Leveraging Machine Learning and X-Ray Fluorescence Spectrometry for Elemental Analysis of Pozzolanic Materials

Abstract Pozzolanic materials, known for their superior performance in enhancing the strength and integrity of well cements along with their significantly reduced carbon footprint, have been receiving much attention. The elemental characterization of pozzolanic materials is crucial because it affects the reactivity of cementitious fluids and thus the overall efficiency of a cementing job. This study investigates the elemental analysis correlation between inductively coupled plasma optical emission spectroscopy (ICP-OES) and X-ray fluorescence (XRF) spectrometry techniques using machine learning (ML) algorithms. ICP-OES is known for its high accuracy and low detection limit, but it can be labor-intensive and time-consuming because of the tedious sample preparation involved, which also raises chemical safety concerns. In contrast, XRF requires minimal sample preparation and offers quick analysis. However, the XRF results are subject to matrix effects. The goal of this study is to harness the power of ML to recognize and reduce the matrix interference in XRF, enabling the substitution of ICP-OES with XRF in practical well cementing applications. A comprehensive dataset of over 100 unique pozzolanic materials, sourced from various suppliers worldwide, was compiled using ICP-OES, XRF, and other material characterization techniques. A family of ML regression models was optimized, primarily focusing on the simple and efficient principal component regression (PCR) and partial least squares (PLS) algorithms, while also exploring other linear, nonlinear, kernel-based, and decision tree-based ML models. These models were trained using full XRF spectra (both raw and after preprocessing), with ICP-OES measurements serving as the ground truth. Model predictions on an independent test set were critically evaluated and compared, resulting in the development of an optimal ML workflow. Rigorous reproducibility analysis was also conducted for both analytical techniques. The ML models demonstrated prediction errors of less than 5% (range-normalized root mean square error) for the 12 main elements of interest when applied to the XRF spectra. This suggests that ML-enhanced XRF spectrometry can effectively replace ICP-OES, significantly reducing the labor and time required for the elemental analysis of pozzolanic materials. This study, pioneering the use of ML to enhance the elemental analysis of pozzolanic materials, offers a robust framework for future practical applications of XRF material characterization in the oilfield industry and beyond. The proposed approach will lead to informed decision-making regarding the formulation of pozzolan-based well cementing slurries. Leveraging advanced analytical techniques to ensure their effective use in cementing applications aligns with the industry's ambition to use environmentally benign materials.