EquFlash

Version: v2025.06.23 Added: 2025-06-23 Published: 2025-06-23 28.7M parameters Missing preds: 0
Repo Paper Files PR
Predictions
κSRMEGeometry OptimizationDiscovery

Convex hull distance prediction errors projected onto elements

1 H 0.12
2 He 0.00
3 Li 0.02
4 Be 0.02
5 B 0.05
6 C 0.05
7 N 0.07
8 O 0.10
9 F 0.10
10 Ne 0.00
11 Na 0.02
12 Mg 0.03
13 Al 0.04
14 Si 0.05
15 P 0.04
16 S 0.06
17 Cl 0.08
18 Ar 0.00
19 K 0.03
20 Ca 0.03
21 Sc 0.02
22 Ti 0.03
23 V 0.06
24 Cr 0.08
25 Mn 0.10
26 Fe 0.09
27 Co 0.04
28 Ni 0.03
29 Cu 0.03
30 Zn 0.03
31 Ga 0.04
32 Ge 0.04
33 As 0.04
34 Se 0.07
35 Br 0.07
36 Kr 0.00
37 Rb 0.03
38 Sr 0.03
39 Y 0.03
40 Zr 0.03
41 Nb 0.04
42 Mo 0.04
43 Tc 0.03
44 Ru 0.04
45 Rh 0.04
46 Pd 0.04
47 Ag 0.03
48 Cd 0.02
49 In 0.04
50 Sn 0.04
51 Sb 0.05
52 Te 0.10
53 I 0.05
54 Xe 0.02
55 Cs 0.03
56 Ba 0.03
57 La 0.03
58 Ce 0.03
59 Pr 0.03
60 Nd 0.02
61 Pm 0.03
62 Sm 0.03
63 Eu 0.06
64 Gd 0.04
65 Tb 0.03
66 Dy 0.03
67 Ho 0.03
68 Er 0.03
69 Tm 0.03
70 Yb 0.04
71 Lu 0.03
72 Hf 0.03
73 Ta 0.06
74 W 0.04
75 Re 0.03
76 Os 0.04
77 Ir 0.05
78 Pt 0.05
79 Au 0.05
80 Hg 0.03
81 Tl 0.03
82 Pb 0.05
83 Bi 0.03
84 Po 0.00
85 At 0.00
86 Rn 0.00
87 Fr 0.00
88 Ra 0.00
89 Ac 0.03
90 Th 0.03
91 Pa 0.04
92 U 0.05
93 Np 0.09
94 Pu 0.19
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. Hyuntae Cho  Materials AI Lab at Samsung Electronics      
  2. Saerom Choi  Materials AI Lab at Samsung Electronics    
  3. Heejae Kim  Materials AI Lab at Samsung Electronics  
  4. Jaehee Jang  Materials AI Lab at Samsung Electronics  
  5. Gunhee Kim  Materials AI Lab at Samsung Electronics  
  6. Heesun Lee  Materials AI Lab at Samsung Electronics  
  7. Hyunwoo Lee  Materials AI Lab at Samsung Electronics  
  8. Yongdeok Kim  Materials AI Lab at Samsung Electronics    

Model Info

  • Model Version v2025.06.23
  • Model Type UIP
  • Targets EFSG
  • Openness CSOD
  • Train Task S2EFS
  • Test Task IS2RE-SR
  • Trained for Benchmark No

Training Set

OMat24: 101M structures from 3.23M materials

MPtrj: 1.58M structures from 146k materials

Subsampled Alexandria: 10.4M structures from 3.23M materials

Description

EquFlash is an E(3)-equivariant model based on the SevenNet-0 architecture, with tensor products accelerated by FlashTP. FlashTP achieves up to 41.6× and 60.8× kernel speedups over e3nn and NVIDIA cuEquivariance, respectively, while reducing memory usage by 6×. Leveraging these gains, we scaled EquFlash to a larger capacity than the original SevenNet-0.

Training

EquFlash, a scaled-up model derived from SevenNet-0 and accelerated with FlashTP, was pretrained on OMat24 and finetuned on MPtrj and sAlex.

Hyperparameters

  • max_force: 0.02
  • max_steps: 500
  • ase_optimizer: "FIRE"
  • cell_filter: "FrechetCellFilter"
  • graph_construction_radius: 6
  • max_neighbors: null

Dependencies