Table 2. Average relative absolute error Δτ (see eqn (1)) of exciton transfer times computed with HEOM and either, predicted by the trained neural networks (with/without PCA selection) or computed with secular Redfield. For all four datasets, we show the results of the training, validation, and test set separately. Smallest errors for each dataset are printed in bold.

Dataset	Model	Δτtrain [%]	Δτvalid [%]	Δτtest [%]
FMO	Network (PCA)	4.53	4.38	7.41
	Network	10.53	10.75	11.56
	Redfield	9.70	9.96	9.60
RC	Network (PCA)	2.71	2.73	3.35
	Network	3.61	3.58	3.76
	Redfield	8.62	8.67	8.60
CP43	Network (PCA)	4.42	4.47	4.72
	Network	4.66	4.71	4.86
	Redfield	4.71	4.66	4.73
CP47	Network (PCA)	12.36	12.32	12.59
	Network	13.36	13.34	13.59
	Redfield	10.48	10.47	10.51