ABSTRACT Pre‐stack seismic inversion, which estimates subsurface elastic parameters ( V p , V s and ρ ) from amplitude‐variation‐with‐offset (AVO) data, remains challenged by ill‐posedness and limited well control. Although deep learning (DL) offers data‐driven non‐linear mapping, its generalization is often constrained by scarce labels and physical inconsistency. This study introduces a semi‐supervised DL framework that integrates a time–frequency joint convolutional neural network (TF‐CNN) with physics‐guided constraints. The framework uniquely combines (1) a time–frequency hybrid attention mechanism that adaptively weights seismic features across domains; (2) exact Zoeppritz equation forward modelling to generate physically consistent pseudo‐labels for semi‐supervised learning and (3) a low‐frequency structural prior to stabilize long‐wavelength components. Validated on synthetic (Marmousi2) and field datasets, the method demonstrates enhanced accuracy, improved noise robustness and tighter uncertainty quantification compared to conventional AVO inversion and standard convolutional neural network (CNN) baselines. Quantitative metrics confirm significant gains in parameter accuracy and uncertainty reliability. The proposed approach effectively mitigates ill‐posedness through multi‐constraint integration, offering a practical and interpretable tool for reservoir characterization in label‐scarce settings.
Yong et al. (Wed,) studied this question.