Analysis of Minerals in Drilling Fluids Using Synergy in X-Ray Fluorescence and Fourier Transform Infrared Spectra Realized with Machine Learning

Abstract

Abstract Analysis of solid contents in drilling muds is part of the API mud testing protocol performed daily by the mud engineer during drilling operations. This is carried out with the laborious manual retort analysis where high-gravity solids (HGS) and low-gravity solids (LGS) are detected. With the vision to improve the process and potential toward online measurement, monitoring of HGS and LGS in water-based mud (WBM) using two fast and reliable analytical techniques, X-ray fluorescence (XRF) spectrometry and Fourier transform infrared (FTIR)-attenuated total reflection (ATR) spectroscopy, have been evaluated. While information on total LGS content in a drilling fluid is useful, understanding the LGS minerology is also undoubtedly valuable, particularly when reactive clays are present and wellbore stability and integrity are at risk. In the present work we describe the dataset of XRF and FTIR-ATR spectra of a series of multicomponent water-based drilling fluids loaded with up to eight dispersed minerals. The minerals were barite (HGS) and LGS as clays (bentonite, illite, and kaolinite), carbonates (calcite and dolomite) and siliceous minerals (quartz and soda feldspar). Machine-learning (ML) analysis techniques such as partial least squares (PLS) and advanced modern PLS methods were applied to XRF and FTIR combined data to analyse these complex multicomponent samples. In this paper, a detailed discussion of the data workflows will cover the spectra preprocessing, ML algorithm selection and results on individual spectra, concatenated data, and data fusion. While quantification of barite (HGS) in drilling fluids with either XRF or FTIR was straightforward (the coefficient of determination, R2, at ~0.99), analysis of some LGS was more challenging. For example, the best soda feldspar and calcite PLS models based on FTIR spectra showed poor performance (R2 ~0.75 and ~0.48, respectively). Interestingly, PLS models based on different combinations of concatenated spectra (joined side-by-side) demonstrated even worse results compared to the single spectra models. Extensions of the PLS method designed to deal with multiblock datasets were then explored, and the sequential and orthogonalized PLS (SO-PLS) was found to be the best, realizing the synergy between spectrum types. SO-PLS models for calcite and soda feldspar gave R2 of 0.97 and 0.77, respectively. Detailed error analysis for all minerals is finally presented in the paper, with prediction errors varied from 1% for barite to 11% for soda feldspar. The two spectrum types, when smartly processed together using data fusion and ML techniques, demonstrate synergy, and provide significantly better mineral quantification accuracy, achieving a comprehensive minerology analysis of solids in drilling fluid.