Abstract Peridotites record the compositional evolution of the lithospheric and asthenospheric mantle across plume‐related lithospheres, convergent margins, and rift‐related volcanic environments; however, their tectonomagmatic affinities are commonly difficult to resolve using traditional low‐dimensional geochemical discrimination diagrams. In this study, we develop an explainable machine‐learning framework that integrates Compositional Data Analysis (CoDA), isometric log‐ratio transformation, and gradient‐boosted decision tree classification to explore multivariate whole‐rock geochemical patterns in a curated global data set of peridotites ( n = 640). Rather than aiming for deterministic assignment to tectonic end‐members, the approach operates in probability space, allowing samples to express mixed or transitional affinities that better reflect natural mantle processes. Model interpretability is achieved using SHapley Additive exPlanations (SHAP), which provides both global rankings of influential variables and local explanations of individual predictions. The results show that rare earth elements (REE) dominate tectonic discrimination, with middle rare earth elements (REEs) characterizing rift‐related volcanic affinities, light REEs driving plume‐related lithosphere signatures, and convergent margin peridotites exhibiting lower‐amplitude, distributed contributions consistent with metasomatized mantle wedge processes. Application of the trained framework to Archean greenstone peridotites reveals preferential alignment with plume‐related and rift‐like REE systematics, while convergent margin—like signatures represent a distinct but subordinate mode. Overall, this study demonstrates that explainable machine learning, when grounded in compositional data principles, provides a transparent and geologically meaningful tool for interrogating complex mantle geochemical data sets and for evaluating tectonomagmatic affinities without imposing rigid modern analogs.
Roy et al. (2026) studied this question.