Table 1.

Molecular property prediction performances of several models on QM9 dataset in terms of MAE

Property	Unit	SchNet	DimeNet	SphereNet	CFFN	Targeta
μ	D	0.0802(0.0300)	0.0542(0.0029)	0.0593(0.0020)	0.0415 (0.0003)	0.1000
α	a03	0.1573(0.0100)	0.0955(0.0100)	0.1070(0.0016)	0.0613 (0.0050)	0.1000
△e	eV	0.0928(0.0080)	0.0870(0.0060)	0.0857(0.0100)	0.0741 (0.0005)	0.0430
Ehomo	eV	0.0752(0.0030)	0.0682(0.0050)	0.0460(0.0040)	0.0301 (0.0020)	0.0430
Elumo	eV	0.0589(0.0030)	0.0465(0.0030)	0.0381(0.0020)	0.0343 (0.0019)	0.0430
ZPVE	eV	0.0033(0.0030)	0.0091(0.0003)	0.0020(0.0002)	0.0019 (0.0001)	0.0012
U0	eV	0.0237(0.0030)	0.0186(0.0024)	0.0164(0.0002)	0.0151 (0.0017)	0.0430
U	eV	0.0272(0.0026)	0.0209(0.0024)	0.0187(0.0002)	0.0163 (0.0006)	/
H	eV	0.0309(0.0022)	0.0162(0.0010)	0.0165(0.0012)	0.0153 (0.0009)	/
G	eV	0.0265(0.0020)	0.0162(0.0016)	0.0174(0.0010)	0.0157 (0.0008)	/
Cv	cal mol−1 K−1	0.0769(0.0050)	0.0462(0.0030)	0.0417(0.0029)	0.0413 (0.0020)	0.0500

^aThe target accuracies are taken from [48]. SDs are in brackets. The SOTA results are shown in bold.