The Journal of Physical Chemistry Letters, Год журнала: 2025, Номер unknown, С. 4907 - 4920

Опубликована: Май 9, 2025

We present an efficient and reliable large-scale non-adiabatic dynamics simulation method based on machine learning Hamiltonian force field. The quasi-diabatic network (DHNet) is trained in the Wannier basis well-designed translation rotation invariant structural descriptors, which can effectively capture both local nonlocal environmental information. Using representative two-dimensional transition metal dichalcogenide MoS2 as illustration, we show that density functional theory (DFT) calculations of only ten structures are sufficient to generate training set for DHNet due high efficiency analysis orbital classification sampling interorbital couplings. demonstrates good transferability, thus enabling direct construction electronic matrices large systems. Compared with DFT calculations, significantly reduces computational cost by about 5 orders magnitude. By combining DeePMD field, successfully simulate electron transport monolayer up 3675 atoms 13475 levels using a state-of-the-art surface hopping method. mobility calculated be 110 cm2/(V s), agreement extensive experimental results range 3-200 s) during 2013-2023. Due performance, proposed methods have great potential applied study charge carrier wide material

Язык: Английский