Machine Learning Algorithms to Predict Source and Reservoir Zones of Abu Roash Formation Using Spectral Gamma‐Ray Logs, Khalda Oil Field, Northern Western Desert, Egypt

Abstract

ABSTRACT Accurate estimation of delineation of radioactive anomalies deviation (DRAD) is critical for characterizing subsurface radioactive element distributions commonly associated with hydrocarbon accumulations. However, DRAD estimation traditionally relies on natural gamma‐ray spectrometry measurements, which are costly, operationally demanding, and not routinely acquired in many wells. This study proposes a physics‐informed machine learning model that aims to predict DRAD using traditional well logs while embedding geological constraints to ensure physically consistent predictions. Conventional logs, including gamma ray, density, neutron porosity, resistivity, and photoelectric factor from the Khalda oil field (Shushan Basin, Western Desert, Egypt), are subjected to systematic preprocessing involving outlier removal and feature standardization. Two predictive models are implemented: random forest (RF) as a data‐driven baseline and a physics‐informed neural network (PINN) incorporating radioactive element relationships within the optimization process. Training of the models is carried out using well KH‐24 with an 80–10–10 split for training, testing, and validation sets, whereas prediction stability is carried out using a three‐point moving average smoothing technique after prediction. The PINN demonstrated superior predictive capability, yielding R 2 values of 0.90, 0.82, and 0.85 for training, testing, and validation datasets, respectively, improving to 0.92, 0.90, and 0.89 after smoothing. The generalization capability of the models is validated with blind wells KH‐17 and KH‐21, where predicted DRAD profiles successfully reproduced anomaly magnitudes and stratigraphic variability consistent with measured data. These results showed that the use of physics‐based learning is more stable in terms of prediction performance and geological reliability in comparison with empirical‐based models, providing a transferable and cost‐effective alternative for estimating DRAD values in wells lacking spectrometric measurements and supporting more efficient hydrocarbon exploration workflows.