53 lines (42 loc) · 1.6 KB

Week 1 Summary: What makes materials data special?

Cross-Book Summary

1. The Concept of Data-Based Modeling

Data-based vs. First-Principle: Top-down ML vs. bottom-up physics.
Traceability: White (explainable), Grey (hybrid), and Black-Box (opaque) models.
Overfitting: Excess complexity fails to generalize.

2. Foundations of Data

Types: Nominal, Ordinal, Cardinal, Binary.
Scales: Nominal, Ordinal, Interval, Ratio.
Uncertainty: Units and measurement error are essential.

3. Materials Science Specifics

PSPP Paradigm: Processing-Structure-Property-Performance dependency.
Noise: Physical noise, aliasing, and instrument bias.
Scarcity: High-quality data is expensive and rare.

90-Minute Lecture Strategy

Part 1: Introduction & Philosophy

AI 4 Materials goals.
Hype vs. Reality.
Convergence of high-throughput, simulation, and ML.

Part 2: Models in Engineering

Prediction vs. Explanation.
Physics-based vs. Data-driven.
Moving toward White-Box ML.

Part 3: Special Materials Data

PSPP graph.
Multi-modal data types.
Small Data challenges.
Physical Priors.

Part 4: Data Quality

Categorizing data.
Metadata and Units.
Error propagation.

Part 5: CRISP-DM for Labs

Adapting industrial standards.
Deployment.
Correlation != Causality.

Quarto Website Update (Summary)

Summary for ML-PC Week 1:

Transitions from physics-based to data-driven modeling.
Highlights multi-modal, scarce materials data.
Covers PSPP relationships, data scales, and uncertainty.
Adapts CRISP-DM for scientific workflows.