Paper submitted to IMAGE 2026, by Ben Lasscock, Meher Gajula, Keyla Gonzalez, Brian Michell, Alejandro Valenciano (TGS).
Abstract
Fusing spatially dense, band-limited seismic data with sparse, high-resolution well logs remains a fundamental challenge in quantitative interpretation. We present a Cross-Modal Masked Autoencoder (CM-MAE), a self-supervised foundation model that learns a shared latent representation of 3D seismic and 1D well-log data. CM-MAE uses a single Vision Transformer encoder-decoder architecture to process a joint sequence of seismic and well tokens, enabling cross-modal reconstruction within a unified framework. The model is pretrained on a heterogeneous mixture of paired seismic-well, seismic-only, and well-only samples, allowing it to learn relationships across modalities. Ablation experiments show that the learned transfer is strongly asymmetric: seismic context substantially improves well-log reconstruction, reducing well reconstruction error by 52%, whereas well-log context provides negligible benefit for seismic reconstruction. Further analysis of well reconstruction reveals that it relies on seismic context and aligns with reflector-scale structures. We also demonstrate the practical value of the learned representation by supervised fine-tuning the pretrained model for seismic-only well reconstruction, where it predicts spatially coherent pseudo-logs directly from seismic data. Collectively, these ablation and downstream prediction results show that early-fusion self-supervised pretraining provides an effective framework for seismic-well integration and seismic-guided prediction of subsurface properties.
Read the full article here.