The Geological Everything Model 3D (GEM) is an integrated generative architecture for Earth's subsurface analysis. While traditional subsurface analysis requires separate models for structural interpretation, stratigraphy, geological body segmentation, and feature modeling, GEM reframes all these tasks with prompted conditional inference, following a potential structural framework derived from subsurface imagery. This enables a shared inference mechanism that propagates user-provided prompts, such as well logs, masks, or structural sketches, along the inferred structural framework to produce geologically consistent outputs. A two-stage training process combining self-supervised representation learning on large-scale field seismic data and adversarial fine-tuning using mixed prompts and labels across various subsurface tasks achieves zero-shot generalization across heterogeneous prompt types without retraining on new tasks or data sources. It is widely applicable to diverse investigations and tasks, including Mars radar stratigraphy, structural interpretation of subduction zones, full seismic stratigraphy, geological body segmentation, and feature modeling.
