IR spectra. In some cases, even a 1% random sampling of pixels is found sufficient for training. While both artificial and convolutional neural networks are probed, performances are similar, but better than linear regression. More important is ensuring the training data are representative of all lithologies to be targeted during prediction. This work implies the feasibility of extrapolating beyond a handful of point or line measurements from indentation, scratch testing, or field surveys to, respectively, cm-scale, m-scale, and potentially outcrop-scale rock faces.