Mishchenko: flat-band materials cannot be picked by a threshold, so his group sorts them into families with a neural network and clustering

What a flat band is and why it is worth finding

Besides energy, an electron in a crystal has a 'crystal momentum' that describes how the phase of its wave is arranged across the atoms. Normally energy changes with it; in a flat band the energy is the same whatever its value. That puts a great many electrons into a very narrow range of energy with nowhere else to go, so they must interact and form collective states, sometimes magnetic order, sometimes superconductivity. He says copper-based, iron-based and nickel-based superconductors 'I think all' have flat bands, and that twisted bilayer graphene at 1.05 degrees has its bands flattened and a superconductor at about 1 K, an example in pure carbon. He grants that 'flat band is necessary' is an experimental observation and that he knows of no deeper theory.

“you want to have this flat band and then on top of that flat band the interactions will happen and that will drive some new physics.”Artem Mishchenko

The old way uses thresholds and hand-written rules, with the person's taste built in

Databases hold computed band structures for hundreds of thousands of materials, and the question is which are flat. One approach says a band varying by less than 50 millielectronvolts is flat, so 49 is flat and 51 is not, which he finds artificial; rules built on eigenvalue gradients need many details written in advance and are also biased. He wants a generic method that can extend to other systems.

Step one: an image network picks flat bands, then structure of the sublattice clusters them

They took 2DMatPedia, a freely downloadable database of two-dimensional materials, stripped the axes from the band-structure images, made them black and white, cut them into small tiles, and had a neural network judge each tile flat or not. He says a simple multilayer perceptron worked at first, a referee asked for something 'more advanced', so they switched to a convolutional network and retitled the paper 'deep learning', and it needed less training. The labels come from researchers, postdocs label better than PhD students, and the bias is still there. They then keep only the atoms that contribute most to the flat band, form a sublattice from them, compute similarity with a structure fingerprint (a vector of about 244 numbers), and cluster with HDBSCAN and plot with t-SNE; the result is about 50 clusters that a person spent weeks reading, with about two thirds of the materials in no cluster, which he says is fine because the aim is to find trends, not to classify everything.

Step two: no structure, the band image itself as the fingerprint

The second idea is to classify by electronic structure rather than crystal structure, but a band image has no ready-made fingerprint. A student proposed an autoencoder: compress the band image to a 64-number vector and rebuild it, training on the rebuild error; the vector is the fingerprint. He says it is a very simple convolutional network, like ResNet-18, not state of the art. Without being told anything about the materials, clustering puts small-gap semiconductors, narrow flat-band systems, intermetallics and selenides into their own clusters, and chemical likeness (the same column of the periodic table) shows up too. Changing one number in the vector slowly generates a series of band images where a gap closes and opens again; whether this can be turned around to find a material for a wanted band, he does not know, 'probably not'.

Scaling to three dimensions: too many clusters for a person, and language models do not help yet

The same pipeline run on about 120,000 three-dimensional materials in the Materials Project gives a picture 'no human researcher can analyse', where in two dimensions they could go point by point. They tried fine-tuned language models or small ones to describe each cluster, but most of the time it is gibberish, which he puts down to a lack of logic and reasoning and expects to improve within a year. The other route is to cut the data: many flat bands are 'trivial', the atoms responsible lie so far apart that they barely overlap, so the band is flat and no interesting physics is expected. They are trying to measure this with how fast the electron density falls along paths between atoms (for example a minimum spanning tree), which leaves two or three thousand materials, but the criterion is not settled.