A Well Log-Driven Approach for Mechanofacies Characterization via Unsupervised Learning

Abstract Recognizing the importance of well logging, especially sonic velocity (DT), and density (RHOB) logs, for estimating rock mechanical properties, this study evaluates the applicability of unsupervised machine learning to classify mechanofacies. To achieve this goal, a multivariate cluster-analysis workflow is proposed, using only geomechanical parameters derived from well logs as input data. This approach seeks to reduce the subjectivity associated with the mechanofacies concept and establish an easily applicable methodology capable of providing fast and consistent results, even when data availability imposes significant constraints. Accordingly, dynamic elastic properties—Young’s modulus ( E d ), shear modulus ( G d ), bulk modulus ( K d ), Poisson’s ratio ( ν )—and unconfined compressive strength (UCS) were calculated through empirical correlations, and their uncertainties were incorporated into the geomechanical model proposed. The analysis was performed on five wells in the south-central Paraná Basin, southern Brazil. The investigated successions are thick and lithologically diverse, comprising clastic and carbonate sedimentary rocks, subvolcanic igneous rocks, and, in some wells, granitic gneisses from the crystalline basement. Cluster analyses identified five clusters in each well, allowing synthesis of rock geomechanical behavior into five mechanofacies (MF1, MF2, MF3, MF4, and MF5) that show lateral continuity across the study area. MF1 and MF2, representing the least resistant rocks, occur mainly near the top of the wells, whereas MF4—with high E d , G d , K d , and UCS—is more common at depth. MF3 appears throughout multiple stratigraphic levels, indicating strong vertical heterogeneity that affects drilling and stimulation performance. MF5 is lithologically controlled and dominated by high-strength igneous rocks. Despite uncertainties and spatial limitations, the model results highlight the potential of unsupervised learning techniques for identifying mechanofacies in data-limited contexts, revealing patterns that can support future drilling and reservoir stimulation projects. A small adjustment is proposed: Mário Palmério University Center (UNIFUCAMP), Monte Carmelo, Minas Gerais, Brazil.