Fig. 1: Schematic representations of machine learning (ML) models.
From: Prediction of transition state structures of gas-phase chemical reactions via machine learning
a Overall ML procedure utilizing reactant, product, and linearly interpolated structures. Two different readout layers derive the molecular properties (molecular energy and entropy values, E and S) and the ratio between interatomic distances of linearly interpolated structures, \({d}_{ij}^{I}\), and transition state, \({d}_{ij}^{TS}\), from the pair features of three structures (\({f}_{ij}^{R,pair}\): reactant, \({f}_{ij}^{I,pair}\): linearly interpolated structure, \({f}_{ij}^{P,pair}\): product). b Illustration of the pair sequence interaction layer consisting of transformer encoder and bidirectional gated recurrent unit(GRU), which are responsible for intermolecular interaction among three structures and interatomic interactions in each structure, respectively. c Predicted interatomic distances are used to reconstruct 3D atomic positions, X*, through nonlinear optimization. For ensemble predictions, the results from multiple models can be utilized to perform a single nonlinear optimization.
