Chemical weathering of mafic and ultramafic rocks plays a crucial role in natural \(\:{\text{C}\text{O}}_{2}\) sequestration through mineral carbonation processes. Quantifying the weathering potential of these lithologies is therefore essential for evaluating their contribution to enhanced weathering strategies and long-term climate regulation. In this study, major oxide geochemical data (172 data sets) of mafic and ultramafic rocks were collected from 15 locations in India based on existing literature. Three machine learning (ML) based models such as SVR, GMDH and GPR, were developed to anticipate WI. Model performance was evaluated using grouped statistical, graphical, and external validation techniques to assess accuracy, robustness, and generalization capability. Among the developed models, GPR demonstrated superior performance in both training and testing stages, achieving \(\:{\text{R}}^{2}=0.8334\) and RMSE = 0.0846 during training, and \(\:{\text{R}}^{2}=0.7869\) and RMSE = 0.1102 during testing. Interpretable ML analyses using SHAP revealed that \(\:{\text{A}\text{l}}_{2}{\text{O}}_{3}\) and \(\:{\text{N}\text{a}}_{2}\text{O}\) are the dominant positive contributors to WI, while \(\:\text{C}\text{a}\text{O}\) and \(\:{\text{K}}_{2}\text{O}\) exert moderate influence. Furthermore, PDPE analysis highlighted strong nonlinear effects of \(\:{\text{A}\text{l}}_{2}{\text{O}}_{3}\) and \(\:{\text{N}\text{a}}_{2}\text{O}\) on WI, with complex interactive behaviour observed for \(\:\text{C}\text{a}\text{O}\) . The proposed interpretable ML framework contributes a robust tool for assessing the \(\:{\:\text{C}\text{O}}_{2}\) sequestration capability of mafic and ultramafic rocks in enhanced weathering applications. It also provides useful insights into the geochemical constraints determining weathering intensity.
