Materials Genomics

1 Course Information

4th Semester – 5 ECTS · 2h lecture + 2h exercises per week, together with ML for Materials Processing & Characterization

2 Course Philosophy

Materials genomics treats the periodic table and all known crystal structures as a giant searchable, computable design space.

Students learn:

• how materials databases are built,
• how to represent matter as numbers, graphs, or fingerprints,
• how to interrogate and predict properties of solids,
• how to use ML as a surrogate for quantum mechanics,
• how to design new materials algorithmically.

3 Week-by-Week Curriculum (14 weeks)

3.1 Unit I — Foundations of Materials Genomics (Weeks 1–3)

3.1.1 Week 1 – What is Materials Genomics?

• Genomics analogy: genes → functions vs atoms → properties.
• Brief history: AFLOW, OQMD, Materials Project, NOMAD.
• PSPP from the structure-first viewpoint.

Exercise: Explore Materials Project; query bandgaps, energies, symmetries.

3.1.2 Week 2 – Crystal structure fundamentals

• Space groups, Wyckoff positions, symmetry operations.
• How symmetry informs descriptors.

Exercise: Using pymatgen / spglib to analyze symmetries.

3.1.3 Week 3 – Materials databases & file formats

• CIF, POSCAR, PDB-like formats.
• Thermodynamic quantities in databases: formation energy, stability, convex hull.

Exercise: Parse CIF files, extract primitive cells, compute density.

3.2 Unit II — Representations of Materials (Weeks 4–6)

3.2.1 Week 4 – Classical descriptors & materials fingerprints

• Magpie, matminer.
• Stoichiometric, elemental, and structural features.

Exercise: Build a small property regressor with Magpie features.

3.2.2 Week 5 – Graph-based representations

• Crystal structures as graphs: nodes, edges, periodic boundary conditions.
• CGCNN, MEGNet architecture intuition (no training from scratch yet).

Exercise: Build a simple CGCNN-like graph featurizer.

3.2.3 Week 6 – Local atomic environments

• Voronoi tessellations, coordination numbers, SOAP descriptors.
• Role in interatomic potentials and ML force fields.

Exercise: Compute SOAP vectors; perform clustering in descriptor space.

3.3 Unit III — High-Throughput Computation & Screening (Weeks 7–9)

3.3.1 Week 7 – Quantum mechanical data and DFT basics

• What DFT gives you: energies, forces, band structures, elastic constants.
• Why it’s expensive; why ML surrogates matter.

Exercise: Run a toy DFT calculation (Quantum Espresso or MP workflows).

3.3.2 Week 8 – High-throughput workflows

• Automation: pymatgen, custodian, FireWorks, Atomate.
• Data generation for building surrogate models.

Exercise: Perform a small FireWorks workflow (or simulate the idea without cluster resources).

3.3.3 Week 9 – Phase stability & the convex hull

• Formation energies, metastability, hull distance.
• Mapping an entire chemical system.

Exercise: Reconstruct phase diagrams from Materials Project data.

3.4 Unit IV — Learning Properties from Atomic Structure (Weeks 10–12)

3.4.1 Week 10 – Regression on crystal data

• Predicting bandgaps, hardness, elastic moduli.
• Comparing different representation families.

Exercise: Benchmark random forest, GPR, CGCNN on a small dataset.

3.4.2 Week 11 – Machine-learned interatomic potentials

• Overview: GAP, SNAP, MTP, NequIP.
• Role in simulating defects, diffusion, mechanical behavior.

Exercise: Fit a tiny ML potential (ACE or simple SNAP-style) to toy data.

3.4.3 Week 12 – Generative models for materials

• VAEs, diffusion models for crystal generation.
• Constraints: symmetry, stability, charge neutrality.

Exercise: Sample a generative model from a pretrained online source; analyze validity.

3.5 Unit V — Mini-Project & Synthesis (Weeks 13–14)

3.5.1 Week 13 – Project workshop

Example projects:

• Predict bandgap from composition + structure representation.
• Identify new stable compounds in a chemical system.
• Build a graph-based model for elastic constants.
• Use ML to approximate formation energies for a ternary subsystem.
• Analyze SOAP fingerprints across polymorphs.

3.5.2 Week 14 – Presentations & Reflection

• Interpreting models: SHAP for materials descriptors.
• Strengths/limitations of materials genomics vs experiment-driven ML.
• How computational and experimental ML meet in modern labs.

4 Learning Outcomes

Students completing this course will be able to:

• Navigate major materials databases and extract relevant structural/property data.
• Represent crystals numerically using descriptors, fingerprints, and graphs.
• Train ML models to predict quantum-mechanical and thermodynamic properties.
• Analyze structural features via symmetry, coordination, and environments.
• Perform high-throughput screening of materials candidates.
• Understand and apply generative models for inorganic crystals.
• Critically evaluate ML results in computational materials discovery.