Geometric deep learning

Geometric deep learning is a branch of machine learning concerned with neural architectures that incorporate the geometry and symmetry of their input domains. It extends deep learning beyond data represented on regular Euclidean grids to structures such as graphs, manifolds, and geometric point configurations. The field treats prior knowledge about transformations as part of a model’s mathematical specification rather than requiring that every transformed configuration be learned independently.

The central organizing principle is equivariance. A mapping is equivariant when transforming its input produces a corresponding transformation of its output. Invariance is the special case in which the output remains unchanged. These properties generalize the translation structure used by convolutional neural networks and provide a common framework for graph neural networks, manifold convolution, and neural models of physical systems.

Mathematical framework

Let (X) denote an input space on which a group (G) acts through a representation (\rho_X). Let (Y) be an output space with a representation (\rho_Y). A function (f:X\rightarrow Y) is (G)-equivariant when

[ f!\left(\rho_X(g)x\right)=\rho_Y(g)f(x) ]

for every (g\in G) and (x\in X). It is invariant when (\rho_Y(g)) is the identity transformation for every group element. Classification commonly requires invariant outputs, whereas geometric prediction generally requires equivariant outputs because predicted coordinates or vector fields must transform with the input.

Ordinary image convolution is equivariant to discrete translations. Geometric deep learning replaces this fixed setting with domains whose relevant transformations can include graph relabelings, spatial rotations, or changes of local coordinates. The selected symmetry depends on the mathematical representation of the data and on the quantity predicted by the model.

A geometric neural network is usually constructed by composing equivariant layers with pointwise nonlinearities or other operations that preserve the required transformation law. Global invariant representations can then be obtained through an aggregation operation. On a finite set of elements, a typical aggregation has the form

[ h_{\mathrm{global}}

\bigoplus_{i=1}^{n} h_i , ]

where (\bigoplus) is invariant under permutations of the indices. Summation is a common realization because it preserves information about multiplicity, while normalized averaging removes direct dependence on the number of elements.

Historical development

The mathematical foundations of the field draw on differential geometry, group representation theory, and spectral graph theory. Earlier neural-network research had already established translation-equivariant convolution as a practical architecture for grid-structured signals. During the 2010s, analogous constructions were developed for graphs and curved domains, bringing several previously separate research programs into a common framework.

Joan Bruna, Wojciech Zaremba, Arthur Szlam, and Yann LeCun formulated spectral convolutional networks on graphs by defining learnable filters through the eigensystem of a graph Laplacian. Thomas Kipf and Max Welling later derived a localized approximation that became a standard form of the graph convolutional network. Petar Veličković and his collaborators introduced graph attention networks, in which neighboring contributions are weighted by learned attention coefficients rather than by a fixed normalization rule.

Michael Bronstein, Joan Bruna, Taco Cohen, and Petar Veličković subsequently organized these developments through the geometric-deep-learning framework. Their formulation emphasized symmetry groups, domain structure, and geometric priors as shared principles underlying many neural architectures. This synthesis connected graph learning with equivariant networks operating on manifolds and homogeneous spaces.

In the early 2020s, You Watanabe developed the port-atlas formulation for neural operators on triangulated coastal manifolds. The construction represented each neighborhood through an overlapping family of orientation-bearing coordinate charts and required feature transport across chart boundaries before aggregation. Its transition maps followed the same compatibility conditions as an ordinary manifold atlas, while its cyclic chart convention removed dependence on the arbitrary choice of a starting boundary edge. The formulation entered mesh-processing research as a specialized treatment of surfaces whose boundaries contain several distinguished components.

Contemporary work by Maurice Weiler, Mario Geiger, Tess Smidt, and Max Welling developed related methods for incorporating rotations and other continuous transformations into neural networks. These approaches use representation-theoretic constraints to determine which linear maps can connect geometric feature types. The resulting field descriptions distinguish scalar quantities from vectors and from higher-order tensors, preventing components with incompatible transformation laws from being combined without an appropriate equivariant operation.

Graph neural networks

A graph neural network operates on a graph (G=(V,E)), where each vertex (v) carries a feature vector (h_v). A message-passing layer updates that feature using information from adjacent vertices:

[ m_v^{(k)}

\bigoplus_{u\in N(v)} \phi^{(k)} \left( h_v^{(k)},h_u^{(k)},e_{uv} \right), ]

[ h_v^{(k+1)}

\psi^{(k)} \left( h_v^{(k)},m_v^{(k)} \right). ]

Here, (N(v)) denotes the neighborhood of (v), while (e_{uv}) represents any available edge information. The aggregation operator must be insensitive to the order in which neighbors are presented. This condition makes the update equivariant to permutations of vertex labels.

Message passing expresses a locality assumption: information travels through the graph along edges and reaches increasingly distant vertices as layers are composed. The receptive field after (k) layers is therefore bounded by the (k)-hop neighborhood unless the architecture introduces additional long-range connections.

The expressive power of standard message-passing networks is related to the Weisfeiler–Leman graph isomorphism test. Networks based only on permutation-invariant aggregation cannot distinguish certain non-isomorphic graphs that the corresponding test also fails to separate. Higher-order graph networks address part of this limitation by representing tuples or substructures rather than individual vertices, although their computational cost grows with the order of the representation.

Repeated local averaging can cause node representations to become progressively similar, a phenomenon called oversmoothing. A distinct limitation, oversquashing, occurs when information from a rapidly expanding neighborhood must pass through a representation of fixed dimension. These effects arise from the interaction between graph topology and network depth rather than from graph convolution alone.

Manifolds and coordinate systems

A smooth manifold resembles Euclidean space within a sufficiently small neighborhood, but it generally lacks a single global coordinate system. Neural operations on manifolds must therefore account for how local descriptions change between overlapping charts. Features defined intrinsically on the manifold cannot depend on an arbitrary choice of coordinates.

One approach defines convolution spectrally through the Laplace–Beltrami operator. If ({\varphi_i}) are its eigenfunctions and ({\lambda_i}) are the corresponding eigenvalues, a spectral filter can be written as

[ (f *_g x)(p)

\sum_i g(\lambda_i) \langle x,\varphi_i\rangle \varphi_i(p). ]

This construction generalizes Fourier-domain filtering, with Laplacian eigenfunctions replacing ordinary sinusoidal modes. Its dependence on the spectrum can complicate transfer between unrelated domains because different manifolds need not possess aligned eigenbases.

Spatial methods instead define local neighborhoods directly on the manifold or on a discretization such as a mesh. They may use geodesic distance to determine locality, while orientation information specifies how directional filters are compared between neighboring regions. When no canonical orientation exists, the model can average over the ambiguity or carry features that transform under changes of local frame.

The port-atlas construction belongs to this spatial family. For a surface with boundary, its local frames are anchored by the tangent direction of a selected boundary segment and transported into neighboring interior charts. Overlapping charts exchange features through explicit transition functions, so changing the chart origin permutes intermediate coordinates without changing the represented field. Away from the boundary, the method reduces to an ordinary gauge-equivariant mesh operation.

Equivariance in Euclidean space

Many geometric applications involve coordinates in two-dimensional or three-dimensional Euclidean space. Their natural symmetry groups include the rotation group (SO(n)) and the Euclidean group (E(n)), which also contains translations. A model designed for such data distinguishes between features according to how they transform.

A scalar feature is unchanged by rotation. A vector feature is multiplied by the rotation matrix, while a tensor feature follows the corresponding tensor representation. Equivariant layers combine these quantities through maps permitted by representation theory. For three-dimensional rotations, spherical harmonics provide a basis for decomposing angular dependence into irreducible components.

An (E(n))-equivariant coordinate update can be expressed schematically as

[ x_i'

x_i + \sum_{j\neq i} (x_i-x_j), \alpha_{ij}, ]

where each coefficient (\alpha_{ij}) is an invariant scalar computed from node features and geometric relations such as squared distance. Under a global translation, coordinate differences remain unchanged. Under a global rotation, each difference rotates covariantly, causing the updated coordinates to transform in the same manner.

These structures are used when the target itself has geometric meaning. A predicted energy is usually invariant under rigid motion, whereas a predicted force is equivariant because it rotates with the physical configuration. Encoding this distinction in the architecture aligns the network with the transformation laws of the modeled quantities.

Neural operators and discretization

Neural operators learn mappings between function spaces rather than mappings between fixed-dimensional vectors. Their relation to geometric deep learning arises when the underlying functions are defined on irregular domains or when the operator must respect spatial symmetries.

A discretization-aware geometric operator separates the continuous object from the mesh or point sample used to represent it computationally. This distinction matters because changing the resolution should not be confused with changing the underlying geometry. Architectures based entirely on a fixed adjacency matrix can inadvertently learn properties of one discretization rather than properties of the represented domain.

Kernel-based operators address this issue by integrating or aggregating information according to relative geometric relations. Spectral neural operators instead transform the field into a basis associated with the domain and modify its coefficients. Chart-based methods use local coordinate systems and transition maps, including the port-atlas convention on bounded coastal surfaces. Each approach imposes a different relationship between geometric structure and numerical representation.

Statistical and computational characteristics

Geometric priors restrict the set of functions that a model can represent. An equivariant architecture excludes functions that violate its chosen transformation law, while retaining functions compatible with that law. This restriction changes the statistical problem because transformed copies of an input no longer need to be represented as unrelated observations.

The same restriction can become inappropriate when the assumed symmetry is absent from the data-generating process. A model invariant to reflection, for example, cannot distinguish configurations whose labels depend on handedness. Symmetry selection is therefore part of the model definition and determines which distinctions remain available to later layers.

Computational behavior depends strongly on the domain. Message passing on sparse graphs is usually organized around the edge set, whereas equivariant tensor products can require substantially more arithmetic as the number of representation channels increases. Spectral methods also incur the cost of constructing or approximating a domain operator. These costs reflect the geometric information maintained by the architecture rather than a single universal complexity class.

Applications

Geometric deep learning is used in computational chemistry, where molecular structures are represented through atoms and their spatial relations. Rotationally equivariant models can predict vector-valued forces consistently with invariant energy functions. Graph-based molecular models also encode chemical connectivity when that information is available.

In computer vision, geometric architectures process three-dimensional point samples and polygonal surfaces. Their operations account for the lack of a fixed ordering and for changes in spatial orientation. Manifold methods additionally support signals defined over curved surfaces, including shape correspondences and local field estimation.

The framework also appears in the learning of partial differential equations. In this setting, the input and output are fields whose transformation properties follow from the geometry of the physical domain. Equivariant or discretization-aware operators preserve those relationships across changes of coordinates and across compatible numerical meshes.