Matej Mezera

[Jorge Bravo Abad @bravo_abad]

100 molecular geometries, nearly the training cost of 10

Stretch a molecular bond slightly, and you have a new quantum calculation to perform. Follow a molecule as it twists after absorbing light, and those calculations multiply.

But nearby geometries have much in common. How much of the computational work can they share?

Schätzle and colleagues tackle this with neural wavefunctions trained using variational quantum Monte Carlo.

A transformer learns across multiple molecular geometries, sharing weights so that progress on one configuration can help the others. The training is driven directly by quantum mechanics, without supervised pretraining on reference calculations.

For a small molecular ion, they expanded the calculation from 10 geometries to a 100-point grid spanning bond stretching and twisting. Training cost increased only marginally, while accuracy on the ten common reference geometries remained essentially unchanged.

That means a much denser picture of how the molecule’s electronic energies change, without paying for every point as an independent problem.

For scientists running expensive calculations across related configurations, sharing the learned solution across them may be as valuable as making each calculation faster.

Schätzle et al., Nature Computational Science (2026). CC BY 4.0. https://t.co/Wcu7VfOGhd

[Frank Noe @FrankNoeBerlin]

Out after a long time: Ab initio simulation of excited-state potential energy surfaces with transferable deep quantum Monte Carlo, with Zeno Schätzle, @PBerntSzab1 , Alice Cuzzocrea, @MatejMezera

https://t.co/B4SjbwNqSq

[Frank Noe @FrankNoeBerlin]

First official release of the DeepQMC code for Quantum Monte Carlo with #NeuralNetworks wave functions, including architectures like PauliNet, FermiNet and DeepErwin. By Zeno Schätzle, @PBerntSzab1, Matej Mezera and @jhrmnn

https://t.co/OwwgU7lJJv

[Mews R&D @MewsRnD]

We promised, we delivered, they built... 🚀
Lucky winner(s) of the Lego version of Saturn V we had at our booth last month at a Career day at @matfyz. Congratulations! 👏👏👏

[Matej Mezera @MatejMezera]

@Tooooooooo16 That is not true. People do not apply any moment of power by jumping. So the Earth's rotation stays constant, just the rotation of people is slowed (it means that everybody would land a bit shifted).