douyimin / RGT-Est

What is RGT-Est?

RGT-Est is a deep learning framework for estimating Relative Geologic Time from 3D seismic data. In plain terms, it takes seismic reflection volumes and outputs a continuous field that represents the depositional age ordering of subsurface strata. The core innovation is a sinusoidal mapping approach that reformulates the problem from direct scalar regression into multi-scale phase optimization, which helps recover thin geological layers that conventional MSE/MAE losses tend to smooth away. Built in Python with PyTorch and PyTorch Lightning, it includes a pretrained model checkpoint ready for inference on field seismic surveys.

Why is it gaining traction?

The main hook here is the horizon quality. Traditional deep learning approaches for RGT estimation struggle with fine-scale stratigraphic details because their loss functions reward smoothness over geological accuracy. RGT-Est fights this by encoding periodic stratigraphic semantics directly into the optimization space, using adversarial, perceptual, and pointwise losses jointly. The optional horizon guidance module is a practical win too: feed in sparse 2D or 3D horizon constraints, and precision improves significantly without breaking the automatic workflow. The demos cover real field surveys from Australia, Costa Rica, Netherlands, and China, so the approach is validated across diverse structural complexity.

Who should use this?

This is squarely for geophysicists, structural geologists, and reservoir engineers working with 3D seismic interpretation. If you are building subsurface models, performing depositional analysis, or characterizing reservoirs, RGT-Est can accelerate horizon extraction and ensure stratigraphic consistency. Seismic imaging specialists evaluating AI-assisted interpretation tools will find the benchmark comparisons useful. General ML engineers exploring novel applications of sinusoidal representations in regression tasks will also find the approach conceptually interesting.

Verdict

RGT-Est solves a niche but real problem in computational geology with a technically sound approach. The 1.0% credibility score reflects the project's early stage: only 19 stars, minimal community engagement, and no training data or dataset code included in the release. Documentation is adequate for running demos, but production deployment requires careful handling of the fixed inference grid size constraint. Worth exploring for specialized seismic interpretation workflows, but treat the pretrained weights as a starting point rather than a turnkey solution.