Developing machine learning-based interatomic potentials from ab initio electronic structure methods remains a challenging task for computational chemistry and materials science. This work studies the capability of transfer learning, in particular discriminative fine-tuning, for efficiently generating chemically accurate interatomic neural network potentials on organic molecules from the MD17 and ANI data sets. We show that pre-training the network parameters on data obtained from density functional calculations considerably improves the sample efficiency of models trained on more accurate ab initio data. Additionally, we show that fine-tuning with energy labels alone can suffice to obtain accurate atomic forces and run large-scale atomistic simulations, provided a well-designed fine-tuning data set. We also investigate possible limitations of transfer learning, especially regarding the design and size of the pre-training and fine-tuning data sets. Finally, we provide GM-NN potentials pre-trained and fine-tuned on the ANI-1x and ANI-1ccx data sets, which can easily be fine-tuned on and applied to organic molecules.
%0 Journal Article
%1 D2CP05793J
%A Zaverkin, Viktor
%A Holzmüller, David
%A Bonfirraro, Luca
%A Kästner, Johannes
%D 2023
%I The Royal Society of Chemistry
%J Physical Chemistry Chemical Physics
%K EXC2075 PN3 PN3-4 PN6 PN6-3 PN6A-1 unclear
%N 7
%P 5383-5396
%R 10.1039/D2CP05793J
%T Transfer learning for chemically accurate interatomic neural network potentials
%U http://dx.doi.org/10.1039/D2CP05793J
%V 25
%X Developing machine learning-based interatomic potentials from ab initio electronic structure methods remains a challenging task for computational chemistry and materials science. This work studies the capability of transfer learning, in particular discriminative fine-tuning, for efficiently generating chemically accurate interatomic neural network potentials on organic molecules from the MD17 and ANI data sets. We show that pre-training the network parameters on data obtained from density functional calculations considerably improves the sample efficiency of models trained on more accurate ab initio data. Additionally, we show that fine-tuning with energy labels alone can suffice to obtain accurate atomic forces and run large-scale atomistic simulations, provided a well-designed fine-tuning data set. We also investigate possible limitations of transfer learning, especially regarding the design and size of the pre-training and fine-tuning data sets. Finally, we provide GM-NN potentials pre-trained and fine-tuned on the ANI-1x and ANI-1ccx data sets, which can easily be fine-tuned on and applied to organic molecules.
@article{D2CP05793J,
abstract = {Developing machine learning-based interatomic potentials from ab initio electronic structure methods remains a challenging task for computational chemistry and materials science. This work studies the capability of transfer learning{,} in particular discriminative fine-tuning{,} for efficiently generating chemically accurate interatomic neural network potentials on organic molecules from the MD17 and ANI data sets. We show that pre-training the network parameters on data obtained from density functional calculations considerably improves the sample efficiency of models trained on more accurate ab initio data. Additionally{,} we show that fine-tuning with energy labels alone can suffice to obtain accurate atomic forces and run large-scale atomistic simulations{,} provided a well-designed fine-tuning data set. We also investigate possible limitations of transfer learning{,} especially regarding the design and size of the pre-training and fine-tuning data sets. Finally{,} we provide GM-NN potentials pre-trained and fine-tuned on the ANI-1x and ANI-1ccx data sets{,} which can easily be fine-tuned on and applied to organic molecules.},
added-at = {2024-03-26T11:56:11.000+0100},
author = {Zaverkin, Viktor and Holzmüller, David and Bonfirraro, Luca and Kästner, Johannes},
biburl = {https://puma.ub.uni-stuttgart.de/bibtex/28d15a27baf03f8569ee071dc0bf12705/testusersimtech},
doi = {10.1039/D2CP05793J},
interhash = {19f110fb84ec7e74c5fcf2a339000daa},
intrahash = {8d15a27baf03f8569ee071dc0bf12705},
journal = {Physical Chemistry Chemical Physics},
keywords = {EXC2075 PN3 PN3-4 PN6 PN6-3 PN6A-1 unclear},
number = 7,
pages = {5383-5396},
publisher = {The Royal Society of Chemistry},
timestamp = {2024-03-26T11:56:11.000+0100},
title = {Transfer learning for chemically accurate interatomic neural network potentials},
url = {http://dx.doi.org/10.1039/D2CP05793J},
volume = 25,
year = 2023
}
