Rgt Seismic Eval

rgt-seismic-eval

Massive-scale unlabeled field and labeled synthetic seismic datasets of global shelf-edge clinothems — Hui Gao et al. (arXiv:2604.17235, 2026)

What this evaluates

Evaluates a model's ability to perform continuous regression for Relative Geologic Time (RGT) estimation from 2D seismic images. It probes the model's capacity to learn stratigraphic continuity and structural consistency across diverse geological settings, testing generalization from synthetic labeled data to unlabeled real-world field data.

Datasets

• Field Seismic Dataset — total 3000; splits: test (3000)
• Synthetic Seismic Dataset — total 4000; splits: train (4000)

Metrics

• regression (primary) — range: other
  • Continuous regression error measuring the difference between predicted and ground-truth Relative Geologic Time (RGT) values. The specific loss function (e.g., MSE or MAE) is not explicitly named in the provided text, but standard regression metrics apply.

Input / output format

Input: Normalized 2D seismic image/profiles representing seismic amplitude data.

Output: Continuous Relative Geologic Time (RGT) values per pixel/voxel or profile.

Scoring recipe

def compute_regression_error(predictions, ground_truth):
    # predictions and ground_truth are tensors of RGT values
    mse = np.mean((predictions - ground_truth) ** 2)
    mae = np.mean(np.abs(predictions - ground_truth))
    return {"MSE": mse, "MAE": mae}

Common pitfalls

• The field dataset lacks RGT labels, requiring unsupervised or self-supervised learning strategies rather than direct supervised evaluation.
• Synthetic data generation involves complex forward modeling with many hyperparameters, which can introduce domain gaps if not carefully tuned.
• Evaluating generalization requires testing on unseen basins or sea-level cycle configurations not present in the training split.

Evidence (verbatim from paper)

Relative geologic time (RGT) estimation is a continuous regression task that is highly sensitive to stratigraphic continuity and structural consistency. Its performance depends not only on network architecture and optimization strategy, but also critically on the representativeness of the training data and the quality of the target labels.

Citation

@misc{gao2026massivescale,
  title={Massive-scale unlabeled field and labeled synthetic seismic datasets of global shelf-edge clinothems},
  author={Hui Gao et al.},
  year={2026},
  note={arXiv:2604.17235}
}

• arXiv: 2604.17235