Improving P-wave velocity prediction of pre-salt carbonates (Santos Basin, Brazil) using thin-section–derived pore geometry, petrophysics, and mineralogy

Several studies have modeled relationships between elastic properties and geological attributes, often combining empirical models with pore geometry. However, the specific effects of pore types and their interactions with other geometric features and porosity on P-wave velocity (Vp) remain unclear. To address this, we applied multiple linear regression (MLR) and Elastic Net machine learning to predict Vp using key variables: plug porosity, mineral compressibility (Kmin), pore geometry (perimeter-to-area ratio and pore radius), and predominant pore types (vuggy, interparticle, and intraparticle). Digital image analysis of scanned thin sections provided data on pore geometry and types. Spearman correlation showed that interparticle pores correlate with higher velocities and intraparticle with lower, and vuggy pores had a slight positive trend. Initial regression with porosity and Kmin yielded weak correlation (coefficient of determination [R2] = 0.323, adjusted R2 = 0.188). Adding pore geometry and pore type greatly improved the model (R2 = 0.934). Elastic Net performed well during training (R2 = 0.879), managing multicollinearity effectively, but showed reduced performance in blind testing (R2 = 0.612). Its lambda (0.378) balanced accuracy and error penalties moderately. Whereas Elastic Net suits data sets with many predictors or strong collinearity, MLR demonstrated superior predictive power and interpretability in this case. These results highlight the importance of selecting models based on data characteristics, with MLR emerging as the more effective method for predicting Vp here. Predicting Vp is crucial for reservoir characterization, directly impacting lithological interpretation and petrophysical evaluation.