Physics-aware Bayesian Vision Transformer enables structure-level uncertainty in seismic inversion

Abstract Seismic inversion methods based on traditional frameworks face fundamental challenges in accurately characterizing spatial structure and quantifying model reliability. Here, we introduce a next-generation approach, systematically evolving from a convolutional physics-informed neural network (PINN), to a Bayesian PINN (BPINN) with uncertainty modeling, and culminating in a Bayesian Physics-Informed Vision Transformer (BPI-ViT) architecture that enables structure-level uncertainty quantification. Consistent evaluation is performed on the Marmousi2 benchmark and validated on field-scale CO₂ EOR monitoring data. The BPI-ViT framework integrates structure-aware self-attention and Bayesian inference, effecting a transition from pixel-level optimization to structural collaboration. Empirical results reveal that BPI-ViT outperforms previous methods in target horizon recovery, fault and anomaly detection, spatial continuity, and uncertainty quantification. This study establishes a structural-intelligent paradigm, advancing seismic inversion beyond error minimization towards structure-aware, reliable, and cognitively informed modeling, and provides a foundation for future multi-physics and complex geological inversion applications.