AI-Driven Subsurface Analysis

Executive Summary

This intensive 24-hour training program equips Abu Dhabi geoscientists with practical AI and machine learning skills for subsurface analysis. The course progresses from ML fundamentals through advanced deep learning to cutting-edge Generative AI and Agentic AI applications, with specific focus on carbonate reservoir characterization common to the region

 

Learning Objectives

Upon completion, participants will be able to:

  • Apply supervised and unsupervised ML algorithms to well log interpretation and lithology classification
  • Build regression models for porosity and permeability prediction
  • Implement rock typing workflows using clustering techniques
  • Design CNN architectures for seismic facies classification
  • Develop LSTM models for production forecasting
  • Understand Reinforcement Learning for well placement optimization
  • Leverage Large Language Models (LLMs) for technical knowledge retrieval
  • Build RAG systems for subsurface document analysis
  • Design Agentic AI workflows for autonomous analysis tasks

 

Course Structure

Day Theme Key Topics
Day 1 ML/DL Fundamentals Classification, Regression, Clustering, Data Prep
Day 2 Deep Learning & RL CNN for Seismic, LSTM for Production, RL Optimization
Day 3 GenAI & Agentic AI LLMs, RAG, Prompt Engineering, Autonomous Agents

Interactive Streamlit Dashboards

Each day includes a dedicated Streamlit dashboard for hands-on exercises:

  • Day 1: day1_ml_petrophysics_dashboard.py – ML algorithms for well log analysis
  • Day 2: day2_deep_learning_dashboard.py – CNN, LSTM, and RL demonstrations
  • Day 3: day3_genai_agentic_dashboard.py – LLM chat, RAG, and agent workflows

 

DAY 01

Machine Learning Fundamentals for Petrophysics

Classification, Regression, Clustering, and Data Preparation

Day 1 Learning Objectives

  • Understand supervised vs. unsupervised learning paradigms
  • Apply Random Forest for lithology classification from well logs
  • Build regression models for porosity and permeability prediction
  • Implement K-Means clustering for rock typing
  • Perform feature engineering on petrophysical data

Morning Session 1: ML Fundamentals (08:00-10:00)

Module 1.1: Introduction to AI/ML for Geoscience (45 min)

Setting the context for AI in subsurface analysis:

  • Why AI Now? Convergence of data, compute, and algorithms
  • AI vs. Traditional Methods: When to use ML vs. physics-based approaches
  • Abu Dhabi Context: Carbonate challenges and opportunities
  • Common Pitfalls: Data quality, overfitting, interpretability

Module 1.2: Supervised Learning Algorithms (60 min)

Core algorithms for petrophysical prediction:

  • Classification: Random Forest, XGBoost, SVM for lithology
  • Regression: Linear Regression, Decision Trees, Neural Networks
  • Model Evaluation: Accuracy, precision, recall, R², RMSE
  • Cross-Validation: Train/test split, k-fold validation

Demo 1.1: ML Algorithm Comparison Dashboard (15 min)

Interactive Streamlit demonstration showing algorithm performance on well log data.

Morning Session 2: Classification & Data Prep (10:15-12:15)

Module 1.3: Lithology Classification (45 min)

  • Feature Engineering: Log combinations, derivatives, ratios
  • Class Imbalance: SMOTE, class weights, stratified sampling
  • Carbonate Minerals: Limestone, dolomite, anhydrite signatures
  • Feature Importance: SHAP values, permutation importance

 

Exercise 1.1: Build a Lithology Classifier (45 min)

Hands-on exercise using the Day 1 Streamlit dashboard:

  • Load synthetic Abu Dhabi well log data
  • Select features and configure Random Forest parameters
  • Train model and evaluate confusion matrix
  • Analyze feature importance for geological interpretation

Case Study 1.1: ADNOC Carbonate Classification (30 min)

Real-world application of ML for lithology prediction:

  • Challenge: Complex Arab Formation mineralogy
  • Approach: Multi-class classification with GR, RHOB, NPHI, DT
  • Results: 92% accuracy, reduced interpretation time by 60%
  • Lessons: Importance of core calibration and quality control

Afternoon Session 1: Property Prediction (13:00-15:00)

Module 1.4: Porosity & Permeability Regression (60 min)

  • Porosity Prediction: Density-neutron relationships, ML alternatives
  • Permeability Estimation: Timur, Coates vs. ML approaches
  • Log Transforms: Why and when to use log-scale for permeability
  • Uncertainty Quantification: Prediction intervals, ensemble methods

Demo 1.2: Property Prediction Crossplots (20 min)

Demonstrate actual vs. predicted crossplots with error metrics.

Exercise 1.2: Poro-Perm Model Development (40 min)

  • Build porosity prediction model from density-neutron logs
  • Develop permeability transform using porosity and other features
  • Compare Random Forest vs. Gradient Boosting performance

Afternoon Session 2: Clustering & Quiz (15:15-17:00)

Module 1.5: Unsupervised Rock Typing (45 min)

  • K-Means Clustering: Algorithm mechanics, choosing K
  • Elbow Method: Optimal cluster selection
  • PCA Visualization: Dimensionality reduction for interpretation
  • Geological Interpretation: Mapping clusters to rock types

Exercise 1.3: Rock Type Identification (30 min)

  • Apply K-Means to porosity-permeability-GR data
  • Visualize clusters in PCA space
  • Interpret rock types and their reservoir significance

Day 1 Quiz & Wrap-up (30 min)

10-question quiz covering:

  1. Supervised vs. unsupervised learning
  2. Random Forest hyperparameters
  3. Feature importance interpretation
  4. Regression metrics (R², RMSE)
  5. Clustering applications

 

DAY 02

Deep Learning & Reinforcement Learning

CNN for Seismic, LSTM for Production, RL for Optimization

Day 2 Learning Objectives

  • Understand neural network architectures (FNN, CNN, RNN/LSTM)
  • Apply CNN for seismic facies classification
  • Build LSTM models for production time-series forecasting
  • Grasp Reinforcement Learning concepts for optimization
  • Implement RL for well placement decisions

Morning Session 1: Neural Networks & CNN (08:00-10:00)

Module 2.1: Deep Learning Fundamentals (45 min)

  • Neural Network Basics: Neurons, layers, activation functions
  • Feedforward Networks: Architecture for tabular data
  • Training: Backpropagation, gradient descent, optimizers
  • Regularization: Dropout, batch normalization, early stopping

Module 2.2: CNN for Seismic Interpretation (60 min)

  • Convolution Operation: Filters, feature extraction
  • CNN Architecture: Conv layers, pooling, flattening
  • Seismic Facies: Channel, shale, carbonate, salt detection
  • Transfer Learning: Leveraging pretrained models

Demo 2.1: CNN Training Visualization (15 min)

Watch a CNN learn to classify seismic patches in real-time.

Morning Session 2: CNN Exercises (10:15-12:15)

Exercise 2.1: Seismic Facies Classification (60 min)

Using the Day 2 Streamlit dashboard:

  • Explore synthetic seismic patches for 4 facies types
  • Configure CNN architecture (layers, filters)
  • Train model and monitor learning curves
  • Evaluate with confusion matrix

 

 

Case Study 2.1: Abu Dhabi Seismic AI (30 min)

Deep learning for carbonate seismic interpretation:

  • Challenge: Distinguishing Arab Formation members
  • Approach: U-Net semantic segmentation
  • Results: 85% IoU, 70% time reduction
  • Integration: Petrel plugin for production use

Discussion: CNN Interpretability (30 min)

  • Feature Visualization: What do filters see?
  • Grad-CAM: Highlighting important regions
  • Geoscientist Trust: Bridging black-box and domain knowledge

Afternoon Session 1: LSTM & Time Series (13:00-15:00)

Module 2.3: RNN and LSTM Networks (45 min)

  • Sequential Data: Why order matters in production
  • RNN Basics: Hidden states, vanishing gradients
  • LSTM Architecture: Forget, input, output gates
  • Applications: Decline curve prediction, pressure forecasting

Module 2.4: Production Forecasting with LSTM (45 min)

  • Data Preparation: Windowing, normalization, train/test split
  • Multivariate LSTM: Incorporating pressure, water cut
  • Uncertainty: Prediction intervals, Monte Carlo dropout
  • Evaluation: RMSE, MAPE, visual inspection

Exercise 2.2: LSTM Production Forecast (30 min)

  • Select well from dashboard
  • Configure lookback window and forecast horizon
  • Generate forecast with confidence intervals
  • Compare to actual production

Afternoon Session 2: Reinforcement Learning (15:15-17:00)

Module 2.5: Introduction to RL (45 min)

  • RL Framework: Agent, environment, state, action, reward
  • Q-Learning: Value function, exploration vs. exploitation
  • Deep RL: DQN, policy gradients
  • Applications: Well placement, injection optimization

Demo 2.2: RL Well Placement Simulation (15 min)

Watch an RL agent learn optimal well locations on a reservoir grid.

Exercise 2.3: RL Optimization (30 min)

  • Run RL agent on NPV maximization task
  • Observe exploration-exploitation trade-off
  • Compare convergence across multiple runs

Day 2 Quiz & Wrap-up (30 min)

10-question quiz covering:

6. CNN architecture components

7. LSTM gate functions

8. Seismic facies detection

9. Production forecasting metrics

10. RL agent-environment interaction

 

 

DAY 03

Generative AI & Agentic AI

Large Language Models, RAG Systems, and Autonomous Agents

Day 3 Learning Objectives

  • Understand LLM architecture and capabilities
  • Apply prompt engineering techniques for geoscience tasks
  • Design and implement RAG systems for technical documents
  • Build Agentic AI workflows for autonomous analysis
  • Navigate safety and governance for AI agents

Morning Session 1: LLM Fundamentals (08:00-10:00)

Module 3.1: Introduction to Large Language Models (45 min)

  • Transformer Architecture: Self-attention, positional encoding
  • Foundation Models: GPT-4, Claude, Gemini, Llama
  • Capabilities: Text generation, reasoning, code, multi-modal
  • Limitations: Hallucinations, context limits, knowledge cutoff

Module 3.2: LLMs for Subsurface Applications (45 min)

  • Technical Q&A: Answering geoscience questions
  • Document Summarization: Well reports, technical papers
  • Code Generation: Python scripts for analysis
  • Report Writing: Drafting interpretations and recommendations

Demo 3.1: LLM Chat Interface (30 min)

Interactive demonstration using Day 3 Streamlit dashboard:

  • Ask questions about Arab Formation
  • Request log interpretation guidance
  • Generate Python code snippets

Morning Session 2: RAG & Prompt Engineering (10:15-12:15)

Module 3.3: Retrieval-Augmented Generation (45 min)

  • RAG Architecture: Embedding, vector DB, retrieval, generation
  • Document Processing: PDFs, technical reports, standards
  • Chunking Strategies: Size, overlap, semantic boundaries
  • Vector Databases: ChromaDB, Pinecone, Weaviate

 

Module 3.4: Prompt Engineering (45 min)

  • Zero-Shot: Direct instructions without examples
  • Few-Shot: Providing examples to guide output
  • Chain-of-Thought: Step-by-step reasoning
  • Structured Output: JSON, tables for downstream processing

Exercise 3.1: RAG System Demo (30 min)

  • Query the knowledge base about carbonate interpretation
  • Observe document retrieval and relevance scoring
  • Evaluate response grounding

Afternoon Session 1: Agentic AI (13:00-15:00)

Module 3.5: Introduction to Agentic AI (45 min)

  • Agent Architecture: Perception, reasoning, action, memory
  • Tool Use: Function calling, API integration
  • Planning: Task decomposition, goal-directed behavior
  • Multi-Agent Systems: Specialist collaboration

Module 3.6: Agentic Workflows for Subsurface (45 min)

  • Well Analysis Agent: Query data, compute, visualize, report
  • Interpretation Agent: Retrieve guidelines, apply cutoffs
  • Multi-Agent Example: Geologist + Engineer + Economist
  • Safety & Governance: Human-in-the-loop, guardrails, audit trails

Exercise 3.2: Build an Analysis Agent (30 min)

Using Day 3 Streamlit dashboard:

  • Define agent task: ‘Analyze all wells’
  • Watch step-by-step execution trace
  • Review agent output and recommendations

Afternoon Session 2: Integration & Quiz (15:15-17:00)

Case Study 3.1: ADNOC GenAI Pilot (30 min)

Enterprise deployment of GenAI for operations:

  • Challenge: Technical knowledge scattered across documents
  • Solution: RAG-based assistant with 50K+ indexed documents
  • Results: 40% faster report drafting, improved consistency
  • Lessons: Data quality, user training, governance

 

Module 3.7: Future Directions (30 min)

  • Multi-Modal Models: Combining logs, seismic, text
  • Foundation Models for Geoscience: Domain-specific pretraining
  • Autonomous Operations: AI-driven field management
  • Responsible AI: Bias, transparency, workforce impact

Day 3 Quiz & Course Wrap-up (30 min)

10-question quiz covering:

  1. LLM capabilities and limitations
  2. RAG system components
  3. Prompt engineering techniques
  4. Agent architecture
  5. Safety considerations

Course completion certificates and next steps discussion.

 

Appendices

Appendix A: Streamlit Dashboard Instructions

To run the interactive dashboards:

  1. Ensure Python 3.8+ is installed
  2. Install dependencies: pip install streamlit pandas numpy plotly scikit-learn
  3. Run dashboard: streamlit run day1_ml_petrophysics_dashboard.py
  4. Access in browser at http://localhost:8501

Appendix B: Glossary

  • CNN: Convolutional Neural Network
  • DL: Deep Learning
  • FNN: Feedforward Neural Network
  • GenAI: Generative Artificial Intelligence
  • LLM: Large Language Model
  • LSTM: Long Short-Term Memory
  • ML: Machine Learning
  • RAG: Retrieval-Augmented Generation
  • RL: Reinforcement Learning
  • SHAP: SHapley Additive exPlanations

Appendix C: Certification

To receive the course completion certificate:

  • Attend all 3 days (minimum 80% attendance)
  • Complete all hands-on exercises
  • Achieve minimum 70% average on daily quizzes
  • Complete course evaluation survey

16 noviembre 2026

Ends: 18 noviembre 2026

24 Hours

Abu Dhabi

USD 4950

Geosciences

Rated 4.5 out of 5

Contacto

Inscripcion al curso

Solicitar in company