GRSG 35th Conference 2024 Abstract

Title: Integrating Single-Point Spectral And Scattered X-Ray Fluorescence Data To Generate Compositional Maps Of Rock Surfaces: A Machine Learning Approach

Author: Raphael Bianchi Hunger

Organisation: University of Campinas

Coauthors: Souza Filho, CR¹; Miranda, LP¹; Scafutto, RDPM¹; Freitas, BT¹; Ducart, DF¹
¹Institute of Geosciences, University of Campinas (UNICAMP)

Determining the surface distribution of minerals and elements in drill cores has become fundamental in hydrocarbon reservoir characterization. Indeed, the implementation of non-destructive techniques to produce high spatial resolution mineralogical and compositional maps of core sections or individual samples has proved invaluable for the oil industry. Among the currently available methods dedicated to creating these maps, those capable of performing pixel-by-pixel analysis through rock surface scans, such as hyperspectral imaging and micro-X-ray fluorescence (µXR), significantly stand out.

However, despite their numerous advantages, including the fastness of data acquisition and remarkable precision, these techniques require relatively expensive analytical equipment, whose transport between locations is considerably complex. Hence, the application of portable devices to establish the mineralogical and compositional variation of rock materials represents an excellent and cost-effective alternative, given their consistent results and ease of operation.

In this work, we integrate single-point spectral and scattered XRF data obtained through handheld devices (Spectral Evolution and Tracer 5g, respectively) to generate chemical element maps of dolostone samples from the Vazante-Paracatu region, in the northwestern portion of the Minas Gerais state (Brazil). This region is a well-known mining district that hosts a broad collection of Mississippi Valley-Type (MVT) Zn-Pb deposits, which have been extensively affected by epigenetic hydrothermal alteration.

The hydrothermal alteration patterns shown by the dolomitic rocks in these deposits – including multiple episodes of dissolution, dolomitization, and silicification – share similarities with those recognized in carbonate reservoir rocks from the Brazilian pre-salt, within the Santos Basin. Analytical procedures conducted in this study initially involved the acquisition of spectral data in the VNIR-SWIR range (350 to 2500 nm) from specific spots of interest in the rock samples, which were used to retrieve local mineralogical and qualitative compositional information. XRF scans were then performed in the exact spots previously determined for the spectral analyses. Compositional maps were finally created using a random forest machine-learning approach and RBG images of the sample surfaces as inputs.

This approach handles the generation of these maps as a regression problem, solved by modeling a data set known as predictor variables, which, in our case, are defined by the XRF scanned points. Therefore, the unknown amounts of each chemical element for the remaining pixels in the rock image characterize the output variable to be estimated by the model. The preliminary outcomes of this work demonstrate that, even with a small number of analytical XRF data spots, the resulting compositional maps provide satisfactory results and can effectively distinguish between distinct dolomite generations with variable compositions, showing good correlation with the spectral information. The spectral-mineralogical and compositional characterization of rocks subjected to hydrothermal process potentially analogous to those registered in the pre-salt, as conducted in this work, could aid in understanding the impacts caused by hydrothermal activity in the porosity, permeability and heterogeneity of petroleum reservoirs.