Optimization of artificial neural network models using genetic algorithm for accurate estimation of water saturation

Abstract Accurate estimation of water saturation ( S w ) is a critical task in reservoir engineering, as it directly influences reservoir characterization, hydrocarbon volume estimation, and decision‐making in field development. Traditional empirical and deterministic models often fail to provide the required accuracy due to the nonlinear and complex relationships between petrophysical parameters. To address this challenge, this study investigates the application of advanced artificial neural network (ANN) models optimized by genetic algorithm (GA) for S w prediction. A total of 42 924 data samples were collected from well log records, containing 13 petrophysical input parameters: depth, total porosity (PHIT), hydrocarbon total neutron porosity (HTNP), true formation resistivity (RT), flushed zone resistivity (RXO), computed gamma ray (CGR), spectral gamma ray (SGR), hole caliper (HCAL), compressional slowness (DTCO), photoelectric factor (PEFZ), Bulk density (RHOZ), Gamma ray density (RHGA), and temperature. To enhance data quality, 14.5% of the samples identified as outliers using a Gaussian distribution‐based method (the 3‐sigma rule) were removed. Five ANN architectures—convolutional neural network (CNN), recurrent neural network (RNN), feedforward neural network (FNN), multi‐layer perceptron (MLP), and radial basis function network (RBFN)—were developed. GA was used to optimize the hyperparameters of each model. Model performance was evaluated using coefficient of determination ( R 2 ) values for training, validation, and test datasets, while prediction errors were assessed using mean absolute error ( MAE ), mean squared error ( MSE ), and root mean squared error ( RMSE ). Among all models, CNN achieved the best results, with R 2 values of 0.9843 (train), 0.9645 (validation), and 0.9637 (test), and RMSE of 0.0019 (train) and 0.0029 (test). These results confirm the effectiveness of CNN with GA optimization in capturing complex relationships within petrophysical data for accurate S w prediction.