Summary Accurate porosity prediction from well-logging data is vital for reservoir characterization and field development yet remains challenging due to noise and nonstationarity in the logging signals. In this paper, we propose an adaptive frequency-domain (AFD) dual-branch network (AFD-DBNet) to improve porosity prediction under such conditions by seamlessly integrating frequency-domain filtering with deep neural sequence modeling. AFD-DBNet introduces an AFD filtering module that dynamically retains the most informative spectral components via an energy-based threshold, effectively suppressing noise while preserving key trend information. Meanwhile, a dual-branch 1D convolutional neural network (CNN) separately processes the main frequency components and residual components to extract complementary trend and detail features, which are fused and fed into a bidirectional long short-term memory (BiLSTM) decoder to capture long-term temporal dependencies. This multiscale feature extraction allows the model to learn both global reservoir trends and fine-scale fluctuations in the porosity sequence. By combining adaptive spectral selection with sequence modeling, we build a unified forward pipeline from input windows to porosity outputs, while AFD provides a data-dependent filtering step embedded in the forward pass. Evaluated on well-log field data, the proposed model achieves consistently improved predictive accuracy, with coefficient of determination (R2) reaching 0.996, mean squared error (MSE) as low as 0.283, mean absolute error (MAE) of 0.352, and symmetric mean absolute percentage error (sMAPE) of only 9.48%. These metrics substantially outperform those of representative baseline models, demonstrating that the AFD approach and dual-branch architecture significantly enhance noise robustness and prediction accuracy.
Li et al. (Sun,) studied this question.