Pioneering next-generation AI materials design through graph representations of molecules and materials
About This Series
This series is an educational content of five chapters that lets you learn, step by step from the fundamentals, the practical skills needed to apply Graph Neural Networks (GNN) to materials science and chemistry.
Graph Neural Networks (GNN) are an innovative technology that represents molecules and crystals as graph structures ("atoms = nodes", "bonds = edges") and predicts their properties with deep learning. GNNs automatically learn the complex interatomic interactions and spatial arrangements that conventional materials descriptors (composition, density, symmetry, etc.) could not capture.
Features:
- ✅ Step-by-step structure: Each chapter can be read as an independent article, and all five chapters together cover the topic comprehensively
- ✅ Practice-oriented: 38 executable code examples and 5 detailed case studies
- ✅ Materials-science focused: Focuses on applications to molecules, materials, and catalysts rather than general graph learning
- ✅ Cutting-edge techniques: Covers the latest equivariant GNN methods such as SchNet, DimeNet, NequIP, and MACE
- ✅ Career support: Provides concrete career paths and learning roadmaps
Total Reading Time: 110-130 minutes (including code execution and exercises)
How to Study
Recommended Learning Path
For beginners (no prior GNN knowledge):
- Chapter 1 → Chapter 2 → Chapter 3 → Chapter 4 → Chapter 5 (all chapters recommended)
- Estimated time: 110-130 minutes
For intermediate learners (with deep learning experience):
- Chapter 2 → Chapter 3 → Chapter 4 → Chapter 5
- Estimated time: 90-110 minutes
Practical skill enhancement (implementation over theory):
- Chapter 3 (focused study) → Chapter 4 → Chapter 5
- Estimated time: 70-90 minutes
Chapter Details
Chapter 1: Why GNN for Materials Science
Difficulty: Introductory
Reading Time: 20-25 min
Code Examples: 6
Learning Content
- What is a graph - Fundamentals of graph theory, graph representation of molecules and materials
- Limitations of conventional materials descriptors - Loss of structural information, the diamond vs. graphite example
- Success stories of GNN - Achievements on QM9, OC20, and Materials Project
- Why GNN is suited to materials science - Translational invariance, rotational equivariance, permutation invariance
Learning Objectives
- ✅ Explain the fundamentals of graph representation
- ✅ List three limitations of conventional materials descriptors with concrete examples
- ✅ Understand the technical background that motivates GNN
Chapter 2: GNN Theory Fundamentals
Difficulty: Beginner to Intermediate
Reading Time: 25-30 min
Code Examples: 10
Learning Content
- Definition of GNN and related fields
- Message Passing Neural Networks (MPNN) - Mathematical definition of message passing
- Major GNN architectures - GCN, GAT, SchNet, DimeNet
- Theory of equivariant GNN - The concept of E(3)-Equivariance
Learning Objectives
- ✅ Understand the three steps of message passing
- ✅ Explain the characteristics and use cases of GCN, GAT, SchNet, and DimeNet
- ✅ Understand the concept of E(3)-Equivariance
Chapter 3: PyTorch Geometric Practice
Difficulty: Intermediate
Reading Time: 30-35 min
Code Examples: 12 (all executable)
Learning Content
- Environment setup - Three options: Anaconda, venv, and Google Colab
- Fundamentals of PyTorch Geometric - Data structures, DataLoader
- Molecular property prediction (QM9 dataset) - Prediction with GCN and SchNet
- Crystal property prediction (Materials Project) - CGCNN implementation
- Model performance comparison
Learning Objectives
- ✅ Set up a Python environment
- ✅ Predict molecular properties on the QM9 dataset
- ✅ Predict crystal properties on Materials Project data
Chapter 4: Advanced GNN Techniques
Difficulty: Intermediate to Advanced
Reading Time: 20-25 min
Code Examples: 8
Learning Content
- Equivariant GNN (E(3)-Equivariant GNN) - SchNet, NequIP, MACE
- Directional message passing - DimeNet, SphereNet, GemNet
- Transformer + GNN - Graphormer, Graph Transformer
- Pretrained models - MolBERT, ChemBERTa, Uni-Mol
- GNNExplainer - Interpretability
Learning Objectives
- ✅ Understand the theory and implementation of E(3)-Equivariant GNN
- ✅ Implement directional message passing such as DimeNet and GemNet
- ✅ Visualize the rationale behind predictions with GNNExplainer
Chapter 5: Real-World Applications and Careers
Difficulty: Intermediate to Advanced
Reading Time: 15-20 min
Code Examples: 6
Learning Content
- Catalyst design (OC20 Challenge)
- Crystal structure prediction (CGCNN, Matformer, MODNet)
- Materials screening - High-throughput exploration workflows
- Industrial application cases - Battery materials, catalysts, drug discovery
- Career paths - Academia, industry, startups
- Learning roadmap - 3-month / 1-year / 3-year plans
Learning Objectives
- ✅ Understand the latest trends in catalyst design (OC20)
- ✅ Build a materials screening workflow
- ✅ Understand the career paths of GNN specialists
- ✅ Grasp the required skill set and learning roadmap
Overall Learning Outcomes
Upon completing this series, you will acquire the following skills and knowledge:
Understanding
- ✅ Explain the historical background of GNN and its necessity in materials science
- ✅ Understand basic concepts such as message passing, equivariance, and attention mechanisms
- ✅ Distinguish and apply major methods such as GCN, GAT, SchNet, DimeNet, and NequIP
- ✅ Describe five or more real-world success stories in detail
Doing
- ✅ Set up a Python environment and install PyTorch Geometric
- ✅ Convert molecules and materials into graph data
- ✅ Predict molecular properties on the QM9 dataset (MAE < 0.1 eV)
- ✅ Predict crystal properties on Materials Project data (R² > 0.9)
- ✅ Implement GCN, SchNet, and CGCNN and compare their performance
Applying
- ✅ Select the appropriate GNN architecture for a new materials problem
- ✅ Leverage pretrained models through transfer learning
- ✅ Evaluate industrial deployment cases and apply them to your own research
- ✅ Concretely plan your future career path
Let's Get Started!
Are you ready? Start with Chapter 1 and begin your journey into the world of GNN!
Chapter 1: Why GNN for Materials Science →
Update History
- 2025-10-17: v1.0 first edition published
Your GNN learning journey begins here!