Rotation-Invariant Machine Learning for Highly Symmetric Molecular Configurations
Authors: Logan Bolton, Tomas Karella, Nicholas Lubbers, Nathaniel Philip Gorski, Emily Shinkle, Alice Elisabeth Anastasia Allen, Pieter J Swart, Roxana Bujack
Organizations: CAI-3: Information Sciences, Los Alamos National Laboratory · Machine Learning in Molecular Dynamics, Max Planck Institute for Polymer Research · T-5: Applied Mathematics and Plasma Physics, Los Alamos National Laboratory
Atomistic machine learning can approximate expensive quantum mechanical calculations at a fraction of their computational cost, but accuracy depends strongly on how the three-dimensional atomic environments are represented. Invariant representations enable fast computation but can fail to distinguish geometrically distinct configurations, especially in highly symmetric environments. We introduce a lightweight benchmark to evaluate expressivity on n-fold symmetric configurations, and use it to expose degeneracies in both invariant and equivariant architectures. Guided by these results, we construct a new invariant basis for HIP-HOP which resolves most of the identified failures. On a methane simulation dataset, the extended basis achieves the lowest error among the HIP-HOP variants and has similar accuracy to EquiformerV3 while running 3.2× faster at inference. These results show that carefully constructed invariant representations can mitigate key expressivity limitations while balancing accuracy and computational efficiency.
Figures & tables
Figure 1: For the rings benchmark, the model is given two rotated versions of the same graph, and must determine which orientation has the nodes closer together.
Figure 2: The n -body benchmark (A) accuracy is saturated for all bases, but the rings benchmark (B) shows the difference in expressivity between bases.
Model / Basis
Prediction Performance
Inference Speed
Model
ℓmax
nmax
# Par.
E MAE
E RMSE
F MAE
F RMSE
μ s/atom
Speedup
HIP-HOP
3
4
25 K
0.21±0.02
0.50±0.35
0.61±0.03
2.63±2.56
2.28
1.07 ×
HIP-HOP
4
3
24 K
0.21±0.00
0.38±0.02
0.62±0.01
1.54±0.33
3.38
1.05 ×
HIP-HOP
4
4
36 K
0.18±0.00
0.32±0.01
0.55±0.01
1.24±0.03
4.42
1.57 ×
EquiformerV3
4
N/A
36 K
0.18±N/A
0.29±N/A
0.51±N/A
1.25±N/A
14.3
N/A
HIP-HOP
4
4
1.13 M
0.04±N/A
0.11±N/A
0.17±N/A
0.61±N/A
12.7
1.72 ×
Table 3
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
Hyperparameter
36.2 K model
1.12 M model
Message-passing layers
1
1
Hidden dimension
16
32
ℓmax
4
4
mmax
ℓmax
ℓmax
Attention heads
2
8
Attention hidden dimension
4
32
Appendix
Table 1: EquiformerV3 hyperparameters used for methane training.
Hyperparameter
∼ 25–36 K models
1.13 M model
Message-passing layers
1
1
Hidden dimension
32
192
Cutoff radius
10.3 Å
10.3 Å
Minimum batch size
256
4096
Maximum batch size
4096
4096
Batch size doubling patience
150 epochs
–
Appendix
Table 2: HIP-HOP-NN hyperparameters used for methane training.
Basis
Energy
Forces
ℓmax
nmax
# Par.
MAE
RMSE
MAE
RMSE
HIP-HOP-NN
3
4
24.7 K
0.213±0.016
0.500±0.347
0.607±0.032
2.631±2.557
3
5
25.8 K
0.225±0.031
0.389±0.046
0.642±0.046
3.842±4.498
4
3
24.4 K
0.213±0.004
0.378±0.024
0.623±0.013
1.537±0.333
4
4
36.3 K
0.184±0.003
0.322±0.006
0.551±0.011
1.244±0.026
Appendix
Table 6
Hyperparameter
Value
Message-passing layers
1
Hidden dimension
32
Sensitivity functions
32
Atom layers per interaction
2
Loss
binary cross-entropy
Optimizer
Adam
Appendix
Table 3: HIP-HOP-NN hyperparameters used for the n-body benchmark.
Hyperparameter
Value
Graphs per configuration
2
Message-passing layers
1
Hidden dimension
32
Sensitivity functions
32
Atom layers per interaction
2
Loss
binary cross-entropy
Appendix
Table 4: HIP-HOP-NN hyperparameters used for the ring benchmark.
Figure 3: Ring benchmark configurations with (a) equal radii for the inner and outer rings and (b) different radii for the inner and outer rings.
Figure 4: Ring benchmark performance across graph configurations for different model and invariant-basis configurations.
Model / Basis
Inference Speed ( μ s/atom)
Model
ℓmax
nmax
# Par.
Baseline
Optimized
Speedup
HIP-HOP
3
4
24.7 K
2.45
2.28
1.07 ×
1.02 M
10.19
9.66
1.05 ×
HIP-HOP
3
5
25.8 K
3.13
2.95
1.06 ×
1.05 M
16.04
15.76
1.02 ×
EquiformerV3
3
N/A
23.1 K
11.3
Appendix
Table 5: Inference speed comparison between HIP-HOP and EquiformerV3 models at different model sizes.
Model
Params
Time / Epoch (s)
EquiformerV3 ( ℓmax=4 )
1.12M
388.8
HIP-HOP ( ℓmax=4 , nmax=4 )
1.13M
234.8
Appendix
Table 12
Basis
Invariant Construction (seconds)
ℓmax
nmax
Baseline
Optimized
Speedup
3
4
0.899
0.079
11.40 ×
3
5
1.791
0.084
21.35 ×
4
3
0.500
0.075
6.69 ×
4
4
4.028
0.138
29.29 ×
4
7
267.811
0.862
310.70 ×
Appendix
Table 13
Figure 5: Invariant tensor-contraction basis for HIP-HOP with angular channels up to ( ℓmax=4 )
Figure 6: Nonzero contractions for the 2/3 -ring configuration as a function of the relative rotation between the two- and three-fold components. The original contractions (left) remain constant, while the additional contraction Bujack et al. (2025) (right) varies with the relative rotation.
State Key Lab of Processors, Institute of Computing Technology, Chinese Academy of Sciences · University of Chinese Academy of Sciences · School of Advanced Interdisciplinary Sciences, UCAS