21 numbers describe the elastic response of a fully general anisotropic solid. For the olivine-rich upper mantle beneath your survey, the honest count is 9, or 5, or 2. 19 of those 21 are not free parameters at all. They are consequences, fixed exactly by symmetry. Whether a model knows that before training starts, or has to infer it from noisy examples, decides both how much data you need and whether the tensor it hands back describes a rock that could exist.
Where 21 comes from
Stress and strain are linked by the fourth-rank stiffness tensor C, which carries 81 components in the abstract. Two facts collapse that quickly. The symmetry of the stress and strain tensors reduces it to a 6x6 matrix in Voigt notation, 36 numbers. The existence of a strain-energy function then forces that matrix to be symmetric, so the upper and lower triangles agree [1].
21 independent constants remain. That is the triclinic case: no crystallographic symmetry at all, every entry free to move. It is the most general elastic solid, and it is the worst case for anyone trying to learn the tensor, because there are 21 knobs and nothing in the problem statement says how they relate. A wrong knob is a wrong predicted splitting time, or a spurious lobe on a transverse receiver function.
Symmetry is an exact prior, and it is free
Real mantle minerals are not triclinic blobs. Every symmetry operation a crystal obeys is a hard constraint on the stiffness matrix. Single-crystal olivine is orthorhombic, and imposing that symmetry drops the count to 9: the shear-normal couplings vanish and only three diagonal-block groups survive [2].
The seismically relevant symmetry is usually coarser than the crystal's, because it is set by the fabric rather than the grain. Mantle flow shears the aggregate and drives lattice-preferred orientation, rotating the olivine a-axes toward the flow direction, and the bulk tensor of the aligned aggregate inherits an effective symmetry [2,3]. Align the a-axes about a single axis and the aggregate becomes transversely isotropic, or hexagonal, at 5 independent constants. Take a texture-free, randomly oriented aggregate and you arrive at full isotropy at 2, the Lamé parameters lambda and mu, from which every other entry follows.
Each step either forces entries to zero or ties them to one another. In the hexagonal case, for instance, one entry is not free but bound:
Triclinic to isotropic saves 19 of the 21 constants. That saving is a prior of the exact kind, handed over by group theory at no cost in data, and it is the same symmetry that decides whether the medium splits a shear wave and by how much.
What the naive network is actually asked to do
Now put a network on the job: map an olivine fabric, or a set of seismic observations, to the effective stiffness tensor of the aggregate. A network with no symmetry structure treats the 21 constants as 21 independent regression targets. It has to discover from examples that C22 equals C11 in a hexagonal medium, that the shear-normal couplings are zero under orthorhombic symmetry, that C12 equals C11 minus twice C66.
Every one of those is a law, not a trend. Asking a finite, noisy sample to reproduce a law returns the law approximately, and approximately is another word for wrong. The predicted tensor then violates the symmetry class it belongs to. Its eigen-velocities and its fast axis are subtly off, and every quantity computed downstream of it, splitting delay time and transverse receiver-function amplitude among them, inherits that error. Rounding the output afterwards does not repair it, because the damage sits in the relations between entries rather than in the digits of any one entry.
Equivariance learns the free constants and none of the constraints
The alternative is architectural. An equivariant network is built so that rotating or reflecting the input rotates or reflects the output in exactly the same way, by construction rather than by training [4,5]. For anisotropy that property is the physics, not a convenience: rotate the fabric and the fast axis, the splitting direction, and the whole tensor must rotate with it under the rank-4 transformation law.
A network equivariant to SO(3) cannot emit a tensor that breaks the symmetry, because the inadmissible region of parameter space is not representable at all. A fabric and a rotated copy of that fabric give consistently rotated splitting predictions rather than two independently learned guesses. And with the relations built in, the network fits only what the physics leaves free: 5 constants for a transversely isotropic mantle fabric, 2 for isotropy. Fewer free parameters means less data to constrain them, faster convergence, and, for a model meant to be scientific rather than decorative, a stiffness tensor that is admissible by construction. The irreducible decomposition of the elasticity tensor [6] is what makes that bookkeeping precise: the free content lives in a small number of pieces, and an equivariant read-out predicts those pieces rather than 21 loose numbers.
That reasoning is the design brief behind the E(3)-equivariant tensor-field surrogate we described in an April method note, which maps an olivine point cloud straight to C and holds its equivariance error at zero for every rotation. The parameter count in this piece is the reason that architecture earns its higher cost per layer.
Why an asset team should care about a constant count
None of this stays theoretical for long. The observables that anisotropy work sells are shear-wave splitting delay time, which reads as a fracture-density gauge, and the backazimuth harmonics of the transverse receiver function, which name the fabric direction. Both are computed from C. A tensor whose anisotropic part is slightly inconsistent gives a fast azimuth that points somewhere other than where the fabric points, and a fast azimuth is what a fracture-strike call rests on.
This is the seam where Tannistha Maiti's doctoral work on receiver-function imaging and upper-mantle anisotropy [7] meets the machine-learning half of the same programme. The thesis measured the observables. The surrogate predicts the tensor that generates them. A parameter count is what tells you how much of that prediction the physics will do for you at no charge, and how much you would otherwise be paying for in training fabrics.
Zooming out: architecture as a way to spend data on the right thing
The elastic-symmetry case is the clean instance of a principle this whole series runs on. The cheapest way to reduce the data a model needs is to give it the physics it would otherwise have to learn. Elasticity makes the point better than most examples because the reduction is exact and countable, 21 to 2, with nothing approximated anywhere in the argument.
The same logic runs through the rest of the stack: a forward operator that knows the wave equation, a deconvolution that knows the source statistics, an inversion that carries a crustal prior. Every symmetry, conservation law and invariance you build into an architecture is a parameter you never have to buy with training examples. The 21-number problem is small enough to hold in one hand, which is exactly why it is the right place to learn the lesson.
Limitations
- The constant counts are exact for ideal symmetry classes. Real aggregates are polyphase and defective, a fabric that is nearly hexagonal is not exactly hexagonal, and imposing an exact class can smooth over a genuine departure from it.
- The exhibit's matrix colouring is a teaching schematic of which entries are free, linked, equal or forced to zero. It is not the numerical tensor of any specific olivine aggregate.
- Symmetry constrains the form of the map, not its accuracy. A well-formed equivariant prediction can still be a wrong prediction.
- The saving quoted here is a parameter count, not a measured data-efficiency result for any particular study. Sample-efficiency claims belong to training experiments, not to group theory.
- Equivariant layers cost more per forward pass than plain matrix multiplies, so the parameter saving buys data efficiency rather than raw speed.
By the numbers
Components in C before symmetry
Independent constants, triclinic
Orthorhombic olivine
LPO-aligned TI fabric
Isotropic aggregate
Key takeaways
- A general anisotropic stiffness tensor has 21 independent constants: 81 components reduced by strain-tensor symmetry and by the existence of a strain-energy function.
- Crystal and fabric symmetry is an exact prior, not an approximation. Orthorhombic olivine needs 9 constants, an LPO-aligned transversely isotropic fabric needs 5, isotropy needs 2, and the trip from triclinic to isotropic saves 19.
- A network without symmetry structure must learn all 21 targets and the laws relating them from finite noisy data, and predicts symmetry-violating tensors when it gets those laws slightly wrong.
- An equivariant architecture makes the rank-4 rotation law structural, so the network fits only the free constants and cannot represent an inadmissible tensor.
- A wrong tensor is not an abstract failure: it moves the fast azimuth, and fast azimuth is what fracture-strike and fabric-direction calls depend on.
References
[1] J. F. Nye. Physical Properties of Crystals: Their Representation by Tensors and Matrices. Oxford University Press, 1957. (Symmetry-class reductions of the elastic constant count.)
[2] D. Mainprice, A. Ismail. An olivine fabric database. Tectonophysics, 1998. doi:10.1016/S0040-1951(98)00141-3
[3] P. G. Silver. Seismic anisotropy beneath the continents. Annu. Rev. Earth Planet. Sci., 1996. doi:10.1146/annurev.earth.24.1.385
[4] N. Thomas, T. Smidt, S. Kearnes, et al. Tensor Field Networks: Rotation- and Translation-Equivariant Neural Networks for 3D Point Clouds. 2018. arXiv:1802.08219
[5] M. Geiger, T. Smidt. e3nn: Euclidean Neural Networks. 2022. arXiv:2207.09453
[6] J. Browaeys, S. Chevrot. Decomposition of the elastic tensor and geophysical applications. Geophys. J. Int., 2004. doi:10.1111/j.1365-246X.2004.02415.x
[7] Source: 2018 PhD thesis (T. Maiti), receiver-function imaging of the Moho and LAB, seismic anisotropy, and geodynamic modelling. The symmetry argument here is a machine-learning reading of that material, not a result reported in it.
