CGCNN+P

Version: v0.1.0 Added: 2023-02-03 Published: 2022-02-28 128k parameters Ensemble of 10 models Missing preds: 2 (0.001%)
Repo Paper DOI Files PR Checkpoint
Predictions
Discovery

ML vs DFT Formation Energies

ML vs DFT Convex Hull Distance

Convex hull distance prediction errors projected onto elements

1 H 0.50
2 He 0.00
3 Li 0.15
4 Be 0.26
5 B 0.35
6 C 0.29
7 N 0.41
8 O 0.49
9 F 0.50
10 Ne 0.00
11 Na 0.18
12 Mg 0.19
13 Al 0.32
14 Si 0.38
15 P 0.29
16 S 0.41
17 Cl 0.42
18 Ar 0.00
19 K 0.19
20 Ca 0.20
21 Sc 0.22
22 Ti 0.35
23 V 0.43
24 Cr 0.28
25 Mn 0.33
26 Fe 0.38
27 Co 0.34
28 Ni 0.29
29 Cu 0.21
30 Zn 0.24
31 Ga 0.25
32 Ge 0.29
33 As 0.27
34 Se 0.33
35 Br 0.38
36 Kr 0.00
37 Rb 0.18
38 Sr 0.19
39 Y 0.20
40 Zr 0.26
41 Nb 0.32
42 Mo 0.26
43 Tc 0.17
44 Ru 0.28
45 Rh 0.25
46 Pd 0.22
47 Ag 0.17
48 Cd 0.18
49 In 0.23
50 Sn 0.24
51 Sb 0.23
52 Te 0.32
53 I 0.32
54 Xe 0.01
55 Cs 0.18
56 Ba 0.19
57 La 0.20
58 Ce 0.18
59 Pr 0.17
60 Nd 0.17
61 Pm 0.16
62 Sm 0.17
63 Eu 0.16
64 Gd 0.17
65 Tb 0.19
66 Dy 0.19
67 Ho 0.19
68 Er 0.19
69 Tm 0.20
70 Yb 0.18
71 Lu 0.19
72 Hf 0.25
73 Ta 0.38
74 W 0.22
75 Re 0.24
76 Os 0.25
77 Ir 0.29
78 Pt 0.26
79 Au 0.22
80 Hg 0.15
81 Tl 0.18
82 Pb 0.22
83 Bi 0.18
84 Po 0.00
85 At 0.00
86 Rn 0.00
87 Fr 0.00
88 Ra 0.00
89 Ac 0.15
90 Th 0.31
91 Pa 0.19
92 U 0.26
93 Np 0.73
94 Pu 0.29
95 Am 0.00
96 Cm 0.00
97 Bk 0.00
98 Cf 0.00
99 Es 0.00
100 Fm 0.00
101 Md 0.00
102 No 0.00
103 Lr 0.00
104 Rf 0.00
105 Db 0.00
106 Sg 0.00
107 Bh 0.00
108 Hs 0.00
109 Mt 0.00
110 Ds 0.00
111 Rg 0.00
112 Cn 0.00
113 Nh 0.00
114 Fl 0.00
115 Mc 0.00
116 Lv 0.00
117 Ts 0.00
118 Og 0.00
57-71 La-Lu Lanthanides
89-103 Ac-Lr Actinides

Model Authors

  1. Jason B. Gibson  University of Florida  
  2. Ajinkya C. Hire  University of Florida  
  3. Richard G. Hennig  University of Florida  

Trained By

  1. Janosh Riebesell  University of Cambridge, Lawrence Berkeley National Laboratory  

Model Info

  • Model Version v0.1.0
  • Model Type GNN
  • Targets E
  • Openness OSOD
  • Train Task S2RE
  • Test Task IS2RE
  • Trained for Benchmark Yes

Training Set

MP v2022.10.28: 155k structures

Description

This work proposes simple structure perturbations to augment CGCNN's training data of relaxed structures with randomly perturbed ones resembling unrelaxed structures that are mapped to the same DFT final energy during training. Step function PES

Long

The model is essentially taught the potential energy surface (PES) is a step-function that maps each valley to its local minimum. The expectation is that during testing on unrelaxed structures, the model will predict the energy of the nearest basin in the PES. The authors confirm this by demonstrating a lowering of the energy error on unrelaxed structures.

Hyperparameters

  • Perturbations: 5
  • graph_construction_radius: 5
  • max_neighbors: null

Dependencies