Inorganic crystals are periodic, highly-symmetric arrangements of atoms in three-dimensional space. Their structures are constrained by the symmetry operations of a crystallographic space group and restricted to lie in specific affine subspaces known as Wyckoff positions. The frequency an atom appears in the crystal and its rough positioning are determined by its Wyckoff position. Most generative models that predict atomic coordinates overlook these symmetry constraints, leading to unrealistically high populations of proposed crystals exhibiting limited symmetry. We introduce Space Group Conditional Flow Matching, a novel generative framework that samples significantly closer to the target population of highly-symmetric, stable crystals. We achieve this by conditioning the entire generation process on a given space group and set of Wyckoff positions; specifically, we define a conditionally symmetric noise base distribution and a group-conditioned, equivariant, parametric vector field that restricts the motion of atoms to their initial Wyckoff position. Our form of group-conditioned equivariance is achieved using an efficient reformulation of group averaging tailored for symmetric crystals. Importantly, it reduces the computational overhead of symmetrization to a negligible level. We achieve state of the art results on crystal structure prediction and de novo generation benchmarks. We also perform relevant ablations.
Figures & tables
Figure 1: The main components of SGFM. (a) Wyckoff position noise prior. General points are sampled randomly and projected according to the conditioned Wyckoff positions. (b) Space group equivariant vector field. The equivariance of the model combined with the G -symmetry of the input crystal ensures that atoms preserve their symmetry structure. (c) Comparison between standard group averaging and our optimized formulation. The G -symmetry of the input crystal allows space group operations to be replaced by their corresponding permutations (denoted using σ ). When combined with the Sn -equivariance of u , this reduces the number of required forward passes per crystal. For visualization purposes, heavy arrows indicate expensive model forward passes.
Figure 2: A 2 D example of a unit cell with p4mm symmetry. Applying a group element to this set permutes “atoms” of the same type (shape and color). The symmetry divides the unit cell (black box) into four Wyckoff positions: the center, horizontal and vertical coordinate axes, diagonal axes, and general position (white space).
Figure 3: Projection of a coordinate f using Wyckoff position w .
Model
NFE
Validity (%)↑
Property ↓
Stable (%)↑
S.U.N. (%)↑
S.S.U.N. (%)↑
Stable (%)↑
S.U.N. (%)↑
S.S.U.N. (%)↑
Structural
Composition
dρ
delem
dcn
Ehull<100 meV/Atom
Ehull<0 meV/Atom
CDVAE
5000
100.00
86.70
0.688
0.278
-
-
-
-
-
-
-
ADiT
500
99.74
92.14
-
-
-
72.0
27.4
18.21
13.0
4.6
3.57
FlowMM
500
96.86
83.24
0.075
0.079
0.443
31.2
19.7
14.38
4.6
2.3
1.85
FlowLLM
250
99.81
89.05
0.660
0.090
-
67.9
21.9
20.18
14.2
3.6
3.27
MatterGen
1000
100.00
82.60
0.206
0.242
-
-
24.3
-
-
-
-
Table 1: DNG evaluation. Models were trained on the MP- 20 dataset. NFE refers to the number of generation steps. We evaluated all the S.U.N. metrics, except those marked with ∗ . DiffCSP++ (empirical) , SGEquiDiff results were computed with only 1,000 samples; others used 10,000 samples. Methods above the dividing line do not explicitly use Wyckoff positions; methods below use them.
Model
Wyckoff Positions
MP- 20
MPTS- 52
Perov- 5
Carbon- 24
Alex-MP- 20
MR (%) ↑
RMSE ↓
MR (%) ↑
RMSE ↓
MR (%) ↑
RMSE ↓
MR (%) ↑
RMSE ↓
MR (%) ↑
RMSE ↓
CDVAE
None
33.90
.1045
5.34
.2106
45.31
.1138
17.09
.2969
-
-
DiffCSP
51.49
.0631
12.19
.1786
52.02
.0760
17.54
.2759
-
-
FlowMM
61.39
.0566
17.54
.1726
53.15
.0992
23.47
.4122
-
-
OMatG
69.83
.0741
27.38
.1970
83.06
.3753
-
-
72.50
.1260
DiffCSP++
Ground truth
80.27
.0295
46.29
.0896
98.44
.0430
-
-
83.18
.0188
Table 2: Crystal Structure Prediction. MR denotes match rate. Best results in bold. Models below the dashed line receive ground-truth space-group and Wyckoff information; comparisons across the line therefore involve different conditioning information.
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 4: 2 D visualization of the deformed geometry induced by moving to fractional coordinates in a non-orthogonal lattice basis. This example demonstrates how a 3 -fold rotational space group becomes a set of special affine transformations when acting on fractional coordinates and, specifically, the mapping of each rotation element to a corresponding special affine transformation. In addition, the figure illustrates how a space group induced orbit transforms under this change of basis. This new representation makes it difficult to use models that rely on Euclidean geometry.
Crystal Family
Space Group No.
Lattice Shape
Constraint of Symmetric Bases
Triclinic
1∼2
No Constraint
No Constraint
Monoclinic
3∼15
α=γ=90∘
k1=k3=0
Orthorhombic
16∼74
α=β=γ=90∘
k1=k2=k3=0
Tetragonal
75∼142
α=β=γ=90∘
k1=k2=k3=0
a=b
k4=0
Hexagonal
143∼194
α=β=90∘,γ=120∘
k2=k3=0,k1=−log(3)/4
Appendix
Table 3: Relationship between the lattice shape and the constraint of the symmetric bases.
Dataset
Number of Layers
d
dt
ds
Activation
Layer Norm
CSP
MP- 20
8
512
256
128
SiLU
✓
MPTS- 52
8
512
256
128
SiLU
✓
Carbon- 24
8
512
256
128
SiLU
✓
Perov- 5
8
512
256
128
SiLU
✓
Alex-MP- 20
8
512
256
128
SiLU
✓
Appendix
Table 4: Hyperparameter details for all the models reported in the paper. Hyperparameter (bottom row) search was conducted on the MP- 20 dataset.
Dataset
Batch Size/GPU
Learning Rate
Epochs
λF
λA
λk
sF′
sk′
CSP
MP- 20
64
0.0005
5000
100
-
1
3
3
MPTS- 52
32
0.0005
5000
100
-
1
3
3
Carbon- 24
64
0.0005
8000
100
-
1
3
3
Perov- 5
256
0.0005
1000
100
-
1
3
3
Alex-MP- 20
32
0.0003
1250
100
-
1
5
3
Appendix
Table 5: Training hyperparameter details for all the models reported in the paper. Hyperparameter (bottom row) search was conducted per experiment.
Model
Generation Time (s/batch)
64
256
500
DiffCSP++
66
268
490
SGEquiDiff
19
75
154
SGFM
22
92
175
Appendix
Table 6: Generation times for various batch sizes.
Model
MP- 20
MPTS- 52
Perov- 5
MR (%) ↑
RMSE ↓
MR (%) ↑
RMSE ↓
MR (%) ↑
RMSE ↓
CSPML
70.51
.0388
36.98
.0664
51.84
.1066
DiffCSP++
70.58
.0272
37.17
.0676
52.17
.0841
SGFM
70.13
.0310
35.09
.0739
54.10
.1508
Appendix
Table 8: CSPML evaluation. Wyckoff positions are obtained via the CSPML heuristic. The reported metrics are MR and RMSE.
Noise Prior \Model
Equivariant
Non-Equivariant
Wyckoff Conditional
82.74
68.16
Uniform
-
64.49
Appendix
Table 9: Noise prior vs. model equivariance comparison. The reported result is MR on the MP- 20 dataset.
Figure 5: Sampling Efficiency Analysis. MR ( ↑ ) and RMSE ( ↓ ) versus integration steps on MP- 20 . We observe that SGFM reaches near-optimal accuracy within just 50–100 steps. DiffCSP++ converges more slowly, requiring up to 1000 steps to approach its best performance—while still showing a notable gap in match rate compared to SGFM. In this same plot, we include a version of SGFM without anti-annealing to assess the effects of anti-annealing on performance.
NVIDIA · Center for Computational Science and Engineering, Massachusetts Institute of Technology · Department of Materials Science and Engineering, Massachusetts Institute of Technology +4
Institute for Functional Intelligent Materials (IFIM), National University of Singapore · School of Computing, National University of Singapore · Southern University of Science and Technology