Generative Model for Constructing Reaction Path from Initial to Final States

TL;DR

The paper tackles the challenge of generating plausible chemical reaction pathways without real-time PES evaluations by introducing a geometry-based neural framework. It combines a transformation guidance field and a denoising field to iteratively construct reaction paths from an initial state to a final state, trained on Transition1x low-energy transition paths. The model demonstrates conditional and random generation capabilities, and ablation studies highlight the essential role of the denoising field and the limits of systematic generalization. This approach offers a fast, scalable means to obtain approximate reaction pathways in organic chemistry, with potential to accelerate sampling and screening in reaction simulations.

Abstract

Mapping the chemical reaction pathways and their corresponding activation barriers is a significant challenge in molecular simulation. Given the inherent complexities of 3D atomic geometries, even generating an initial guess of these paths can be difficult for humans. This paper presents an innovative approach that utilizes neural networks to generate initial guesses for reaction pathways based on the initial state and learning from a database of low-energy transition paths. The proposed method is initiated by inputting the coordinates of the initial state, followed by progressive alterations to its structure. This iterative process culminates in the generation of the guess reaction path and the coordinates of the final state. The method does not require one-the-fly computation of the actual potential energy surface, and is therefore fast-acting. The application of this geometry-based method extends to complex reaction pathways illustrated by organic reactions. Training was executed on the Transition1x dataset of organic reaction pathways. The results revealed the generation of reactions that bore substantial similarities with the test set of chemical reaction paths. The method's flexibility allows for reactions to be generated either to conform to predetermined conditions or in a randomized manner.

Figures (11)

• Figure 1: Model image of the learned flow. Consider the center of the figure as the IS. The three curves emanating from the IS are considered optimized RPs. The FS is the endpoint of the RPs on the side opposite to the IS. The black arrows extending from each point in the figure represent the denoising field ($\mathbf{t}_\mathrm{d}$), and the red arrows represent the transformation guidance field ($\mathbf{t}_\mathrm{t}$).
• Figure 2: The whole model architecture.
• Figure 3: Diagrams of the interaction of the $E(3)$-attention network. (a) The whole diagram. (b) Details of the EdgeFeat block used in (a).
• Figure 4: Model architecture of the readout part.
• Figure 5: The mean of the norm of the difference vector between the predicted and ground truth data for both the transformation guidance field and the denoising field during the training steps. Here, train-t relates to the transformation guidance field within the training dataset, and valid-t pertains to the transformation guidance field within the validation dataset. Similarly, train-d is associated with the denoising field in the training dataset. At the same time, valid-d is connected to the denoising field in the validation dataset.