Direction-Aware Masked Pretraining for 3D Seismic Representation Learning and Transfer to Cross-Area Acoustic Impedance Inversion

Abstract

Large archives of unlabeled three-dimensional seismic data offer opportunities for self-supervised representation learning and subsequent transfer to acoustic impedance inversion. However, conventional masked pretraining often treats three axes equivalently, overlooking differences between lateral reflector structure and vertical waveform characteristics. We propose a direction-aware masked autoencoder for three-dimensional post-stack seismic data, combining anisotropic tokenization, direction-aware representation, trace-aligned tube masking, and reconstruction constraints designed for reflector continuity and waveform characteristics. We evaluate reconstruction quality and downstream transferability using field data through masked reconstruction and cross-area acoustic impedance inversion. Reconstruction is more sensitive to lateral token resolution than to moderate changes in vertical patch length. Within the evaluated configurations, increasing encoder capacity does not fully compensate for reconstruction fidelity loss associated with coarser tokenization. Preferred token scales and masking strategies differ between reconstruction and inversion, indicating that reconstruction fidelity alone is not a reliable indicator of transferability. For cross-area inversion, the pretrained model is fine-tuned in the source area and applied to the target area without further parameter updates. With limited target-area well control, the proposed framework reduces normalized root-mean-square error by 20.8% relative to a pretrained conventional masked autoencoder across eight target-area test wells under matched tokenization and downstream settings. These results demonstrate the value of direction-aware masked pretraining for field seismic inversion and show that token scale and masking strategy should be selected according to downstream-task requirements rather than reconstruction accuracy alone.