A glass study built around 42,572 atomistic simulations tested predictions against 55 synthesised glasses. Mixing those measurements into larger simulated datasets diluted their influence: the larger hybrid set contained 2,549 simulation-derived targets alongside the 55 measurements.[1]
The team’s SCALE method instead learns the ratio between measured and predicted density, varying that multiplier with the recipe. Selection begins with uncertain predictions, then favours large corrections.[1]
What the multiplier does
Consider a made-up example. A model predicts 2.50 g/cm³ and a laboratory measurement gives 2.40. Their ratio is 0.96. Applying that multiplier to a prediction of 2.60 gives 2.496 g/cm³. These numbers illustrate the arithmetic; they are not the study’s measurements.
A recipe that was already predicted correctly should keep a factor near 1.00. Giving it the same 4% reduction would introduce an error. The useful correction must follow where the error occurs.
The physical demonstration used three batches of six new compositions. Corrections reached 0.3 g/cm³ relative to the original model; an initially flat prediction captured falling density as boron oxide increased in one glass series. The cycle had not converged after 18 measurements. Density was the tested property.[1]
Sources & further reading
- SCALE glass-density paper: published full-text mirror npj Computational Materials · 24 Jul 2026 · Accessed 20 Oct 2026