CarbonFormer3D: Enhanced CO2 Plume Segmentation from Time-Lapse Seismic Data

Summary

Time-lapse (4D) seismic monitoring is a critical tool for tracking the migration and long-term containment of injected carbon dioxide (CO2) in subsurface geological storage reservoirs. Accurate delineation of CO2 plume evolution is essential for assessing storage performance, identifying unexpected migration pathways, and supporting regulatory reporting in carbon capture and storage (CCS) projects. However, conventional seismic interpretation workflows rely heavily on manual delineation of plume extents, which is time-consuming, subjective, and difficult to scale across large volumetric data sets and multiple monitoring vintages. These challenges are further complicated by acquisition-related noise, survey-to-survey repeatability issues, and processing artifacts, which can hinder consistent plume tracking over time.

In this study, we introduce CarbonFormer3D, a transformer-convolutional neural network (CNN) hybrid network designed for full-volume 3D segmentation of CO2 plumes from time-lapse seismic data. The architecture combines a patch-based transformer encoder, which captures long-range spatial context and global plume connectivity, with a lightweight convolutional decoder and multiscale skip connections that preserve fine-grained spatial detail. This hybrid design aims to improve segmentation accuracy by integrating global and local information, which is crucial for resolving complex plume geometries and layered migration patterns. CarbonFormer3D is evaluated using the publicly available Sleipner CO2 storage benchmark data set across multiple seismic vintages. The results demonstrate consistent improvements over conventional 3D U-Net baselines, including more accurate delineation of plume boundaries, improved preservation of fine-scale features, and robust performance across time-lapse surveys. At its core, CarbonFormer3D provides a computationally efficient framework that combines global context modeling with local spatial reconstruction, enabling scalable and reliable segmentation of complex CO2 plume structures. This approach supports more accurate monitoring and assessment of CO2 storage performance in operational CCS settings.