Porosity is a crucial parameter in characterizing reservoir properties, with accurate prediction being essential for the exploration and development of oil and gas reserves. Traditional approaches to porosity prediction have predominantly relied on empirical formulas and linear relationships, which are often limited in scope and accuracy. To overcome these limitations, this study introduces a deep learning algorithm that combines long short-term memory (LSTM) neural networks and Extreme Gradient Boosting (XGBoost) using an integrated learning approach. The hybrid model, termed PSO-LSTM-XGB, employs the particle swarm optimization (PSO) algorithm to fine-tune hyperparameters, enhancing prediction accuracy. The study uses well logging data from the Yanchang Formation of the Changqing Oilfield in the Ordos Basin. These logs include acoustic time difference (AC), rock density (DEN), compensated neutron log (CNL), spontaneous potential (SP), resistivity (RT), and porosity (POR). In the data preprocessing stage, Pearson and Spearman correlation analyses are applied to assess the relationships between the variables. Four key features—RT, AC, DEN, and CNL—are selected as input features, while POR serves as output feature. A sensitivity analysis of PSO parameters is conducted to determine the optimal configuration, improving the efficiency of hyperparameter optimization. Experimental results demonstrate that the PSO-LSTM-XGB model significantly outperforms the benchmark LSTM-XGBoost model, achieving an R 2 of 0.891. Furthermore, the hybrid model exhibits superior performance compared to other machine learning and deep learning algorithms tested in this study. The study concludes that the PSO-LSTM-XGB provides superior accuracy in modeling porosity prediction. This method shows great potential for accurate porosity prediction in various reservoirs.
Zhang et al. (2026) studied this question.