Machine learning-driven porosity prediction: A case study from the Cretaceous Mata Series, Canterbury Basin, New Zealand

Muhammad Ubaid Umar, Shenghe Wu, S.M. Talha Qadri, Wakeel Hussain, Du Wei, Shaukat Khan,
Machine learning-driven porosity prediction: A case study from the Cretaceous Mata Series, Canterbury Basin, New Zealand,
Physics and Chemistry of the Earth, Parts A/B/C,
Volume 141, Part 2,
2025,
104159,
ISSN 1474-7065,
https://doi.org/10.1016/j.pce.2025.104159.
(https://www.sciencedirect.com/science/article/pii/S1474706525003092)
Abstract: Predicting and verifying porosity measurements in the offshore section of the Cretaceous Mata Series in the Canterbury Basin is challenging. Key factors include the thick Mata Series hosting a number of reservoir formations with extremely heterogeneous lithologies, depositional environments, and highly variable thicknesses and depths of potential reservoir formations. A new approach using machine learning techniques can efficiently predict porosity in complex settings of the Canterbury Basin. This study employed several efficient machine learning methods including Multiple Linear Regression (MLR), Artificial Neural Network (ANN), Long Short-Term Memory (LSTM), and Extreme Gradient Boosting (XGBoost) to predict porosity in reservoir intervals. Five offshore wells were selected, and four well logs (GR, DT, NPHI, RHOB) were used as inputs. Data preparation included training and testing set creation, followed by porosity evaluation using closeness of fit (Cfit) curves, with coefficient of determination (R2) and root mean square error (RMSE) as the primary evaluation metrics. Here, Cfit curves represent the direct overlay of measured and predicted porosities along the depth through a visual measure of how closely the two porosity curves match. Among the models tested, XGBoost demonstrated the best performance with an R2 of 0.9976 and an RMSE of 0.0028, while MLR showed the weakest performance (R2 = 0.9384, RMSE = 0.0143). ANN and LSTM also performed well, achieving R2 values of 0.9751 and 0.9876, respectively. All models yielded R2 values above 0.90, indicating a high predictive accuracy. The results confirm the feasibility of applying machine learning models for porosity prediction in this offshore region, marking the first such application in the Mata Series. These models also hold promise for predicting other petrophysical properties in intervals where log data is incomplete.
Keywords: Machine learning; Porosity prediction; Mata series; Canterbury basin; New Zealand