Geological
Everything Model

Scientific imaging requires the inference of spatial objects, geometric boundaries, and continuous properties from limited observations. Some tasks in medical imaging, remote sensing, and subsurface imaging share the challenges of missing observations and target ambiguity: high-dimensional data and large-scale training do not necessarily supply all the information needed to determine a target. Local annotations, domain interpretations, and independent measurements can provide additional target constraints. An important question is how these heterogeneous conditions can form representations jointly with observations, enabling one model to infer different targets and revise its predictions. We propose Geological Everything Model 3D (GEM), which explores unified three-dimensional interactive inference through target-conditioned representations. Object membership, geometric relationships, and measured values are encoded as local conditions and supplied with observations to a three-dimensional backbone, allowing conditions to participate in feature extraction from the input stage. Multitask joint training brings discrete object recognition and continuous field estimation into a common conditional prediction interface. An interactive workflow allows conditions to be accumulated, revised, and revisited. Adding, replacing, or withdrawing conditions updates candidate results while retaining valid historical constraints, enabling continued inference with fixed model parameters. Subsurface imaging serves as the principal validation setting, where complex three-dimensional structures, diverse targets, and sparse constraints provide a challenging test. Synthetic benchmarks, field surveys, and blind-well evaluations demonstrate applications to fault and horizon interpretation, geobody identification, stratigraphic coordinates, and physical property estimation. GEM outperforms the compared baselines on field fault slices and some tasks constrained by well logs, whereas specialized models retain advantages on some synthetic tasks, indicating that the benefits depend on the setting. A Martian radar example provides exploratory evidence for application across imaging modalities. These results support the feasibility of the approach for subsurface interpretation and modeling. This work provides an approach to unified modeling and interactive revision for heterogeneous target inference under insufficient observations, with potential extensions to medical imaging and other remote-sensing tasks.