Development of a Machine‑Learning Predictive Model for Radiogenic Heat Production in Nigerian Basin Rocks

Abstract Radiogenic heat production (RHP) is a key control on the lithospheric thermal structure, yet direct measurements are sparse and unevenly distributed. We assemble 1,232 published radiometric measurements from six Nigerian geological provinces—encompassing sedimentary, metamorphic, and igneous terrains—and use potassium, uranium, and thorium concentrations as predictors in a suite of machine learning algorithms, including generalized linear models (GLM), support vector regression (SVR), decision tree regression (DTR), gradient-boosted regression trees (GBRT), random forests (RFR), extreme gradient boosting (XGBR), and a voting-based ensemble model (EM). Models were trained on 60% of the data and tested on the remaining 40%. Performance was assessed via RMSE, MAE, and R². The EM achieved the best balance of accuracy and generalization (R² = 0.98, MAE = 0.03 µW m⁻³, RMSE = 0.22 µW m⁻³), outperforming classical RHP formulae and single-algorithm predictors. We derive a new empirical RHP equation as a function of rock dry density and radionuclide concentrations and generate a high-resolution RHP map for the Nigerian Basin. This dataset reveals coherent geothermal anomaly patterns with greater spatial completeness than previous models. Our ML framework offers a robust approach for RHP estimation in data-poor regions.