Seismic inversion is vital for reservoir characterization but faces significant challenges in complex fluvial-deltaic systems due to strong heterogeneity and thin-bedded formations. Current methods, including convolution-based, geostatistical, and artificial neural network (ANN) approaches, are often limited by wavelet stationarity assumptions, spatial uncertainty, and physical implausibility. This study develops a novel artificial intelligence (AI) seismic inversion algorithm that integrates a convolutional physical model with data-driven learning to overcome these drawbacks. The proposed physics-guided hybrid model employs a multi-wavelet inversion framework, incorporating 8–10 spatially variable wavelets per inversion cell to account for lateral wavelet variability. These physically constrained inversion candidates are then intelligently fused using a computationally efficient neural network, which maintains a 3.4% training error and 4.7% validation accuracy. This integrated approach achieves remarkable improvements: a 30% enhancement in vertical resolution enabling 1–3m thin-bed detection, a 40% improvement in lateral continuity (with correlation coefficients increasing from 0.85), and 70% better noise suppression. Application in a complex fluvial-deltaic system covering 7.2 km ² with 80 wells confirmed the method’s robustness, delivering over 80% accuracy in sandbody prediction while significantly reducing geologically implausible results.