Interpretable intelligent gas-bearing reservoir prediction using time–frequency analysis and manifold-regularized semi-supervised GAN

As hydrocarbon exploration advances toward deep and complex reservoirs, the identification accuracy of traditional time–frequency (TF) analysis is constrained by strongly heterogeneous geological conditions. Concurrently, while deep learning has shown great potential, mainstream supervised models commonly face the dual challenges of scarce labeled samples and the lack of interpretability in their black-box decision-making processes. To address these challenges, this study proposes an innovative, intelligent prediction framework integrating high-precision TF analysis, manifold-regularized semi-supervised generative adversarial networks, and SHapley Additive exPlanations (SHAP) for interpretability. First, the Fourier-based synchrosqueezing transform was employed to extract two-dimensional TF features with superior energy concentration, effectively overcoming the resolution limits imposed by the Heisenberg uncertainty principle. Subsequently, the manifold-regularized semi-supervised generative adversarial network was developed. By incorporating manifold regularization constraints, the discriminator captures the intrinsic topological structure of large-scale unlabeled samples, effectively leveraging data geometry to significantly enhance generalization capability under sparse-label conditions. Finally, the SHAP method was utilized to conduct a post hoc interpretation. Experimental results on the Marmousi II model demonstrate a remarkable testing accuracy of 98.4%. In a real-world application to deep marine reservoirs in the Sichuan Basin, the framework achieved an 85.0% testing accuracy using only 5% labeled samples. Compared to baseline models, the semi-supervised strategy and manifold regularization contributed accuracy gains of 18.8% and 5.0%, respectively. SHAP analysis further confirms the model’s adaptive capability to extract geophysical features, enabling it to deconstruct the tuning-effect patterns in synthetic data and the low-frequency enhancement/high-frequency attenuation patterns in real data, respectively. This validation of geophysical consistency provides a theoretical foundation for the application of artificial intelligence in complex hydrocarbon exploration.