Flow for Generating Reaction Pathways and Validation of the Trained Neural Network

The reaction pathway and the corresponding activation barrier are closely related to the reaction rate of a chemical reaction. The activation barrier is considered in various scenarios of material discovery, such as studying the synthesis methods of materials and examining changes over time. Activation barriers obtained by computational chemistry using density functional theory or machine learning potentials are also expected to be useful. However, it is challenging to calculate many appropriate reaction pathways that are useful for material exploration. Due to its inherent complexity and non-linearity, estimating these pathways algorithmically and comprehensively has been difficult. In addition, since chemical reactions are intrinsically rare events, exploration based on statistical mechanics, such as molecular dynamics, has not been efficient.<br/>To solve this problem, we proposed a method using neural networks to generate an initial estimate of these reaction pathways [1]. The proposed method starts by inputting the coordinates of the initial state and gradually changing its structure. This iterative process generates an approximate representation of the reaction pathway and the coordinates of the final state. Using this method, it is possible to generate complex reaction pathways such as those seen in organic reactions.<br/>A neural network that learned a dataset including organic chemical reaction pathways was created. It was demonstrated that our neural network has the ability to generate reactions similar to the corresponding test data. Furthermore, it can generate reactions either randomly or according to predetermined conditions. Using the model trained with Transition1x, typical chemical reactions such as the Diels-Alder reaction and the rotation reaction between the NN in azobenzene could be generated in about 10-30 seconds.<br/>We have verified our proposed model works with the Transition1x dataset. To expand its applicability, we are now focusing on dataset expansion. We are now constructing an automated reaction data correction framework using an improved reaction pathway analysis algorithm and universal neural network potential PFP [2].<br/><br/>[1] A. Hayshi, et al., https://arxiv.org/abs/2401.10721<br/>[2] S. Takamoto, et al., Nat Commun 13, 2991 (2022).