Table 2. Transfer learning and fine-tuning of SorbNet pre-trained on MFI-C5-W system.

Sorption system	Initialization	Branched layers	Training MSE a (×10–4)	Test MSE a (×10–4)
MFI-C4-W	Pre-trained	Fixed	2.9 ± 0.6	5.0 ± 1.2
	Pre-trained	Trainable	2.6 ± 0.5	4.6 ± 1.1
	Random	Trainable	2.5 ± 0.4	4.5 ± 0.7
	Random	Fixed	105 ± 6	113 ± 8
MFI-C5-E	Pre-trained	Fixed	2.5 ± 0.5	6.4 ± 1.4
	Pre-trained	Trainable	1.6 ± 0.6	4.4 ± 1.4
	Random	Trainable	1.9 ± 0.4	5.6 ± 0.9
	Random	Fixed	150 ± 3	155 ± 5
LTA-C5-W	Pre-trained	Fixed	3.5 ± 0.3	9.2 ± 0.7
	Pre-trained	Trainable	2.8 ± 0.5	7.5 ± 1.5
	Random	Trainable	2.8 ± 0.6	7.6 ± 1.6
	Random	Fixed	(2.0 ± 1.7) × 102	(2.2 ± 1.7) × 102

^aStandard deviations were measured from 8 training runs. 8 models independently pre-trained on MFI-C5-W system used as initialization in transfer learning experiments.