Quantitative estimation of reservoir parameters from seismic observations is a primary objective of reservoir characterization. Accurate prediction of reservoir elastic and petrophysical parameters is of great significance for reservoir evaluation, sweet-spot identification, and subsequent development deployment. However, conventional stepwise seismic inversion methods are prone to error accumulation during the propagation from elastic-parameter inversion to petrophysical-parameters prediction. Although various simultaneous inversion methods have been proposed, deep learning-based approaches that simultaneously estimate elastic and petrophysical properties while incorporating rock physics constraints and seismic forward consistency remain relatively limited. We propose an intelligent multi-parameter joint inversion method (MPJ-Inv) for seismic reservoirs to jointly predict P-wave velocity, S-wave velocity, density, porosity, gas saturation, and shale volume. Specifically, a rock-physics model that accounts for stiff and soft pores was introduced, and a forward operator from petrophysical parameters to seismic data was constructed by combining this model with a linearized approximation of the Zoeppritz equation. In terms of network architecture, an end-to-end joint inversion network composed of an elastic-parameters prediction module and a petrophysical-parameters prediction module was developed. After the elastic-parameters prediction module obtained the elastic parameters, they were used as guiding information and combined with the initial petrophysical model to simultaneously predict petrophysical parameters. During network training, a semi-supervised strategy was adopted. At well locations, well-log labels were used to monitor both elastic and petrophysical parameters; at non-well locations, synthetic seismic data were generated using the forward operator and compared with the observed seismic records. In addition, a rock-physics consistency constraint was introduced to ensure agreement between the elastic and petrophysical parameters predicted by the network, thereby reducing the ill-posedness of the inversion to some extent. Tests on a synthetic seismic dataset and a tight sandstone gas reservoir dataset from Southwest China showed that, compared with conventional sequential inversion, the proposed method achieved higher accuracy, providing an effective reference for joint prediction of sandstone reservoirs.

Seismic impedance inversion is a fundamental task in exploration geophysics. With the development of artificial intelligence, deep learning has been increasingly adopted for post-stack seismic impedance inversion. This paper proposes a frequency-aware deep hybrid framework that integrates continuous wavelet transform (CWT) representations, a transformer encoder, and atrous spatial pyramid pooling (ASPP) to improve both impedance fidelity and resolution. The CWT provides time–frequency components of seismic traces; the transformer aggregates non-local context to stabilize structural consistency; and ASPP captures local multi-scale features. In addition, we introduce a frequency-aware loss that rebalances optimization toward components that are more difficult to reconstruct, improving detail recovery in thin beds. Network training used a semi-supervised strategy. At well locations, well-log impedance values supervised the network prediction; at non-well locations, synthetic seismic data were generated by seismic forward modeling and compared with the observed seismic records. Validation on the Marmousi II model and a field dataset indicated that the proposed method achieves consistently higher accuracy than representative deep-learning baselines and a commercial post-stack inversion workflow.

Due to the limitations of seismic data resolution, conventional inversion methods are commonly limited by the band-limited nature of seismic data, noise interference, and insufficient resolution for subtle geological targets. This study proposes a continuous wavelet transform (CWT)-assisted multiscale stepwise learning (MSL) workflow for seismic impedance inversion. The proposed workflow transforms seismic traces into time–frequency representations using CWT and progressively learns different frequency components through convolutional modules with different kernel sizes. A seismic forward-modeling constraint is further introduced to improve the consistency between the predicted impedance and the observed seismic response at non-well locations. The network takes a single seismic trace represented by CWT coefficients in the 1–80 Hz frequency range and a corresponding low-frequency initial impedance model as inputs, and outputs the acoustic-impedance trace at each time sample. The main methodological novelty is the progressive learning of different frequency components using frequency-dependent convolutional kernels. The method is evaluated on a 3D seismic dataset from the Sichuan Basin, China, using leave-one-well-out validation involving seven wells. In addition, we provide detailed ablation experiments and quantitative analysis. Across all seven blind-well tests, the proposed method consistently achieves the highest mean correlation coefficient (0.8509) and the lowest mean MSE (0.0636) among the compared methods. These results indicate that the proposed workflow can effectively integrate multiscale seismic information and provide effective support for seismic impedance inversion.