Materials informatics: a practical guide
What materials informatics is, how it turns data into material property predictions, the tools and databases behind it, and where it quietly breaks.
Materials informatics is the practice of using data and machine learning to predict the properties of materials and to decide which ones are worth making next. Instead of discovering materials one slow experiment at a time, you learn patterns from thousands of materials that are already known, then use those patterns to rank, screen, or design new candidates.
This is the pillar for that topic. Under it sit the hands-on articles on this site: the band-gap prediction series, the Materials Project API tutorial, and the pieces on why representation decides your result. If you want the practice, follow the links as you go.
What materials informatics actually is
The field sits at the meeting point of materials science, computational chemistry, and data science. The core idea is simple: material properties are not random. They follow from composition and structure, and those relationships can be learned from data. Once learned, a model can estimate a property for a material nobody has measured yet, in milliseconds rather than months.
That does not replace experiment or physics. It reorders the work. You use a model to narrow millions of possibilities down to a shortlist, then spend your expensive lab time only on that shortlist.
The workflow, end to end
Almost every materials informatics project moves through the same four stages, and knowing them is most of the battle:
- Data. You start from a database of materials with known properties. The best known is the Materials Project, which holds computed properties for over 150,000 inorganic materials. Others include OQMD, AFLOW and JARVIS. The full landscape, with references, is on the resources page.
- Representation. You turn each material into numbers a model can read. This is the step that decides everything, and it is covered below.
- Model. You fit a model, from a simple linear or tree model up to a graph neural network, to map representation to property.
- Decision. You use the model to screen, rank, and recommend, then confirm the top candidates against a real source or a lab.
Representation is where accuracy is won or lost
The most common mistake is to describe a material only by its composition, which elements and in what fractions. That throws away structure, and structure is often the whole story. Two materials with identical composition can crystallise into different structures with very different properties. A composition-only model cannot tell them apart, so it is capped before training even begins.
This is not a small effect. It is the difference between a model that works and one that has a hidden ceiling. Two articles here go deep on it: why properties belong to structures, not formulas, and how graph neural networks finally let a model read structure.
A worked example: predicting band gaps
The clearest way to learn the whole pipeline is to run it once. The band-gap prediction series does exactly that, from raw Materials Project data through features, models, leakage, and honest error analysis, across eight parts. If you want to pull the data yourself first, start with the Materials Project API tutorial.
The tools you will actually use
The ecosystem is mostly open source and Python. The core is pymatgen for structures and data access, matminer for featurisation, and graph-model libraries like MatGL, CHGNet and MACE for structure-aware prediction. Each of these, with links, is listed on the resources page.
The honest part: where it breaks
A materials informatics model is only as trustworthy as the data under it and the split you test it on. Three failure modes recur:
- Data leakage. If near-identical materials sit in both training and test sets, the score looks great and means nothing. This is subtle, because whether a split leaks depends on what the model can see.
- Distribution shift. A model trained on one family of materials often collapses on a different family. High accuracy in-sample is not the same as a useful prediction out-of-sample.
- A prediction is not a material. A promising number does not mean the material can be made, is stable, or is safe. Most predicted materials are never synthesised, for reasons worth understanding: why predicted materials rarely get made.
The habit that separates good work from demos is auditing the data before trusting the model. That is its own skill: how to audit a database entry.
Start here
A reading path through this site, in order:
- Properties belong to structures, the core idea.
- Materials Project API tutorial, get the data.
- Band-gap prediction, part 1, run the full pipeline.
- Graph neural networks for materials, break the composition ceiling.
- Why predicted materials rarely get made, keep it honest.
Materials informatics is not magic and it is not a shortcut around physics. It is a way to spend your scarce experimental effort on the few candidates most likely to work, and to be honest about the uncertainty on the rest.
Found a mistake? Good, tell me. This publication flags its own suspect values. Reach me on LinkedIn.