Nequix MP

Version: 0.1.0 Added: 2025-08-17 Published: 2025-08-17 708k parameters Missing preds: 0
Repo Paper Docs DOI Files PR Checkpoint
Predictions
κSRMEGeometry OptimizationDiscovery

Convex hull distance prediction errors projected onto elements

1 H 0.29
2 He 0
3 Li 0.0499
4 Be 0.117
5 B 0.134
6 C 0.108
7 N 0.151
8 O 0.161
9 F 0.169
10 Ne 0
11 Na 0.0606
12 Mg 0.0671
13 Al 0.11
14 Si 0.122
15 P 0.127
16 S 0.14
17 Cl 0.162
18 Ar 0
19 K 0.0679
20 Ca 0.0715
21 Sc 0.0811
22 Ti 0.0924
23 V 0.112
24 Cr 0.135
25 Mn 0.157
26 Fe 0.153
27 Co 0.0944
28 Ni 0.0909
29 Cu 0.0695
30 Zn 0.0779
31 Ga 0.103
32 Ge 0.108
33 As 0.104
34 Se 0.15
35 Br 0.15
36 Kr 0
37 Rb 0.0697
38 Sr 0.0682
39 Y 0.0796
40 Zr 0.0962
41 Nb 0.122
42 Mo 0.11
43 Tc 0.0974
44 Ru 0.129
45 Rh 0.106
46 Pd 0.0932
47 Ag 0.0611
48 Cd 0.063
49 In 0.105
50 Sn 0.101
51 Sb 0.0968
52 Te 0.166
53 I 0.134
54 Xe 0.128
55 Cs 0.0734
56 Ba 0.0675
57 La 0.0685
58 Ce 0.0727
59 Pr 0.0674
60 Nd 0.0669
61 Pm 0.0633
62 Sm 0.0681
63 Eu 0.0848
64 Gd 0.0715
65 Tb 0.0693
66 Dy 0.0734
67 Ho 0.0709
68 Er 0.0735
69 Tm 0.075
70 Yb 0.0709
71 Lu 0.0741
72 Hf 0.0959
73 Ta 0.154
74 W 0.109
75 Re 0.11
76 Os 0.118
77 Ir 0.131
78 Pt 0.109
79 Au 0.105
80 Hg 0.0616
81 Tl 0.0752
82 Pb 0.0987
83 Bi 0.0887
84 Po 0
85 At 0
86 Rn 0
87 Fr 0
88 Ra 0
89 Ac 0.0681
90 Th 0.0971
91 Pa 0.123
92 U 0.103
93 Np 0.152
94 Pu 0.361
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. Teddy Koker  Massachusetts Institute of Technology  Massachusetts Institute of Technology logo    
  2. Tess Smidt  Massachusetts Institute of Technology  Massachusetts Institute of Technology logo    

Trained By

  1. Teddy Koker (Massachusetts Institute of Technology)

Model Info

  • Model Version 0.1.0
  • Model Type UIP
  • Targets EFSG
  • Openness OSOD
  • Train Task S2EFS
  • Test Task IS2RE-SR
  • Trained for Benchmark Yes

Training Set

MPtrj: 1.58M structures from 146k materials

Description

Nequix is a simplified version of NequIP with added RMSnorm, implemented in JAX, and trained with Muon optimizer.

Hyperparameters

  • max_force: 0.02
  • max_steps: 500
  • ase_optimizer: "FIRE"
  • cell_filter: "FrechetCellFilter"
  • hidden_irreps: "128x0e+64x1o+32x2e+32x3o"
  • graph_construction_radius: 6
  • max_neighbors: null
  • lmax: 3
  • radial_basis_size: 8
  • radial_mlp_size: 64
  • radial_mlp_layers: 2
  • radial_polynomial_p: 6
  • layer_norm: true
  • optimizer: "muon"
  • weight_decay: 0.001
  • warmup_epochs: 0.1
  • warmup_factor: 0.2
  • grad_clip_norm: 100
  • batch_size: 256
  • n_epochs: 100
  • loss: {"energy":"mae","force":"l2","stress":"mae"}
  • loss_weights: {"energy":20,"force":20,"stress":5}
  • ema_decay: 0.999

Dependencies