More data isn’t the same as better data
A new preprint argues that the largest materials datasets are 99% redundant. It backs the way we build Quantanomous: fewer, well-chosen structures, computed with more accurate methods.
A new preprint, “Atoms that Matter” by Moerman and colleagues, argues that the largest materials datasets are 99% redundant, and that local chemistry can pick out the 1% a model actually needs.
It fits a growing body of work. Li and colleagues (Nature Communications, 2023) removed up to 95% of training data with little loss of in-distribution accuracy, and the compact MAD dataset, with under 100,000 structures, trains universal potentials that rival those built on 100 to 1,000 times more data.
There are vastly many possible compounds, but far fewer kinds of local bonding environments, and those are what a machine-learned potential has to learn.
Spend compute on accuracy, not volume
That is the principle we are building Quantanomous on: fewer, well-chosen structures, computed with higher-level methods than standard density functional theory, whose systematic errors a model would otherwise learn, and a calibrated uncertainty on every prediction.

