EuroSciPy 2025: Simulation-Based Inference Tutorial

Beyond Likelihoods: Bayesian Parameter Inference for Black-Box Simulators with sbi

📍 Kraków, Poland | 90 minutes | Intermediate Track

🎯 Learning Objectives

By the end of this tutorial, you will:

• Understand when and why to use Simulation-Based Inference (SBI)
• Run parameter inference for any Python simulator using the sbi package
• Diagnose whether your inference results are trustworthy
• Apply SBI to your own scientific problems

📚 Pre-Tutorial Homework (Required)

1. Environment Setup

# Install uv if you haven't already
curl -LsSf https://astral.sh/uv/install.sh | sh

# Clone this repository
git clone https://github.com/janfb/euroscipy-2025-sbi-tutorial.git
cd euroscipy-2025-sbi-tutorial

# Create environment and install dependencies
uv venv
source .venv/bin/activate  # On Windows: .venv\Scripts\activate
uv sync
```text

### 2. Test Your Setup

```python
# Run this to verify installation
python -c "import sbi; import torch; print('✅ Setup complete!')"

Running the notebooks

• You can run the notebooks in VSCode or a similar IDE, selecting the euroscipy-2025-sbi-tutorial python kernel
• Alternatively, you can start jupyter notebook in the terminal and work on the notebooks in the browser.

Open notebooks in Colab

• notebook 1:
• notebook 2:
• notebook 3:

3. Optional but Recommended

• Read a bit about SBI:

  • Last year's EuroSciPy Talk, slides
  • SBI overview blog post
  • SBI: A Practical Guide

• Bring Your Simulator: If you have a scientific simulator you'd like to apply SBI to, prepare a simplified version that:

  • Takes parameters as input (numpy array or torch tensor)
  • Returns observations as output (numpy array or torch tensor)
  • Runs in < 1 second per simulation
  • Example: def simulator(params): return observations

📋 Tutorial Outline

Part 1: Why SBI? (15 min)

• The Environmental Monitoring Challenge
• Point estimates vs. uncertainty quantification
• When traditional methods fail

Part 2: Core Intuition (10 min)

• Rejection sampling in 5 lines of code
• Neural Posterior Estimation concept
• Learning parameter-data relationships

Part 3: Hands-On Exercises (55 min)

Exercise 1: Your First Inference (15 min)

• Lotka-Volterra predator-prey model
• Running NPE with sbi
• Visualizing posterior distributions

Exercise 2: Trust but Verify (20 min)

• Posterior predictive checks
• Coverage diagnostics
• Interpreting warning signs

Exercise 3: Your Own Problem (20 min)

• Adapt template to your simulator
• OR use provided examples (ball throw, SIR model)
• Specify priors and run inference

Part 4: Next Steps (5 min)

• Advanced SBI methods (NLE, NRE, sequential)
• Resources and community

Q&A (5 min)

🗂️ Repository Structure

euroscipy-2025-sbi-tutorial/
├── README.md                     # This file
├── pyproject.toml                # Project dependencies
└── materials/
    ├── 2024-08-30-EuroSciPy2024-SBI.pdf  # Optional pre-reading
    └── references.md                     # Further resources
├── slides/                       # Presentation slides
│   └── sbi_tutorial.md
    ├── sbi_tutorial.pdf
├── src/
│   ├── 00_setup_test.py          # Verify installation
│   ├── 01_first_inference.ipynb  # Exercise 1
│   ├── 02_diagnostics.ipynb      # Exercise 2
│   ├── 03_your_sbi_problem.ipynb     # Exercise 3
│   └── simulators/               # Example simulators
│       ├── lotka_volterra.py
│       ├── ball_throw.py
│       └── sir_model.py
    └── utils.py                  # Plotting utils

💻 Technical Requirements

• Python 3.10+
• Basic familiarity with PyTorch tensors
• Understanding of your own simulator (if bringing one)

👥 Instructors

• Jan Teusen (Boelts), TransferLab, appliedAI Institute for Europe
• Janos Gabler, TransferLab, appliedAI Institute for Europe
• Kristof Schröder, TransferLab, appliedAI Institute for Europe

📧 Support

• Before the tutorial: Open an issue on GitHub
• During the tutorial: Raise your hand
• After the tutorial: Join the sbi community on GitHub Discussions

🔗 Links

• sbi documentation
• sbi paper
• SBI overview blog post
• SBI Tutorial paper (preprint)

Acknowledgments

• SBI community: https://github.com/sbi-dev/sbi/graphs/contributors
• Funding: TransferLab, appliedAI Institute for Europe
• Organization: EuroSciPy 2025

📝 License

This tutorial is licensed under Apache 2.0.