关键词: 反应过渡态, 反应物几何构象, 人工神经网络, 反应势能面, 一阶鞍点, 交叉验证

Abstract:

We found an alternative method for the derivation of transition state structure energy in chemical reactions which would be less dependent on the starting geometry of reactants by combining a mathematical tool and artificial neural networks (ANN) with conventional transition state optimization algorithms. When two reactants approach each other, every geometric structure corresponds to a system energy value. The purpose of this investigation was to collect as many energy values on the reaction energy surface as possible. By simulating the energy surface using the geometric parameters as independent variables, the first order saddle point in the energy surface corresponding to the transition state structure was derived. The nucleophilic attack step of a classical Aldol reaction was studied using acetaldehyde anion and formaldehyde as reactants. The intrinsic reaction coordinate (IRC) path calculation started with 3 different sets of starting reactant geometries and 96 points on the reaction energy surface were derived. The energy surface was simulated using ANN. Cross-validation was applied to evaluate the result and avoided a possible overfitting of the ANN.

Key words: Reaction transition state, Geometry of reactant, Artificial neural network, Reaction energy surface, First order saddle point, Cross-validation