Why AI-predicted materials rarely get made
Models have predicted millions of new materials, yet only hundreds exist. Understanding the gap between a predicted structure and a real one.
In 2023, Google DeepMind’s GNoME model predicted 2.2 million new crystalline materials, with about 381,000 flagged as stable enough to be worth making. Three years later, the number actually synthesised in laboratories sits somewhere under a thousand. This article explains where that enormous gap comes from, because the reasons teach you most of what matters about computational materials data.
What “predicted” actually means
A predicted material is a crystal structure that exists only inside a simulation. A model proposes an arrangement of atoms, and physics-based calculations (usually density functional theory, DFT) estimate whether that arrangement would hold together and what properties it would have. No powder, no crystal, no sample. A prediction is a hypothesis with coordinates.
What “stable” actually means
Stability in these databases is measured by a quantity called energy above hull, in electronvolts per atom. Think of it as the material’s height above the energy floor: the “hull” is the set of lowest-energy combinations that composition can form.
- 0 eV/atom: the material sits on the floor. Thermodynamically stable; nature has no cheaper alternative.
- Up to ~0.05 eV/atom: metastable but plausible. Diamond, famously, is metastable (graphite is carbon’s floor), yet diamonds exist because the conversion is impossibly slow.
- Above ~0.1 eV/atom: very unlikely to ever be made. Nature will prefer to form something else instead.
Here is the catch: “stable” in a database means computed to sit on the hull, within the error bars of the method that computed it. DFT’s errors are a few hundredths of an eV/atom, which is exactly the scale that separates “makeable” from “fantasy.” Many predicted-stable materials are within the noise.
Stability is necessary, not sufficient
Even a genuinely stable structure needs a synthesis route: a set of precursors, temperatures, and steps that arrive at that structure rather than a competing one. Predicting the structure says nothing about the route. Some stable materials have no known way to be made; some metastable ones (diamond again) are made industrially at scale. This is why the honest question about any database entry is not “is it stable?” but “has anyone ever made it?”, and databases do record evidence for that, in the form of experimental structure matches and text-mined synthesis recipes from the literature.
The quieter problem: the training data
Every predictive model learns from existing databases, and those databases contain systematic errors that the models inherit. Three classes worth knowing:
- Method artifacts. Standard DFT famously underestimates band gaps and fails outright on some materials families, recording well-known semiconductors with a gap of zero. A model trained on those values learns physics that is wrong in a consistent, invisible way.
- Silent incompleteness. Most properties are computed for only a fraction of entries. Absence of a value means “not calculated,” but pipelines that fill gaps with zeros teach the model something false.
- Duplicate and near-duplicate structures. The same composition appears as many entries (polymorphs), and careless dataset construction lets near-identical structures land on both sides of a train/test split, inflating reported accuracy.
None of these are scandals; they are documented limitations. The problem is that nothing in the data warns you. A value fetched from a database API looks equally authoritative whether it is a measured constant or a known artifact.
What to take from this
When you read a claim that a model discovered N thousand new materials, translate it: N thousand hypotheses, of which some fraction are within computational error of being unmakeable, judged against data with known systematic biases, awaiting synthesis routes that nobody has proposed. That is still valuable, screening hypotheses cheaply is the entire point, but it is a different claim from “discovered.”
And when you use computed data yourself, three habits cover most of the risk: check the energy above hull before treating an entry as real, check whether experimental evidence exists for it, and treat any suspiciously convenient value (a zero, a round number, an outlier) as a question rather than an answer.
The bottleneck of AI materials discovery is not generating candidates. It is knowing which numbers deserve trust, and that is a data problem before it is a physics one.
Found a mistake? Good, tell me. This publication flags its own suspect values. Reach me on LinkedIn.