TorchSeis: A Unified 3D Deep Learning Framework for Seismic and Subsurface Modeling

TorchSeis is an open research framework that explores the intersection of deep learning, seismic data analysis, and subsurface intelligence. It provides reusable models, operators, and experimental tools aimed at advancing AI-driven understanding of the Earth, with applications across:

• Seismic interpretation and structural analysis (volumetric)
• Faults, stratigraphy, channels, karst, salt bodies, and other geologic features
• Efficient whole-volume inference without patching artifacts for ultra-large 3D datasets

TorchSeis is intended for both research and industrial deployment, enabling scalable, reproducible, and AI-driven subsurface understanding.

---

Contents

📌 Full-Volume Inference Framework

TorchSeis provides a retraining-free full-volume inference framework for 3D dense prediction, enabling seamless processing of datasets up to 1024³ or larger on a single GPU, while preserving spatial continuity and structural coherence.

This work originates from the paper "Memory-Efficient Full-Volume Inference for Large-Scale 3D Dense Prediction without Performance Degradation", which presents an operator-level optimization strategy that enables high-resolution inference on modern GPU hardware without degrading model accuracy.

📄 Details can be found at docs/infer25ce.md

🧭 Seismic Operators & Attributes

• Impedance to Seismic with different wavelets
• Structure Tensor Orientations
  • see scripts/orientations.ipynb for serval examples.
• Local Slope

🔬 Citation

If this repository is useful for your research, please citing the relevant citations below:

@article{li2025infer,
  title={Memory-Efficient Full-Volume Inference for Large-Scale 3D Dense Prediction without Performance Degradation},
  author={Li, Jintao and Wu, Xinming},
  journal={Communications Engineering},
  year={2025}
}