Google DeepMind's Graph Networks for Materials Exploration (GNoME) predicted 381,000 stable inorganic compounds in 2023, the vast majority never synthesized. The predictions were thermodynamic: this combination of atoms, at this stoichiometry, should form a stable crystal. Whether any specific compound would be interesting — magnetic, superconducting, catalytic — was not part of the prediction. GNoME said “this should exist,” not “this should matter.”
Naganuma and Kitagawa synthesized one: MnFeCo₄Si₂. It forms a rhombohedral single-phase structure, confirming GNoME's structural prediction. But the experimentally discovered property is what matters: it's a soft ferromagnet with a Curie temperature of 1039 K — 766°C. That's remarkably high. For context, iron's Curie temperature is 1043 K. A compound nobody had ever made, predicted by a neural network trained on crystal stability, turns out to be ferromagnetic at nearly the same temperature as elemental iron.
The structural point: the existence prediction and the interesting-property discovery are decoupled. GNoME had no magnetic training data for this compound. The Curie temperature is emergent — a property of the electronic structure that the stability model couldn't see. The AI found the haystack; the experimentalists found the needle. This decoupling means the value of large-scale computational materials prediction isn't in predicting which materials have useful properties. It's in massively expanding the space of materials that exist to be tested. The prediction is of opportunity, not of outcome.
One compound from 381,000 predictions. If even a fraction of one percent have interesting properties, the yield is thousands of new functional materials. The bottleneck shifts from “what could exist” to “what's worth testing.”