DFT data for "Molecular Simulations with a Pretrained Neural Network and Universal Pairwise Force Fields"

This repository contains quantum mechanical datasets used for training and testing the SO3LR machine-learned force field model in Molecular Simulations with a Pretrained Neural Network and Universal Pairwise Force Fields. The combined training set, provided in extxyz format and compressed as so3lr_train.tar.gz, includes GEMS general protein fragments, QM7-X small organic molecules, AQM large drug-like molecules, SPICE dipeptides, and DES15k dimers (see the "T – Optimization on diverse training data" subsection of the original manuscript for details). To maintain consistent references, the last two datasets were recomputed at the PBE0+MBD/tight level of theory (see FHI-aims input file control.in). Test sets from Table I of the original manuscript are available in so3lr_test.tar.gz. Structures from the MD22 and TorsionNet500 datasets were also recomputed at the PBE0+MBD/tight level of theory.