Transferability of MACE Graph Neural Network for Range Corrected Δ-Machine Learning Potential QM/MM Applications DOI
Timothy J. Giese, Jinzhe Zeng, Darrin M. York

и другие.

The Journal of Physical Chemistry B, Год журнала: 2025, Номер unknown

Опубликована: Май 26, 2025

We previously introduced a "range corrected" Δ-machine learning potential (ΔMLP) that used deep neural networks to improve the accuracy of combined quantum mechanical/molecular mechanical (QM/MM) simulations by correcting both internal QM and QM/MM interaction energies forces [J. Chem. Theory Comput. 2021, 17, 6993-7009]. The present work extends this approach include graph networks. Specifically, is applied MACE message passing network architecture, series AM1/d + models are trained reproduce PBE0/6-31G* model phosphoryl transesterification reactions. Several designed test transferability varying amount training data calculating free energy surfaces reactions were not included in parameter refinement. compared DP use DeepPot-SE (DP) architecture. found target even instances where exhibit inaccuracies. train "end-state" only from reactant product states 6 Unlike uncorrected profiles, method correctly reproduces stable pentacoordinated phosphorus intermediate though did structures with similar bonding pattern. Furthermore, mechanism hyperparameters defining varied explore their effect on model's performance. 28% slower than when ΔMLP correction performed graphics processing unit. Our results suggest architecture may lead improved transferability.

Язык: Английский

Transferability of MACE Graph Neural Network for Range Corrected Δ-Machine Learning Potential QM/MM Applications DOI
Timothy J. Giese, Jinzhe Zeng, Darrin M. York

и другие.

The Journal of Physical Chemistry B, Год журнала: 2025, Номер unknown

Опубликована: Май 26, 2025

We previously introduced a "range corrected" Δ-machine learning potential (ΔMLP) that used deep neural networks to improve the accuracy of combined quantum mechanical/molecular mechanical (QM/MM) simulations by correcting both internal QM and QM/MM interaction energies forces [J. Chem. Theory Comput. 2021, 17, 6993-7009]. The present work extends this approach include graph networks. Specifically, is applied MACE message passing network architecture, series AM1/d + models are trained reproduce PBE0/6-31G* model phosphoryl transesterification reactions. Several designed test transferability varying amount training data calculating free energy surfaces reactions were not included in parameter refinement. compared DP use DeepPot-SE (DP) architecture. found target even instances where exhibit inaccuracies. train "end-state" only from reactant product states 6 Unlike uncorrected profiles, method correctly reproduces stable pentacoordinated phosphorus intermediate though did structures with similar bonding pattern. Furthermore, mechanism hyperparameters defining varied explore their effect on model's performance. 28% slower than when ΔMLP correction performed graphics processing unit. Our results suggest architecture may lead improved transferability.

Язык: Английский

Процитировано

0