Development of a deep machine learning interatomic potential for metalloid-containing Pd-Si compounds

Abstract

Interatomic potentials based on neural-network machine learning (ML) approach to address the longstanding challenge of accuracy versus efficiency in molecular-dynamics simulations have recently attracted a great deal of interest. Here, utilizing Pd-Si system as a prototype, we extend the development of neural-network ML potentials to compounds exhibiting various types of bonding characteristics. The ML potential is trained by fitting to the energies and forces of both liquid and crystal structures firstprinciples calculations based on density-functional theory (DFT). We show that the generated ML potential captures the structural features and motifs in Pd82Si18 and Pd75Si25 liquids more accurately than the existing interatomic potential based on embedded-atom method (EAM). The ML potential also describes the solid-liquid interface of these systems very well. Moreover, while the existing EAM potential fails to describe the relative energies of various crystalline structures and predict wrong ground-state structures at Pd3Si and Pd9Si2 composition, the developed ML potential predicts correctly the groundstate structures from genetic algorithm search. The efficient ML potential with DFT accuracy from our study will provide a promising scheme for accurate atomistic simulations of structures and dynamics of complex Pd-Si system.

Molecular Dynamics Simulations of Molten Magnesium Chloride Using Machine‐Learning‐Based Deep Potential

“…The DP is adopted in this work since the descriptor proposed by Zhang et al [15,16] for DP is considered to be more flexible. To our knowledge, the DP has been successfully applied in the investigations of various systems such as water, [17,18] (Zr 0.2 Hf 0.2 Ti 0.2 Nb 0.2 Ta 0.2 )C high entropy materials, [19] solid-state electrolytes, [20] alloy materials, [21][22][23] etc.…”

Primary radiation damage in silicon from the viewpoint of a machine learning interatomic potential

“…On the other hand, the fact that the ML potentials calculate the energies and forces by interpolating the training data makes the them significantly faster than DFT calculations (yet noticeably slower that classical potentials) [29]. Hence, the simulation capability that has been made available by ML potentials makes larger scale molecular dynamics simulations with DFT accuracy reachable [30][31][32][33][34][35][36][37][38][39]. That is why they have received a great amount of interest in the community and their implementation in different fields matures rapidly [40][41][42][43][44].…”

Deep machine learning potentials for multicomponent metallic melts: development, predictability and compositional transferability

“…Ryltsev) chtchelkatchev@hppi.troitsk.ru ( N.M. Chtchelkatchev) ORCID(s): 0000-0003-1746-8200 ( R.E. Ryltsev); 0000-0002-7242-1483 ( N.M. Chtchelkatchev) have been recently developed [17,18,19,20,21,22,23,24,25]. However, these examples are either unary or binary systems; multicomponent systems are studied relatively weak.…”

“…It is difficult to build such universal interactions using simple models like EAM ones; more flexible (so-called mathematical) potentials are needed here. There are three major classes of regressors, which are used to build MLIPs: deep neural networks [26,23,27,28,29,15,30,31,32,33], kernel methods [34,35,36,37] and generalized linear models [38,39,40,41]. There are also promising alternative approaches, which are not widely accepted so far [13,9].…”

“…The main advantages of DeePMD are: 1) highly effective and automatic procedure to map the particles coordinates to the space of structural descriptors, which includes many-body interactions and provides invariance with respect to translations, rotations, and permutations; 2) the use of the TensorFlow platform to build models allows using GPU to effectively train DNNPs and than utilize them in large-scale molecular dynamics simulations [16]; 3) the possibility to apply active learning strategy using the DPGEN tool [28,26]. This approach has been successfully utilized for simulating systems of different nature [22,21,19,20,23,24,45,46]. However, there are no implicit receipts how to choose values of DNNP hyperparameters (such as the number of hidden layers and the number of neurons) as well as the learning parameters (such as learning rates, batch size, etc.)…”