🌍 DeepEarthVision

Physics-Informed Generative Architectures for Seismic Inversion and Geoenergy

---

📌 Project Overview

This repository hosts the doctoral research project:

Relational Bias and Multi-Path Reasoning for Physics-Informed GANs (PI-GANs)

The project integrates Generative Adversarial Networks (GANs), Physics-Informed Neural Networks (PINNs), Relational Inductive Biases, and Vectorial Chain-of-Thought (CoT) to address inverse problems in seismic inversion and geoenergy applications.

---

🎯 Objectives

• Design a Physics-Informed Generative Architecture combining PI-GANs and PINNs.
• Introduce Relational Attention Biases to capture geological connectivity.
• Implement Multi-Path Vectorial Reasoning (CoT) to improve exploration of solution space.
• Validate on benchmark datasets such as OPENFWI.
• Extend to applications in energy exploration, geothermal systems, and CO₂ storage.

---

🔑 Innovation & Contribution

• Hybrid framework combining generative modeling and physical PDE constraints.
• Relational inductive bias to encode topological/geological structures.
• Quantized computation (FP4, INT4) for efficiency and scalability.
• Doctoral framework (2025–2028) with planned expansion toward industrial and international applications.

---

📍 Impact

• Academic: Publications in high-impact journals (GJI, Commun. Comput. Phys., The Innovation Energy).
• Industrial: Reduction of seismic inversion costs and improved precision in Tabasco’s energy sector.
• Social: Technology transfer and local talent development in AI + Geoscience.
• International: Positioning within global consortia on AI for Science.

---

📊 Project Roadmap

1. Phase 1 (2025–2028): Doctoral Research

   • Develop and validate PI-GAN + PINNs architecture.
   • Benchmark on OPENFWI datasets.

2. Phase 2 (2028–2030): Technology Transfer

   • Application in local reservoir exploration in Tabasco.
   • Collaboration with energy industry partners.

3. Phase 3 (2030–2032): International Consolidation

   • Expansion to geothermal and CO₂ storage.
   • Establishment of a regional AI for Geoscience Hub.

---

📂 Repository Structure

DeepEarthVision/
│── docs/            # Project documentation (proposals, slides, papers)
│── data/            # Links / scripts for OPENFWI and generated datasets
│── src/             # Core source code (models, training, evaluation)
│── notebooks/       # Jupyter notebooks for experiments
│── results/         # Outputs, trained models, and logs
│── figures/         # Diagrams, plots, and visualizations
│── README.md        # Global project overview
│── LICENSE          # License for use and contributions

---

📚 Key References

• Raissi et al. (2019). Physics-Informed Neural Networks. JCP.
• Jagtap & Karniadakis (2020). XPINNs: Domain Decomposition for PINNs. Commun. Comput. Phys.
• Tronci et al. (2024). Physics-Informed Machine Learning Strategies. Springer
• Wang et al. (2025). PIDL for Geoenergy Development. The Innovation Energy
• Deng et al. (2022). OPENFWI Benchmark Datasets. NeurIPS

---

👉 This repository is both an academic framework and a living research project evolving with the doctoral journey.

---