Machine Learning Approaches for Compositional Fluid Analysis in Logging While Drilling Using Near-Infrared Data

Accurate quantification of light hydrocarbon components (C1–C6) in downhole fluids is critical for reservoir evaluation and production planning. This study presents a comparative analysis of nonnegative least squares (NNLS) regression, indirect hard modeling (IHM) and convolutional neural networks (CNNs) for compositional analysis of near-infrared (NIR) spectra of downhole fluids. Quantification of mass fractions of hydrocarbon fractions from methane (C1) over volatile alkanes (C2-C5) to liquid and heavy alkanes(C6+), and of carbon dioxide (CO2) is particularly challenging due to overlapping spectral features associated with aliphatic hydrocarbons. The predictive performance of all methods is evaluated using a curated dataset of NIR spectra from pure and mixed fluid samples. The NNLS model attempts direct linear spectral unmixing; IHM attempts spectral unmixing via a close fit of the measured spectrum; while CNNs are trained end-to-end on augmented spectra mimicking phenomena expected during logging-while-drilling acquisitions. Results indicate that all three methodologies exhibit good accuracy in predicting pure components and mixtures. IHM developed within this works as a mixture-centric model, shows better generalization to complex mixtures while lacking precision on pure components. NNLS and CNNs correctly identify pure components with root mean squared error (RMSE) being well below 5%, with similar accuracy on fluid mixtures. Hence, both NNLS and CNN are promising approaches for real-time, in-situ compositional analysis, offering a scalable solution for embedded downhole applications.