Low-cost prediction of molecular and transition state partition functions via machine learning

This dataset contains the vibrational, rotational, translational, and electronic partition functions for 35,883 organic chemistry molecular structures taken from the Grambow et. al dataset [1]. It was used to train ML deep neural networks to predict unknown transition state partition functions as well as partition functions for known molecular structures [2]

The partition functions were computed at temperatures in the range T= [50, 2000] K with the rigid rotor, rigid body, harmonic oscillator approximations. Reactions involve no more than 7 C, N, or O atoms.

Frequencies for the vibrational partition functions were taken from [1] where they were computed with DFT at the ωB97X-D3/def2-TZVP level of theory.

For the rotational partition function, symmetry numbers were obtained by evaluating proper and improper invariant rotations of the structures. We note that structures involving two molecules were not separated: vibrational frequencies and symmetry numbers were computed for the aggregate structure.

For each reaction, partition functions were calculated at 50 temperatures sampled uniformly from the inverse temperature range 1/T = [1/2000, 1/50] K^-1. This corresponds to 11,961 reactions, 35,883 total structures, and 1,794,150 total partition function examples.

The file Partition_Functions.tar.gz contains directories entitled “rxnXXXXXX” where XXXXXX is a reaction number identifier. Each contain three files “rXXXXXX.csv”, “pXXXXXX.csv”, “tsXXXXXX.csv” corresponding to data from the reactant (“r”), product (“p”), and transition state (“ts”) for reaction XXXXXX. Note that the directory structure and the reaction identifiers are the same as used in the original structure dataset by Grambow et al. and the corresponding structures can easily be extracted from that dataset. Each comma separated value (csv) file contains 50 rows and the following columns:

[1]      C. A. Grambow, L. Pattanaik, and W. H. Green, “Reactants, products, and transition states of elementary chemical reactions based on quantum chemistry,” Sci. Data, 7:1–8, 2020.

[2]      Komp, E. Valleau, S. “Low-cost prediction of molecular and transition state partition functions via machine learning”, arXiv:, 2022.