Many fields of physical interest on the sphere, such as wind, the polarisation of the cosmic microwave background and the shear of weak gravitational lensing, are spin fields, whose values change phase when the local frame rotates. We extend spherical Fourier neural operators (SFNOs) to fields of any spin, through lifted spherical convolutions that are exactly rotationally equivariant. Lifting spin fields to Mackey functions on the rotation group provides the conceptual framework, while computation is performed exactly and efficiently through their harmonic representation and by appealing to sampling theorems on the sphere and rotation group. We construct spin SFNOs (S2FNOs) from these building blocks. We demonstrate empirically that rotational equivariance holds to machine precision for these layers. We achieve state-of-the-art performance on existing benchmarks, and introduce new benchmarks on which spin-aware models outperform their scalar counterparts.
Figures & tables
Figure 1: The spin spherical convolution (a), one multiplier per degree between spin spherical harmonic transforms (SHTs) without forming the lift, and the convolution on the rotation group (b), each followed by ζ and computed in harmonic space below. Spheres denote fields on S2 , spheres with a cone denote functions lifted to SO(3) , triangles denote spin spherical harmonic coefficients and prisms Wigner coefficients, and i indexes output spins si in (a) and kernels in (b).
Figure 2: The S 2 FNO block ( 10 ), with C the spin spherical convolution or the convolution on the rotation group, W1 and W2 harmonic 1×1 convolutions and ζ the activation (in detail in Fig. 10 ).
Equivariance error
Operation
Spin
S2
SO(3)
Conv- S2
0
5.9×10−15
–
Conv- S2
1
5.2×10−15
–
Conv- SO(3)
2
–
6.0×10−15
Harmonic 1×1 conv.
5.4×10−15
5.4×10−15
S 2 FNO block
{0,±1}
2.1×10−14
2.4×10−14
Table 1: Equivariance errors in double precision on the sphere ( S2 ) and the rotation group ( SO(3) ), measured by the mean relative equivariance error ( MREE ) between rotating the input and rotating the output (Appendix E ). Conv- S2 and Conv- SO(3) are the spin spherical convolution Cs→s′ and the directional convolution Cs→SO(3) , ReLU and Swish are applied directly to complex data, whereas the complex activations act on the modulus with the phase kept ( 9 ), and the complex ReLU is measured at initialisation, where it is the identity up to ϵ . The S 2 FNO blocks and the UNets of the spin spherical MNIST architecture have no activation or normalisation, are at L=32 , with N=8 on SO(3) , and reach the round-off floor of the transforms. All other rows are at L=16 , with N=5 on SO(3) .
Sobel
Harmonic
Model
Params
spin 1
spin 1
spin 2
Conv- S2
50k
99.15
99.19
99.17
S 2 FNO- S2
182k
98.82
99.16
99.34
Conv- SO(3)
134k
99.01
99.10
99.11
S 2 FNO- SO(3)
136k
99.17
99.21
99.28
Esteves et al.
58k
98.2
–
–
Table 2: Spin spherical MNIST classification: NR/R accuracy (%) of our spherical ( S2 ) and rotation-group ( SO(3) ) networks (best accuracy and fewest parameters in bold).
MSE ×103↓
Sobel field
Harmonic field
Model
Params
0→1
1→0
0→1
1→0
Conv- S2
23k
3.681
3.212
5.580
3.229
S 2 FNO- S2
29k
4.253
4.544
3.944
4.468
Conv- SO(3)
14k
3.016
3.144
2.876
3.101
S 2 FNO- SO(3)
20k
3.866
2.948
3.063
3.005
Table 3: Spin-changing prediction on spin spherical MNIST: NR/R MSE from the spin- 0 image to its spin- 1 field ( 0→1 ) and back ( 1→0 ), with the best model and the fewest parameters in bold.
Figure 3: Prediction of the spin- 1 Sobel field from the spin- 0 image of a rotated test digit, chosen in advance, by the Conv- SO(3) and Conv- S2 UNets, with their relative errors. Both predictions follow the rotated frame of the target (more in Fig. 13 ).
Model
Params
MSE ↓
SNR ↑
PSNR ↑
PCC ↑
(dB)
(dB)
Global wind forecasting (MSE in m 2 s -2 )
Spin UNet
48k
9.52
14.02
25.06
0.979
Scalar UNet
74k
17.27
11.43
22.64
0.969
Persistence
–
14.55
12.17
23.19
0.962
Weak gravitational lensing (MSE ×106 )
Table 4: Wind forecasting at 50 hPa on the held-out days and mass mapping from the noisy spin- 2 shear of weak lensing at L=256 , averaged over 600 maps (best results and fewest parameters in bold; all metrics in Tables 8 and 9 ). All models are spherical, and the scalar UNet takes the spin field as two scalar channels. Persistence forecasts the 12:00 UTC field for 18:00 UTC.
Figure 4: Wind forecast for one held-out day: (a) the target speed, (b–d) the vector errors of the spin and scalar UNets and persistence, and (e–h) the same in the outlined region, with true (white) and predicted (coloured) vectors. The scalar UNet leaves larger, coherent patches of error (more in Fig. 14 ).
Appendix figures & tables15 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 6: What spin means at a point. Each panel is the tangent plane at x , with eθ to the right. A spin- s value is drawn as s equal spokes, and rotating the frame by χ multiplies it by e−isχ , as in ( 1 ). The ring, the fibre of SO(3) over x (Sec. 3.1 ), is coloured by the phase of the lift of the field, which winds s times around it.
Figure 7: Spin fields on the sphere: a scalar field f and the fields ðf , ð2f and ð3f obtained from it by the spin raising operator ð of ( 11 ), drawn with glyphs whose symmetry is their spin (Fig. 6 ). Wind has spin 1 , polarisation and shear spin 2 , and the second flexion of weak lensing spin 3 ( Bacon et al., 2006 ) .
Figure 8: The rotation group as a principal SO(2) -bundle over the sphere. The fibre over x , the rotations R(α,β,γ) with R(α,β,γ)η=x , is drawn as the rim of a cone at x , coloured by the fibre coordinate γ , with the section s(x)=R(α,β,0) as a red dot and a general element g=s(x)Rz(γ) in the inset.
Method
Year
Signals
Linear layer
Cost
Spherical CNN ( Esteves et al., 2018 )
2018
S2 , scalar
harmonic zonal filter, one weight per ℓ
L3
Spherical CNN ( Cohen et al., 2018 )
2018
SO(3) , scalar
grid filter via generalised FFT
L4
Clebsch–Gordan net ( Kondor et al., 2018 )
2018
SO(3) , harmonic
per- ℓ maps, CG products
L5
Spin-weighted CNN ( Esteves et al., 2020 )
2020
S2 , spins
harmonic filters interpolated over ℓ
L3
Generalised spherical CNN ( Cobb et al., 2021 )
2020
S2 , SO(3)
per- ℓ maps, azimuthal bandlimit N
NL3
FNO ( Li et al., 2021 )
2020
plane
one multiplier per mode
nlogn
Appendix
Table 5: Spherical convolutional layers as neural operators. A layer is a neural operator in the sense of Kovachki et al. (2023) when its parameters do not depend on the sampling of the sphere. Representative costs at bandlimit L (azimuthal bandlimit N≤L , n planar grid points) assume fixed channel and retained spin counts and, for DISCO, bounded local neighbourhood size.
Figure 9: An S 2 FNO block as implemented. The convolution C , here the spin spherical convolution ( 33 ), and a harmonic 1×1 convolution ( W1 ) with one channel matrix and one spin matrix shared across degrees act in parallel, and their sum passes through a second such convolution ( W2 ), the modulus non-linearity ζ and group normalisation (GN). Only the rotation-group construction forms the lift-and-sum Λ .
Figure 10: The generic S 2 FNO block ( 10 ) in detail. Stacked spheres with a dashed cone denote a tuple of spin fields or its lift to SO(3) , and the normalisation is not shown.
Figure 11: Composition of blocks into the UNets used for dense prediction. Every box is a block of a lifting convolution (LConv), the lift-and-sum (LSum), the activation (CAct) and group normalisation (GNorm). The first block maps the data spin onto the internal spin set S and the last layer returns the output spin. Brackets give the five-level spin UNet of the mass-mapping benchmark (bandlimit; channels), whereas the MNIST UNets are shallower and the S 2 FNO UNets keep L=32 at every level.
Figure 12: The two mass-mapping UNets, with five levels from L=256 to L=16 . Each box gives the bandlimit and the number of channels of a block, and the dashed lines carry the encoder features, which are concatenated ( ∥ ) with the decoder output at equal bandlimit. (a) The spin UNet of the rotation-group construction, whose first block lifts the spin- 2 shear to the internal spin set S and whose last layer, a projecting convolution CG , returns the spin- 0 convergence. (b) The scalar UNet, of DISCO convolutions and residual layers, which takes the real and imaginary parts of the shear as two channels and returns the convergence as its single channel.
Task
Architecture
Trainable parameters
Classification
Conv- S2
49,674
Classification
Conv- SO(3)
133,562
Classification
S 2 FNO- S2
182,016
Classification
S 2 FNO- SO(3)
136,088
Prediction
Conv- S2
23,330
Prediction
Conv- SO(3)
13,890
Appendix
Table 6: Trainable parameter counts of the architectures evaluated in Sec. 7 .
Model
N
Spins
Channels
Params
Equivariance error ( MREE )
Identity on S2
1
1
0
2.20×10−14
Identity on SO(3)
8
1
0
1.99×10−13
Positive control
1
1
0
8.03×10−2
S 2 FNO- S2 block
0→0→0
1,1,1
15
2.27×10−14
S 2 FNO- S2 block
1→1→1
1,1,1
15
2.28×10−14
S 2 FNO- S2 block ‡
0→{−1,0,1}→{−1,0,1}
1,8,8
138
2.00×10−14
Appendix
Table 7: Equivariance errors of linear models at L=32 , measured by the MREE ( 34 ) over the 104 pairs of signal and rotation. Spins and channels are given as input, hidden and output, or as one set of spins used throughout, and c is the base width of a UNet, with encoder widths c , 2c , 4c and 4c . ‡ The first block of the S 2 FNO- S2 UNet.
Figure 13: Prediction by the Conv- S2 and Conv- SO(3) UNets at L=32 from the spin- 0 image to the spin- 1 Sobel field (a–c) and back (d–f), for the first three test digits under one rotation. Errors (prediction minus target) are on the scales of the target, spin- 1 fields are drawn as arrows over their modulus, and percentages give the relative error MSE/∣f∣2 .
Model
Params
LR
Input shape
MSE ↓
RMSE ↓
MAE ↓
SNR ↑
PSNR ↑
PCC ↑
Spin UNet
48,446
9×10−4
20×129×256×3
9.52
3.09
2.67
14.02
25.06
0.979
Scalar UNet
74,398
8×10−4
50×129×256×6
17.27
4.16
3.56
11.43
22.64
0.969
Persistence
–
–
–
14.55
3.82
3.31
12.17
23.19
0.962
Appendix
Table 8: Wind forecasting at 50 hPa and L=128 : configurations and all metrics on the held-out days, in physical units (MSE in m 2 s -2 , RMSE and MAE in m s -1 , SNR and PSNR in dB). Input shapes are batch ×θ×φ× channels, and best results are in bold.
Figure 14: Wind forecast for one held-out day at L=128 : the speed ∣u+iv∣ of the inputs, the target and the forecasts (a–f), the vector error ∣wtrue−wpred∣ of the spin UNet, the scalar UNet and persistence with the area-weighted MSE of this day (h–j), and a zoom into the region outlined in (g), with true (white) and predicted (coloured) vectors and the RMSE over the region (k–n).
MSE ↓
RMSE ↓
MAE ↓
SNR ↑
PSNR ↑
PCC ↑
Model
Params
LR
Input shape
×106
×103
×103
(dB)
(dB)
Spin UNet
80,702
5×10−4
1×257×512×1
1.28
1.13
0.90
7.53
23.47
0.907
Scalar UNet
93,198
6×10−4
20×257×512×2
2.06
1.43
1.14
5.45
21.40
0.846
Appendix
Table 9: Weak lensing mass mapping at L=256 with shape noise: configurations and all metrics, averaged over the 600 evaluation maps. Input shapes are batch ×θ×φ× channels, and best results are in bold.
Figure 15: Mass mapping for one evaluation map at L=256 , with the true convergence (a), its reconstructions by the spin and scalar UNets (b, c) and their one-point distributions with that of the target (d), the noisy input shear (e, f) and the residuals κ−κ^ of the two UNets with their MSE on this map (g, h). The scalar residual retains large-scale structure that the spin residual does not.
Fourier Neural Operators learn mappings between input and solution fields using spectral convolutions, enabling fast surrogates for parameterized PDEs.
Department of Physics & Astronomy, University College London, Gower Street, London, WC1E 6BT, UK · Department of Physics & King’s Institute for Artificial Intelligence, King’s College London, Strand, London WC2R 2LS, United Kingdom