Machine Learning in Materials: Data, Features and Physical Meaning
A disciplined introduction to datasets, descriptors, validation, interpretability and responsible use of machine learning.
Begin with the decision
Machine learning is useful when a clear prediction, classification or optimisation task exists. The target should connect to a real materials or process decision.
Data quality sets the ceiling
Small, biased or inconsistent datasets produce fragile models. Metadata, units, test conditions, missing values and repeated samples must be controlled before algorithm selection.
Features encode assumptions
Composition, process parameters, microstructural descriptors and physics-based quantities can all serve as features. Feature design determines which relationships the model can discover.
Validation must reflect deployment
Random train-test splits may exaggerate performance when related alloys or repeated experiments appear in both sets. Grouped, temporal or external validation is often more realistic.
Working rule
Use machine learning to complement mechanisms and experiments. Report uncertainty, domain limits and failure cases—not only the best score.