A gold prospector doesn't wash every handful of sand by hand. First he sifts the ore through a coarse sieve, tossing out the empty pebbles, and only then carefully pans the remaining concentrate. Much the same approach can be used to hunt for useful defects in silicon carbide, a hard crystal that already sits inside power electronics and may one day become the basis of quantum computers.
A defect is a point disruption of the crystal lattice: either an empty spot where an atom is missing, or a foreign atom sitting in the wrong place. Such defects sometimes interfere with a device's operation, and sometimes, conversely, they turn into qubits, the quantum memory cells capable of storing and processing information. The trouble is that there are thousands of defect variants, and an exact calculation of each on a supercomputer takes hours or days. You need a sieve.
The researchers trained three types of models: a simple linear regression, a kernel regression, and a neural network. It turned out that even the simplest model does a decent job if the defect is described properly. The best-performing description used atomic numbers, distances, charge, and spin, with no complex electronic-structure calculations. For vacancies and substitutions the error was 0.437 eV, while for interstitials (extra atoms squeezed between lattice sites) it was 1.101 eV, which is worse but still useful for coarse screening.
It used to seem that you couldn't get by without a full calculation of the defect structure: you had to "relax" the crystal, that is, let the atoms settle into new positions around the defect, and that is expensive. Now it has become clear that even without relaxation you can get reasonable estimates. This approach lowers the barrier to entry: you don't need a huge computing cluster, just an ordinary computer with a trained model.
The sieve doesn't replace the panning, but it saves effort. The prospector still carefully checks what remains after sifting. In the same way, exact quantum calculations remain the final word, but now they are reserved for promising candidates rather than thousands of random variants. And who knows, maybe the next quantum nugget has already passed through such a sieve.
In 4H-SiC there are more than 200 polytypes, but it is 4H that is used in most commercial devices. Point defects in it can emit single photons, which makes them candidates for quantum cryptography.