materials informatics
For most of history, finding a better material meant slow trial and error in the lab — mix, heat, test, repeat, for years. Materials informatics asks a different question: what if we treat the accumulated results of all those experiments and calculations as DATA, and let a computer learn the patterns hidden in it? It is the application of data science, statistics and machine learning to materials — mining large databases of structures and properties to spot relationships, predict the properties of untested materials, and suggest what to make next. It is the 'learn from data' partner to the 'compute from physics' approach of structure prediction.
The engine is the pattern-finding machine-learning model, and the fuel is databases. Over the last decade, projects computed the properties of hundreds of thousands of compounds and gathered them into open repositories (the Materials Project, OQMD, AFLOW and others) — the practical realisation of the Materials Genome Initiative's idea that a searchable 'genome' of materials data could accelerate discovery. A model is trained on this data: you describe each material by numerical 'descriptors' or 'features' (its composition, its structure encoded numerically, its bonding), and the model learns the mapping from those features to a property — hardness, band gap, melting point, whether it is stable at all. Once trained, it can estimate that property for a brand-new candidate in milliseconds, far faster than either experiment or full quantum calculation, letting you screen millions of possibilities and flag the few worth real study.
Materials informatics matters because it dramatically speeds the search through the vast space of possible materials, and it excels exactly where structure prediction is expensive — providing quick, cheap property estimates to prioritise candidates. Be honest, though, about what it is and is not. A machine-learning model interpolates patterns in its training data; it is only as good as that data, and it typically fails when asked to extrapolate to genuinely new chemistry it has never seen. It finds correlations, which are not always causes, and it usually cannot explain the physics behind its predictions. The best practice today pairs it with physics-based calculation and experiment — informatics to narrow the field fast, first-principles and the lab to verify — rather than trusting the model alone.
To hunt for a new battery cathode, a model trained on the Materials Project database estimates the voltage and stability of tens of thousands of hypothetical lithium compounds in an afternoon. It ranks them, and only the top few dozen are handed to expensive quantum calculation and then the lab — a funnel that turns years of blind trial into weeks of targeted work.
Materials informatics: learn structure-property patterns from big databases to screen millions of candidates fast — the Materials Genome idea.
A model is only as good as its training data and interpolates rather than extrapolates — it often fails on genuinely new chemistry, finds correlation not causation, and cannot explain the physics. Best used to prioritise candidates for physics-based calculation and experiment, not to replace them.