. 2025 Sep 15;4(11):3208–3216. doi: 10.1039/d5dd00240k

Table 2. MAE and RMSE for the five descriptors tested. For each descriptor, a hyperparameter search using Bayesian optimization with at least 200 individual runs was performed. All values are in kcal mol⁻¹.

Architecture	Model	MAE	RMSE
no positional information		3.58	6.68
SOAP		4.74	6.99
aniAEV		3.93	6.09
EquiformerV2	omat	3.21	4.50
MACE	mace-mp0	2.44	4.09
MACE	mace-off	2.46	4.17

Table 2. MAE and RMSE for the five descriptors tested. For each descriptor, a hyperparameter search using Bayesian optimization with at least 200 individual runs was performed. All values are in kcal mol−1.