Table 2.

Comparison of accuracy and features between machine learning models

Model	Input type	Target type	D–MAE (Å)
Makoś et al.38	${\mathcal{C}}_R$, ${\mathcal{C}}_P$	CMa	0.170
Jackson et al.37	${\mathcal{C}}_R$, ${\mathcal{C}}_P$	Positions	0.244b
Pattanaik et al.39	${\mathcal{C}}_R$, ${\mathcal{C}}_P$, ${\mathcal{G}}_R$, ${\mathcal{G}}_P$	Distances	0.225b
Choi40	${\mathcal{C}}_R$, ${\mathcal{C}}_P$, (${\mathcal{C}}_R$+${\mathcal{C}}_P$)/2	Distances	0.095b
TSDiff	${\mathcal{G}}_R$, ${\mathcal{G}}_P$	Positions	0.137c, 0.063d, 0.067d

The input type, target type, and accuracy of the models are compared. ${\mathcal{C}}_R$ and ${\mathcal{C}}_P$ denote the geometries of reactants and products, respectively, and ${\mathcal{G}}_R$ and ${\mathcal{G}}_P$ denote the 2D graphs of reactants and products, respectively. (${\mathcal{C}}_R$+${\mathcal{C}}_P$)/2 denotes the interpolated geometry between ${\mathcal{C}}_R$ and ${\mathcal{C}}_P$.

^aCM indicates the Coulomb matrix.

^bThe values are borrowed from Choi⁴⁰.

^cThe value was evaluated without considering conformer matching, meaning that a single generated transition state (TS) for each reaction was used to evaluate the mean absolute error of interatomic distance (D–MAE).

^dThese values were calculated for TSs generated from single and eight sampling rounds, only if they matched the corresponding reference geometry after saddle point optimization, covering 53.2% and 84.6% of the test reactions, respectively.