Paper submitted to IMAGE 2026, by Vi Ly, Altay Sansal, Ben Lasscock, Alejandro Valenciano (TGS).
Abstract
Seismic foundation models offer a scalable method to improve interpretation tasks through self-supervised pretraining on large unlabeled seismic datasets. In this study, we analyze how scaling impacts facies segmentation using the Parihaka facies benchmark, focusing on three factors: pretraining data scale, parameter-efficient adaptation, and model dimensionality. We show that increasing pretraining scale slightly improves 2D seismic foundation models and enhances downstream segmentation compared to previous SFM baselines. Low-Rank Adaptation (LoRA) provides additional benefits by updating only a small subset of model parameters. However, 2D models are inherently limited because they interpret geology through isolated slices, whereas facies architecture and structural continuity are fundamentally three-dimensional. The most substantial improvement results from extending the foundation model to 3D, enabling direct incorporation of volumetric context into representation learning and segmentation. On the Parihaka facies benchmark, the top 2D model attains an mIoU of 0.8115, while the 3D SeisFM reaches 0.9005. These results demonstrate that larger-scale pretraining and efficient adaptation improve seismic foundation models, but 3D representation learning remains essential for geologically consistent facies segmentation.
Read the full article here.