Materials Discovery With Machine Learning: Predicted Is Not Made

A materials model does not predict that a compound exists. It predicts a number from a simulation — usually how far a candidate crystal sits above the lowest-energy combination of its constituents. Everything confusing about this field follows from the gap between that number and a jar of powder on a bench.

What the label actually is

The standard training signal is density functional theory. DFT is an approximate quantum mechanical calculation that returns, among other things, the energy of a given crystal structure. Run it over many candidates and you can compute the energy above the convex hull: zero means the composition is the most stable arrangement of those elements at that stoichiometry, and a small positive value means it is metastable but plausible. Large public databases of these calculations exist, and they are what graph neural networks in this field are trained on.

Two consequences, and both are routinely dropped. First, a model trained on DFT can at best reproduce DFT. Where DFT has known systematic problems — strongly correlated electron systems, van der Waals interactions, certain oxides — the model inherits them exactly, and no amount of additional training data fixes an error that is in every label. Second, thermodynamic stability is not synthesisability. A structure can be the global energy minimum and have no accessible kinetic route to it from any set of precursors you can buy, and plenty of materials people use every day are metastable, which the hull criterion would score as unfavourable.

The quieter and larger result

The application that has changed daily practice most is not candidate generation at all. It is machine-learned interatomic potentials: models trained on DFT forces and energies that then predict them for new configurations at a tiny fraction of the cost. That turns molecular dynamics simulations which were previously restricted to a few hundred atoms and a few picoseconds into simulations of far larger systems over far longer times, at accuracy that is meaningfully closer to DFT than the hand-fitted classical force fields they replace.

This is a surrogate model in the exact technical sense, with the same strengths and the same failure mode: excellent inside the distribution it was trained on, and confidently wrong outside it. A potential trained on equilibrium structures asked about a bond-breaking transition state is being asked to extrapolate, and it will answer.

From predicted to real

These are separate achievements. Reporting one as another is the characteristic error of the field.

• Proposed. A composition and structure exists as a file. Generation is cheap and effectively unlimited.
• Computationally stable. DFT places it on or near the hull. This is a statement about a calculation.
• Synthesised. Somebody made a sample. This requires a route, precursors, a temperature profile and, usually, iteration.
• Characterised. The sample was shown to be the intended phase, at reasonable purity, by diffraction and composition analysis. This is the step that decides whether “made” is true.
• Property measured. The thing you wanted it for — conductivity, capacity, hardness, transition temperature — was measured on that sample.
• Useful. It can be made at scale, from affordable elements, stably enough to put in a product. Most compounds that clear every rung above stop here.

What the disputes were about

Large-scale generation efforts — most visibly the GNoME work on predicted stable crystals, and autonomous synthesis laboratories such as the A-Lab — attracted published critiques from experimental materials chemists. It is worth being precise about the substance, because it generalises well beyond those specific papers.

The criticisms clustered on two rungs above. On the generation side, the argument was that a great many predicted structures were compositional substitutions into known structure types — swapping one element for a chemically similar one in a framework already in the literature — which is a legitimate computational result and a weak novelty claim, and that stability under DFT had been presented as though it settled whether the compound could be made. On the automated synthesis side, the argument was about characterisation: identifying a phase from a powder diffraction pattern is an inverse problem with real degeneracies, automated fitting can return a plausible match that a crystallographer would reject, and if the identification is wrong then the claim to have synthesised the target is wrong regardless of how well the rest of the pipeline ran.

Notice that neither criticism is about the models being bad at what they do. Both are about a rung being reported as the rung above it, and both were made by domain experts reading the supplementary material. That is the normal, healthy operation of a field, and it is also the best available argument for keeping raw characterisation data public.

Evaluating a discovery claim

• Was it already known? Ask whether the compound appears in an experimental structure database. A prediction that recovers a known compound is a validation result, not a discovery.
• Is it a substitution or a new structure type? Replacing an element in a known framework and finding a genuinely new topology are different kinds of new.
• Who did the characterisation, and is the raw data available? Refined diffraction patterns and composition analysis, not a summary table. This is where the argument will happen, so it is what has to be published.
• Was the property measured or predicted? A predicted band gap on a predicted structure is two models deep and zero experiments deep.
• What was the hit rate? If a thousand candidates were proposed and three were made, that ratio is the informative number and it is frequently not reported.