Physics-Informed Machine Learning Framework for Sulphide Mineralization Mapping in Maitengwe Greenstone Belt Northeastern Botswana

Vae Onalethata¹, Boniface Kgosidintsi¹, Bokani Nthaba¹, Elisha M. Shemang¹, Abid Yahya², Mohamed Yasin Abdul Salam^3,*, Yar Muhammad⁴, Enerst Edozie⁵, Asiimwe Eva⁵
1 Department of Geology and Geological Engineering, Botswana International University of Science and Technology, Palapye, Botswana
2 Department of Electrical and Communications Systems Engineering, Botswana International University of Science and Technology, Palapye, Botswana
3 Department of Chemical, Materials & Metallurgical Engineering, Botswana International University of Science and Technology, Palapye, Botswana
4 School of Computer Science and Engineering, Beihang University, Beijing, China
5 Department of Electrical, Telecommunication and Computer Engineering, Kampala International University, Kampala, Uganda
* Corresponding Author: Mohamed Yasin Abdul Salam. Email: email

Computer Modeling in Engineering & Sciences https://doi.org/10.32604/cmes.2026.084977

Received 03 May 2026; Accepted 08 July 2026; Published online 07 September 2026

Download PDF

Abstract

Conventional geophysical inversion approaches typically have a limited capacity to map disseminated sulphide mineralization in complex greenstone belts. Here, we describe GeoPhysML, a physics-informed machine-learning approach that combines aeromagnetic and IP–ERI data to map sulphide mineralization in the Maitengwe Greenstone Belt, Botswana. Labels constrained by boreholes MTW1 and MTW2, including the mineralised intervals (MTW1: 110–160 m; MTW2: 118–153 m), produced 247 labelled grid cells (78 positive and 169 negative) that were split using a 70/15/15 train/validation/test split with spatial blocking. Borehole MTW4 was reserved exclusively for independent validation. GeoPhysML uses a three-hidden-layer neural network with auxiliary resistivity and chargeability outputs regularized by Laplacian smoothness and gradient-consistency regularization. The uncertainty of our predictions was calculated using Monte Carlo dropout with 100 draws. The model achieved 92.10% accuracy, 89.80% corrected F1-score, and 0.96 ROC–AUC, improving accuracy by 2.60 percentage points and F1-score by 2.93 percentage points over XGBoost. Probability-based error metrics produced MAE = 0.042, MSE = 0.0031, and RMSE = 0.056 on the evaluated set. The predicted high-probability zones follow the dominant NE–SW structural trend and coincide spatially with the mineralized interval in borehole MTW4. These results indicate that the proposed framework improves probabilistic target ranking within the surveyed corridor by combining inversion-derived physical fields, structural attributes, and borehole control in a unified mineralization mapping workflow.