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 0.115
2 He 0
3 Li 0.0223
4 Be 0.0191
5 B 0.0493
6 C 0.0454
7 N 0.0669
8 O 0.0968
9 F 0.0961
10 Ne 0
11 Na 0.0224
12 Mg 0.0307
13 Al 0.038
14 Si 0.0497
15 P 0.0434
16 S 0.0578
17 Cl 0.0784
18 Ar 0
19 K 0.0279
20 Ca 0.0318
21 Sc 0.0249
22 Ti 0.0307
23 V 0.0591
24 Cr 0.0784
25 Mn 0.104
26 Fe 0.0851
27 Co 0.0406
28 Ni 0.0334
29 Cu 0.0317
30 Zn 0.0308
31 Ga 0.038
32 Ge 0.0413
33 As 0.0416
34 Se 0.0738
35 Br 0.0662
36 Kr 0
37 Rb 0.0302
38 Sr 0.0288
39 Y 0.0331
40 Zr 0.0305
41 Nb 0.0407
42 Mo 0.0432
43 Tc 0.0316
44 Ru 0.0426
45 Rh 0.0415
46 Pd 0.0377
47 Ag 0.0287
48 Cd 0.0237
49 In 0.044
50 Sn 0.0366
51 Sb 0.0456
52 Te 0.0995
53 I 0.0499
54 Xe 0.022
55 Cs 0.0276
56 Ba 0.0288
57 La 0.0271
58 Ce 0.027
59 Pr 0.0279
60 Nd 0.0249
61 Pm 0.0287
62 Sm 0.0264
63 Eu 0.0566
64 Gd 0.0378
65 Tb 0.0264
66 Dy 0.0284
67 Ho 0.025
68 Er 0.0252
69 Tm 0.0251
70 Yb 0.043
71 Lu 0.0264
72 Hf 0.0315
73 Ta 0.0628
74 W 0.0379
75 Re 0.0332
76 Os 0.0425
77 Ir 0.051
78 Pt 0.0487
79 Au 0.0505
80 Hg 0.0262
81 Tl 0.0285
82 Pb 0.0474
83 Bi 0.0316
84 Po 0
85 At 0
86 Rn 0
87 Fr 0
88 Ra 0
89 Ac 0.0279
90 Th 0.0338
91 Pa 0.0392
92 U 0.0464
93 Np 0.0863
94 Pu 0.192
95 Am 0
96 Cm 0
97 Bk 0
98 Cf 0
99 Es 0
100 Fm 0
101 Md 0
102 No 0
103 Lr 0
104 Rf 0
105 Db 0
106 Sg 0
107 Bh 0
108 Hs 0
109 Mt 0
110 Ds 0
111 Rg 0
112 Cn 0
113 Nh 0
114 Fl 0
115 Mc 0
116 Lv 0
117 Ts 0
118 Og 0
57-71 La-Lu Lanthanides
89-103 Ac-Lr Actinides

Model Authors

  1. Hyuntae Cho  Materials AI Lab at Samsung Electronics  Materials AI Lab at Samsung Electronics logo    
  2. Saerom Choi  Materials AI Lab at Samsung Electronics  Materials AI Lab at Samsung Electronics logo  
  3. Heejae Kim  Materials AI Lab at Samsung Electronics  Materials AI Lab at Samsung Electronics logo
  4. Jaehee Jang  Materials AI Lab at Samsung Electronics  Materials AI Lab at Samsung Electronics logo
  5. Gunhee Kim  Materials AI Lab at Samsung Electronics  Materials AI Lab at Samsung Electronics logo
  6. Heesun Lee  Materials AI Lab at Samsung Electronics  Materials AI Lab at Samsung Electronics logo
  7. Hyunwoo Lee  Materials AI Lab at Samsung Electronics  Materials AI Lab at Samsung Electronics logo
  8. Yongdeok Kim  Materials AI Lab at Samsung Electronics  Materials AI Lab at Samsung Electronics logo  

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