Summary Accurate estimation of total organic carbon (TOC) is fundamental for evaluating hydrocarbon source rock potential; however, laboratory-based geochemical measurements are often sparse, discontinuous, and costly. Although machine learning (ML) approaches enable continuous TOC prediction from well logs, most existing models are purely data-driven and lack physical consistency and uncertainty quantification. In this study, a Bayesian physics-informed neural network (B-PINN) framework is developed to predict TOC from conventional wireline logs while embedding the ΔlogR geochemical relationship as a soft physical constraint. The method is applied to Early Jurassic formations of the Mandawa Basin, southeast Tanzania, and benchmarked against group method of data handling (GMDH) and Gaussian process regression (GPR). The proposed B-PINN achieves testing performance of root mean square error (RMSE) = 0.371 and coefficient of determination (R2) = 0.9735, outperforming benchmark models. Bayesian inference via Hamiltonian Monte Carlo (HMC) enables calibrated uncertainty quantification with 95.8% empirical coverage for the 95% credible interval. The integration of physical consistency and probabilistic inference enhances prediction robustness in data-limited settings. The framework provides an uncertainty-aware and geologically consistent approach for TOC estimation in sedimentary basins.
Mulashani et al. (Wed,) studied this question.