ComENet hyperparameters

	A	B	C	D	E
1	For ComENet	OC20: Different from the values listed in the paper, as we retrain our model more times and made a submission to the OCP leaderboard
2
3		OC20	Molecule3D	QM9
4	Number of layers	4	6, 8	4
5	Cutoff	6.0	5.0	5.0
6	Hidden dim	256	256	256
7	Hidden dim in Self-Atom layer		256	256
8	Number of layers in Self-Atom layer		3	3
9	Number of layers of the MLP in Interaction layer		3	3
10	Distance embedding dim	3	6	12
11	Angle embedding dim	2	3	3
12
13
14		OC20	Molecule3D	QM9
15	Epochs	50	250	1000
16	Batch size	32	256	32
17	Learning rate	1e-3	2e-4	5e-4
18	Learning rate decay factor		0.5	0.5
19	Learning rate decay epochs		30	200
20
21
22
23	For baseline methods on Molecule3D
24		GIN-Virtual	SchNet	DimeNet++	SphereNet
25	Number of layers	5	6	4	4
26	Cutoff	-	10.0	6.0	6.0
27	Hidden dim	600	256	128	128
28
29
30		GIN-Virtual	SchNet	DimeNet++	SphereNet
31	Epochs	250	250	100	100
32	Batch size	256	256	32	32
33	Learning rate	1e-3	5e-4	5e-4	5e-4
34	Learning rate decay factor	0.25	0.5	0.5	0.5
35	Learning rate decay epochs	30	30	10	10
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100