Machine-learning ensemble modeling of radiogenic heat production in crustal rocks in Nigerian geological provinces

Abstract

Radiogenic heat production (RHP) is a fundamental control on lithospheric thermal structure, yet direct measurements remain sparse across many regions, including the Nigerian geological terranes. To improve spatial coverage, a compilation of 1,232 published radiometric measurements from six geological provinces—spanning sedimentary, metamorphic, and igneous settings—was assembled and analysed. Potassium (K-40), uranium (U-238), thorium (Th-232), and dry-rock density were used as predictors within a machine-learning (ML) framework comprising generalized linear models (GLM), support vector regression (SVR), decision tree regression (DTR), random forest regression (RFR), gradient boosting regression trees (GBRT), extreme gradient boosting regression (XGBR), and a voting ensemble model (VEM). Among the evaluated models, the VEM achieved the strongest predictive performance, with a coefficient of determination (R²) of 0.98, a mean absolute error (MAE) of 0.02 µW m⁻³, and a root mean square error (RMSE) of 0.16 µW m⁻³. These results surpass both traditional radiogenic heat-production equations and individual ML models by effectively capturing nonlinear and lithology-dependent relationships among the predictors. Based on the ensemble outputs, a new empirical RHP equation was derived, enabling accurate heat-production estimation directly from radionuclide concentrations and dry-rock density. Application of the ensemble model produced a spatially continuous RHP map across the Nigerian geological provinces and sedimentary basins, revealing coherent geothermal anomalies associated with major basement complexes and contrasting lower-heat sedimentary domains. Overall, the proposed workflow offers a robust and transferable approach for RHP estimation in geologically heterogeneous, data-limited regions, with direct relevance for geothermal exploration and crustal thermal modelling.