In recent years, the use of deep learning (neural network) potential energy surface (NNPES) in molecular dynamics simulation has experienced explosive growth as it can be as accurate as quantum chemistry methods while being as efficient as classical mechanic methods. However, the development of NNPES is highly non-trivial. In particular, it has been troubling to construct a dataset that is as small as possible yet can cover the target chemical space. In this work, an ESOINN-DP method is developed, which has the enhanced self-organizing incremental neural-network (ESOINN) and a newly proposed error indicator at its core. With ESOINN-DP, One can construct the NNPES with little human intervention, and this method ensures that the constructed reference dataset covers the target chemical space with minimum redundancy. The performance of the ESOINN-DP method has been well validated by developing neural network potential energy surfaces for water clusters and by de-redundancy of a sub-data set of the ANI-1 database. We believe that the ESOINN-DP method provides a novelty idea for the construction of NNPES and especially, the reference datasets, and it can be used for MD simulations of various gas-phase and condensed-phase chemical systems.

In recent years, the use of deep learning (neural network) potential energy surface (NNPES) in molecular dynamics simulation has experienced explosive growth as it can be as accurate as quantum chemistry methods while being as efficient as classical mechanic methods. However, the development of NNPES is highly non-trivial. In particular, it has been troubling to construct a dataset that is as small as possible yet can cover the target chemical space. In this work, an ESOINN-DP method is developed, which has the enhanced self-organizing incremental neural-network (ESOINN) and a newly proposed error indicator at its core. With ESOINN-DP, One can construct the NNPES with little human intervention, and this method ensures that the constructed reference dataset covers the target chemical space with minimum redundancy. The performance of the ESOINN-DP method has been well validated by developing neural network potential energy surfaces for water clusters and by de-redundancy of a sub-data set of the ANI-1 database. We believe that the ESOINN-DP method provides a novelty idea for the construction of NNPES and especially, the reference datasets, and it can be used for MD simulations of various gas-phase and condensed-phase chemical systems.

The reactive molecular dynamics is widely used in the field of computational chemistry to study the reaction mechanisms in molecular systems. However, complex trajectories that are difficult to analyze have become a major obstacle to its application in large-scale systems. In this work, a new approach named ReacNetGen is developed to obtain reaction networks based on reactive MD simulations. Molecular species can be automatically generated from the 3D coordinates of atoms in the trajectory. The hidden Markov model is used to filter the noises in the trajectory, which makes the analysis process easier and more accurate. Compared with manual analysis, the advantage of this method in terms of efficiency is very obvious for large-scale simulation trajectories. It has been successfully used in the analysis of the simulated oxidation of 4-component RP-3 and methane.

This study examined the formation mechanisms of singlet (rhombic) and triplet (linear) C4 with acetylene by using accurate ab initio CCSD(T)/cc-pVTZ/B3LYP/6-311G(d,p) calculations, followed by a kinetic analysis of various reaction pathways and computations of relative product yields in combustion and planetary atmospheres. These calculations were combined with the Rice–Ramsperger–Kassel–Marcus (RRKM) calculations of reaction rate constants for predicting product-branching ratios, which depend on the collision energy under single-collision conditions. The results show that the initial reaction begins with the formation of an intermediate t-i2, with entrance barriers of 3.8 kcal/mol, and an intermediate s-i1 without entrance barriers. On the triplet surface, the t-i2 rearranged the other C6H2 isomers, including t-i3, t-i4, and t-i6, through hydrogen migration; the t-i2, t-i3, t-i4, t-i5, and t-i6 isomers lost a hydrogen atom, and produced the most stable linear isomer of C6H, with an overall reaction exothermicity of 11 kcal/mol. Hydrogen elimination from the t-i10 isomer led to the formation of the annular C6H isomer, HC3C3 + H, at 23.9 kcal/mol above l-C4 + C2H2. On the singlet surfaces, s-i1 rearranged the other C6H2 isomers, including s-i2 and s-i4, through carbon–carbon bond cleavage. The s-i6 and s-i11 isomers also lost a hydrogen atom, and produced the linear C6H radical. Hydrogen elimination from the s-i4 isomer led to the formation of the annular C6H isomer. The s-i5 lost a hydrogen atom, and produced the six-member ring c-C6H isomer, at 2.1 kcal/mol higher than l-C4 + C2H2. The 1,1-H2 loss from the s-i10 isomer produced the linear hexacarbon l-C6 + H2 product, with an endothermicity of 2.3 kcal/mol and a 1,1-H2 loss from the s-i11 isomer, producing in the cyclic hexacarbon c-C6 + H2 product, with an exothermicity of 11.2 kcal/mol. The product-branching ratios obtained by solving kinetic equations with individual rate constants calculated using the RRKM and VTST theories for determining the collision energies between 5 kcal/mol and 25 kcal/mol show that l-C6H + H is the dominant reaction product, whereas HC3C3 + H, l-C6 + H2, c-C6H + H, and c-C6 + H2 are minor products with branching ratios. The s-i6 isomer was calculated to be the most stable C6H2 species, even more favorable than t-i3 (by 76 kcal/mol).

<p>Artificial neural network provides the possibility to develop molecular potentials with both the efficiency of the classical molecular mechanics and the accuracy of the quantum chemical methods. In this work, we developed ab initio based neural network potential (NN/MM-RESP-MBG) to perform molecular dynamics study for metalloproteins. The interaction energy, atomic forces, and atomic charges of metal binding group in NN/MM-RESP-MBG are described by a neural network potential trained with energies and forces generated from density functional calculations. Here, we used our recently proposed E-SOI-HDNN model to achieve the automatic construction of reference dataset of metalloproteins and the active learning of neural network potential functions. The predicted energies and atomic forces from the NN potential show excellent agreement with the quantum chemistry calculations. Using this approach, we can perform long time AIMD simulations and structure refinement MD simulation for metalloproteins. In 1 ns AIMD simulation of four common coordination mode of zinc-containing metalloproteins, the statistical average structure is in good agreement with statistic value of PDB Bank database. The neural network approach used in this study can be applied to construct potentials to metalloproteinase catalysis, ligand binding and other important biochemical processes and its salient features can shed light on the development of more accurate molecular potentials for metal ions in other biomacromolecule system. </p>