Random Sampling Versus Active Learning Algorithms for Machine Learning Potentials of Quantum Liquid Water DOI
Nore Stolte, János Daru, Harald Forbert

и другие.

Journal of Chemical Theory and Computation, Год журнала: 2025, Номер unknown

Опубликована: Янв. 14, 2025

Training accurate machine learning potentials requires electronic structure data comprehensively covering the configurational space of system interest. As construction this is computationally demanding, many schemes for identifying most important structures have been proposed. Here, we compare performance high-dimensional neural network (HDNNPs) quantum liquid water at ambient conditions trained to sets constructed using random sampling as well various flavors active based on query by committee. Contrary common understanding learning, find that a given set size, leads smaller test errors not included in training process. In our analysis, show can be related small energy offsets caused bias added which overcome instead correlations an error measure invariant such shifts. Still, all HDNNPs yield very similar and structural properties water, demonstrates robustness procedure with respect algorithm even when few 200 structures. However, preliminary potentials, reasonable initial avoid unnecessary extension covered configuration less relevant regions.

Язык: Английский

Random Sampling Versus Active Learning Algorithms for Machine Learning Potentials of Quantum Liquid Water DOI
Nore Stolte, János Daru, Harald Forbert

и другие.

Journal of Chemical Theory and Computation, Год журнала: 2025, Номер unknown

Опубликована: Янв. 14, 2025

Training accurate machine learning potentials requires electronic structure data comprehensively covering the configurational space of system interest. As construction this is computationally demanding, many schemes for identifying most important structures have been proposed. Here, we compare performance high-dimensional neural network (HDNNPs) quantum liquid water at ambient conditions trained to sets constructed using random sampling as well various flavors active based on query by committee. Contrary common understanding learning, find that a given set size, leads smaller test errors not included in training process. In our analysis, show can be related small energy offsets caused bias added which overcome instead correlations an error measure invariant such shifts. Still, all HDNNPs yield very similar and structural properties water, demonstrates robustness procedure with respect algorithm even when few 200 structures. However, preliminary potentials, reasonable initial avoid unnecessary extension covered configuration less relevant regions.

Язык: Английский

Процитировано

2