DPA-4.0-Pro-MPtrj

Version: v2026.05.20 Added: 2026-05-20 Published: 2026-05-20 32.7M parameters Missing preds: 0
Repo Paper Docs DOI Files PR Checkpoint
Predictions
κSRMEGeometry OptimizationDiscovery

ML vs DFT Formation Energies

ML vs DFT Convex Hull Distance

ML vs DFT Lattice Thermal Conductivity

Convex hull distance prediction errors projected onto elements

1 H 0.18
2 He 0.00
3 Li 0.03
4 Be 0.06
5 B 0.08
6 C 0.07
7 N 0.10
8 O 0.13
9 F 0.12
10 Ne 0.00
11 Na 0.04
12 Mg 0.05
13 Al 0.06
14 Si 0.08
15 P 0.08
16 S 0.09
17 Cl 0.11
18 Ar 0.00
19 K 0.05
20 Ca 0.05
21 Sc 0.04
22 Ti 0.05
23 V 0.08
24 Cr 0.10
25 Mn 0.13
26 Fe 0.11
27 Co 0.07
28 Ni 0.07
29 Cu 0.05
30 Zn 0.05
31 Ga 0.06
32 Ge 0.07
33 As 0.07
34 Se 0.10
35 Br 0.10
36 Kr 0.00
37 Rb 0.05
38 Sr 0.04
39 Y 0.05
40 Zr 0.06
41 Nb 0.07
42 Mo 0.06
43 Tc 0.05
44 Ru 0.08
45 Rh 0.07
46 Pd 0.06
47 Ag 0.04
48 Cd 0.04
49 In 0.07
50 Sn 0.06
51 Sb 0.07
52 Te 0.12
53 I 0.08
54 Xe 0.03
55 Cs 0.05
56 Ba 0.05
57 La 0.05
58 Ce 0.05
59 Pr 0.05
60 Nd 0.04
61 Pm 0.04
62 Sm 0.05
63 Eu 0.07
64 Gd 0.05
65 Tb 0.05
66 Dy 0.05
67 Ho 0.05
68 Er 0.05
69 Tm 0.05
70 Yb 0.05
71 Lu 0.04
72 Hf 0.05
73 Ta 0.10
74 W 0.08
75 Re 0.06
76 Os 0.08
77 Ir 0.09
78 Pt 0.07
79 Au 0.07
80 Hg 0.05
81 Tl 0.05
82 Pb 0.07
83 Bi 0.06
84 Po 0.00
85 At 0.00
86 Rn 0.00
87 Fr 0.00
88 Ra 0.00
89 Ac 0.05
90 Th 0.06
91 Pa 0.10
92 U 0.08
93 Np 0.15
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. Tiancheng Li  AI for Science Institute, Beijing; Peking University  
  2. Duo Zhang  AI for Science Institute, Beijing  
  3. Linfeng Zhang  AI for Science Institute, Beijing; DP Technology  
  4. Han Wang  Beijing Institute of Applied Physics and Computational Mathematics (IAPCM)  

Trained By

  1. Tiancheng Li  AI for Science Institute, Beijing; Peking University  

Model Info

  • Model Version v2026.05.20
  • 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

DPA-4.0-Pro-MPtrj is a DeePMD-kit universal interatomic potential in the DPA4/SeZM architecture family, trained only on the MPtrj dataset for this Matbench Discovery submission.

Architecture

DPA4 uses local-frame SO(2)-equivariant message passing with envelope-gated attention over invariant channels. The submitted model has 8 interaction blocks, lmax=5, mmax=1, 96 channels, 16 radial basis functions, and a 6.0 Å cutoff with up to 384 selected neighbors.

Training

The model was trained for 3,000,000 steps using the Muon optimizer, WSD learning-rate schedule, MAE energy/force/virial loss weights 20/20/5, with AMP, TF32, and compile enabled.

Hyperparameters

  • max_force: 0.02
  • max_steps: 500
  • ase_optimizer: "FIRE"
  • cell_filter: "FrechetCellFilter"
  • graph_construction_radius: 6
  • max_neighbors: 384
  • architecture: {"type":"DPA4/SeZM","n_blocks":8,"lmax":5,"mmax":1,"channels":96,"n_radial":16,"so2_layers":4,"radial_so2_mode":"none","radial_so2_rank":0,"n_focus":1,"n_atten_head":1,"ffn_blocks":1,"sandwich_norm":[true,false,true,false],"s2_activation":[false,true],"lebedev_quadrature":false,"activation_function":"silu","glu_activation":true,"precision":"float32"}
  • learning_rate_schedule: {"type":"WSD"}
  • loss: {"loss_func":"mae","f_use_norm":true,"loss_prefactor":{"energy":20,"force":20,"virial":5}}
  • optimizer: "Muon"
  • weight_decay: 0.001
  • training: {"numb_steps":3000000,"batch_size":500,"gradient_max_norm":5}

Dependencies