Geometric deep learning (GDL) is an area of deep learning that designs neural networks around the structure and symmetries of their data. It includes learning on graphs, sets, manifolds, meshes, spheres, point clouds, and particle systems, as well as familiar grid data viewed through translation symmetry. A 2017 survey described GDL as an umbrella term for extending structured neural models beyond ordinary Euclidean grids; a 2021 proto-book broadened that view into a common design program for grids, groups, graphs, geodesics, and gauges.[1][2]

GDL is not one algorithm or a settled mathematical theory encompassing every structured model. It brings together methods from group representation theory, differential geometry, graph signal processing, harmonic analysis, and machine learning. Different branches make different assumptions about the domain, transformations, feature types, and task. The shared idea is to encode suitable geometric inductive biases instead of asking an unrestricted neural network to relearn all structure from examples.[2][3]

Scope and development

Several research lines converged under the GDL name. Spectral networks proposed at ICLR 2014 defined filters through the eigenvectors of a graph Laplacian, extending convolution-like operations to graph-supported signals.[4] Cohen and Welling's 2016 group-equivariant convolutional neural networks generalized image convolution from translations to discrete rotations and reflections.[5] Bronstein and coauthors' 2017 survey organized work on graphs and manifolds under the term geometric deep learning, while MoNet supplied a spatial framework for local kernels on both kinds of domain.[1][6]

The scope then expanded beyond the original emphasis on non-Euclidean data. Kondor and Trivedi gave a group-theoretic account of convolution and equivariance for compact-group actions under stated assumptions.[7] Work on spherical, steerable, gauge-equivariant, and three-dimensional Euclidean networks made transformations of intermediate scalar, vector, and tensor features explicit.[3] The 2021 proto-book proposed a broad geometric blueprint that also interprets ordinary convolutional networks, graph neural networks, set models, and some attention or recurrent systems through their domains and symmetries.[2]

This later formulation should be read as a useful synthesis, not as evidence that all of those architectures are mathematically identical. A convolutional neural network on a fixed pixel grid, a message-passing network on a changing graph, and a gauge-equivariant network on a curved surface solve different problems and may require different definitions of locality, stability, and admissible transformations.[2][3]

Invariance and equivariance

Symmetry is the central organizing concept. Let a transformation g from a group G act on an input x. A function is invariant when f(g . x) = f(x): applying the transformation does not change the output. It is equivariant when f(g . x) = rho(g) . f(x), where rho(g) describes how the output should transform. Invariance is the special case in which rho(g) acts as the identity.[2][3]

The correct property depends on the target, not just the input. A molecule's scalar energy should be unchanged by a rigid translation or rotation, but its force vectors should rotate with the molecule. A graph-level label should not depend on arbitrary node names, while node-level predictions should be permuted in the same way as the input nodes. Image classification often seeks position-insensitive output, whereas image segmentation requires the output map to move with the input. Equivariance can preserve transformation information through intermediate layers before an invariant readout removes it when the final task requires that.[2][3]

Domain	Common transformation	Typical equivariant quantity	Typical invariant quantity
Image grid	Translation	Feature or segmentation map	Whole-image class
Unordered set	Element permutation	Per-element features	Set-level prediction
Graph	Node relabeling	Node or edge predictions	Graph-level prediction
Molecule or point cloud	Rotation and translation	Coordinates, vectors, or forces	Energy or object class
Sphere	Three-dimensional rotation	Spherical feature field	Global category
Manifold or mesh	Change of local frame	Tangent vector or tensor field	Intrinsic scalar output

An architectural guarantee is narrower than an empirical statement about robustness. A layer may be exactly equivariant in an ideal continuous or algebraic model yet only approximately so in code because of sampling, finite boundaries, interpolation, numerical precision, neighborhood construction, or pooling. Ordinary discrete convolution, for example, is translation equivariant under compatible shifts and boundary conventions, but striding and edge effects can break that property. The distinction between translational invariance and equivariance is therefore important.[5][7]

Geometric priors and the blueprint

The 2021 GDL blueprint groups three principles: symmetry, geometric stability, and scale separation. Symmetry describes exact transformations under which outputs stay fixed or transform predictably. Geometric stability asks that small deformations of a signal or its domain cause controlled changes, even when those deformations do not form a group. Scale separation uses local operations and coarsening to build representations across multiple spatial or relational scales.[2]

In that blueprint, a network combines local equivariant linear maps, pointwise nonlinearities, coarsening or local pooling, and an invariant global readout. Stacking local layers enlarges the receptive field. Coarsening provides a multiscale route to more distant interactions, and the final readout produces an invariant target when needed. These components describe a family of designs rather than prescribing one implementation.[2]

The framework also distinguishes fixed and varying domains. Images usually share one regular grid, while graphs and meshes can change from example to example. On a varying domain, robustness to changes in connectivity, sampling, or discretization becomes part of the learning problem. A model can obey an exact relabeling symmetry and still be unstable when a few graph edges or mesh vertices change.[2]

Main domains and methods

Grids, groups, and homogeneous spaces

Convolution on a regular grid shares a filter across translated locations, making translation equivariance a built-in property under suitable boundary conditions. Group-equivariant CNNs replace translations alone with a larger transformation group. Cohen and Welling demonstrated this construction for planar rotations and reflections, while later work treated spheres and other homogeneous spaces using group correlation, harmonic analysis, and constrained kernels.[5][7]

Intermediate features need not all be scalars. Vectors, orientations, and higher-order tensors transform according to representations of the chosen group. A steerable network constrains how filters and feature channels mix so that these transformation laws are preserved. This can reduce the need to show every transformed example during training, but it also makes the architecture specific to a proposed symmetry and feature representation.[3][7]

Graphs and sets

Graphs have no canonical node order. A graph neural network therefore normally shares update rules across nodes and aggregates neighbors with an order-independent operation. Nodewise outputs are permutation equivariant, and a sum, mean, or another symmetric readout can produce a permutation-invariant graph representation. Sets follow the same permutation principle without a predefined adjacency structure.[2][8]

Graph methods are commonly divided into spectral and spatial families. Spectral networks define filters using a graph Laplacian or polynomials of it. Spatial methods pass messages along edges and aggregate local neighborhoods. Gilmer and coauthors formalized many molecular models as message-passing neural networks, with repeated message and update phases followed by a graph-level readout.[4][8]

Graph deep learning is one branch of GDL, not a synonym for it. The live Graph Machine Learning Models topic focuses on graph tasks and architectures. GDL additionally includes continuous groups, curved domains, local frames, and physical transformation laws that are not captured by node relabeling alone.[1][2]

Manifolds, meshes, and gauges

A manifold is locally similar to Euclidean space but can be globally curved. Neural operations on a surface or mesh need a way to compare nearby features even though different points have different tangent spaces. MoNet used local pseudo-coordinates and Gaussian mixture kernels to define learnable patches on graphs and manifolds, including 3D shape correspondence experiments.[6]

Local coordinates or tangent frames are choices, not properties of the underlying object. A gauge transformation changes these local reference frames without changing the geometry. Gauge-equivariant layers require their outputs to transform consistently when the frames change. Cohen and coauthors implemented such a network on an icosahedral approximation of the sphere and evaluated it on omnidirectional-image and climate-pattern segmentation.[9] This use of gauge is a coordinate-consistency principle and should not be confused with learning a physical gauge theory in every application.

GDL overlaps with manifold learning but addresses a different primary problem. Manifold learning typically infers a low-dimensional representation from high-dimensional samples. Geometric deep learning may instead begin with a known graph, mesh, sphere, or manifold and design trainable operators that respect its structure.[1][3]

Point clouds and Euclidean geometry

Point clouds, molecular graphs, and particle systems combine permutation symmetry with transformations of coordinates. Models may target SE(3), the group of three-dimensional translations and proper rotations, or E(3), which also includes reflections. These are different assumptions: a task sensitive to chirality may not be reflection invariant.[10][11]

The SE(3)-Transformer combined equivariant feature maps with attention for 3D point clouds and graphs and evaluated the method on particle dynamics, object classification, and molecular regression.[10] E(n)-equivariant graph neural networks updated coordinates and scalar features while remaining equivariant to rotations, translations, reflections, and permutations; their authors designed the method to avoid computationally expensive higher-order intermediate representations.[11] Other model families use spherical harmonics, tensor products, and multiple irreducible feature types to retain richer angular information.[3]

These approaches express different accuracy, efficiency, and expressivity tradeoffs. An invariant model based only on pairwise distances can be simple and cheap, but it may discard directional or handed information. A tensor-valued equivariant model can preserve more of that information at the cost of more complex kernels and operations.[11][15]

Representative architectures and applications

Work	Geometric structure	Reported application examples
Spectral networks (2014)	Graph Laplacian spectrum	Graph-supported classification [4]
Group-equivariant CNNs (2016)	Discrete planar rotations and reflections	Rotated digits and image classification [5]
MoNet (2017)	Local coordinates on graphs and manifolds	Citation graphs and 3D shape correspondence [6]
Message-passing neural networks (2017)	Molecular graphs and permutation symmetry	Molecular property prediction [8]
Gauge-equivariant CNN (2019)	Local frames on a spherical discretization	Omnidirectional images and climate segmentation [9]
SE(3)-Transformer (2020)	3D roto-translation equivariance	Point clouds, particle dynamics, and molecules [10]
EGNN (2021)	E(n) coordinate and permutation equivariance	Dynamics, graph autoencoding, and molecular properties [11]
NequIP (2022)	E(3)-equivariant atomic graphs	Interatomic energies, forces, and molecular dynamics [12]

Chemistry and materials science are prominent applications because scalar energies, vector forces, atomic permutations, and rigid motions have clear transformation laws. NequIP used E(3)-equivariant convolutions for interatomic potentials and reported data-efficient accuracy across several molecular and materials systems.[12] Those results support that particular model and evaluation; they do not establish that every equivariant model will outperform every invariant baseline.

Structural biology supplies another kind of geometric domain. MaSIF represented protein molecular surfaces as local geodesic patches carrying geometric and chemical features, then learned descriptors for ligand, site, and interaction-related tasks.[13] The method illustrates that a surface representation can expose information different from a residue sequence or a simple contact graph. It does not imply that surface learning replaces sequence models or experimental validation.

Other published examples cover three-dimensional computer vision, spherical imagery, climate fields, particle systems, and some robotics problems.[3][9][10] In each case, the useful symmetry is task-dependent. A camera mounted upright may have different nuisance transformations from an unconstrained 3D object, and an external field can break a physical system's otherwise natural rotational symmetry.

Evaluation

There is no single benchmark that measures geometric deep learning as a whole. Graph classification, molecular force prediction, spherical segmentation, and point-cloud recognition use different targets and metrics. A credible evaluation should report ordinary predictive quality alongside the property the architecture was designed to enforce, such as equivariance error under sampled transformations, performance with transformed inputs, sample efficiency, runtime, memory, and sensitivity to discretization.[3][18]

Comparisons also need matched data, parameter budgets, optimization, and transformation coverage. Data augmentation can teach approximate robustness without an exact architectural guarantee, while an equivariant model can share parameters across a group by construction. Either approach may win under a particular compute or data budget. Results obtained for a finite rotation group do not automatically extend to continuous rotations, and results for scalars do not establish performance on vector or tensor targets.[3][5]

Scientific benchmarks increasingly test behavior outside the training distribution. The GeSS benchmark presented 30 settings across particle physics, materials science, and biochemistry, covering covariate, conditional, and concept shifts and multiple levels of access to test-distribution information.[18] Its motivation and results underscore that respecting a known symmetry is not, by itself, proof of reliable out-of-distribution generalization.

Limitations

Choosing the wrong symmetry. A built-in symmetry restricts the functions a network can represent. If the data only approximately obey it, or if the target changes under a transformation that the model treats as irrelevant, the constraint can reduce accuracy. Wang, Walters, and Yu showed this problem for imperfectly symmetric dynamics and proposed approximately equivariant alternatives.[17] Symmetry can also be broken by boundaries, sensors, external fields, labels, or chirality.

Computational cost. Group lifting can add spatial or orientation dimensions. Spherical harmonics, tensor products, and higher-order representations can increase memory, arithmetic, and implementation complexity. Graph and point-cloud models must also construct and process neighborhoods. An architecture that is mathematically elegant may be slower than a simpler invariant feature set or a well-augmented baseline.[3][10][11]

Expressivity. Common message-passing GNNs cannot distinguish every graph. Xu and coauthors related the discriminative power of broad GNN classes to the Weisfeiler-Lehman graph isomorphism test.[14] For geometric graphs, later analysis found meaningful differences between invariant and equivariant layers and between scalar and higher-order representations.[15] These are model-class results, not a single ceiling shared by all GDL methods.

Long-range information. Locality improves parameter sharing but can impede distant interactions. In message-passing networks, information from many remote nodes may be compressed through narrow graph bottlenecks, a phenomenon called over-squashing. Theory and experiments connect its severity to network width, depth, and graph topology.[16] Adding layers alone does not necessarily solve the problem.

Discretization and numerical error. Mesh quality, graph construction, sampling density, neighbor cutoffs, interpolation, and boundary rules can change both predictions and measured equivariance. Gauge or group consistency in a continuous derivation does not guarantee exact consistency in every discrete implementation. Evaluation should therefore measure the implemented system rather than infer behavior solely from its architecture label.[2][3][9]

Limited guarantees. Geometric priors do not ensure causal reasoning, calibrated uncertainty, interpretability, fairness, data quality, or safe extrapolation. They can narrow a hypothesis space and encode known transformation laws, but performance still depends on training data, objectives, optimization, and domain validation. GDL is best understood as a collection of principled design tools, not a universal replacement for empirical testing.[2][17][18]

Relationship to neighboring fields

Field or topic	Relationship to geometric deep learning
Graph deep learning	Major GDL branch focused on relational data and graph tasks
Equivariant deep learning	Major branch centered on explicit transformation laws; not every GDL method implements a group action
Manifold learning	Primarily infers low-dimensional structure or embeddings; GDL often designs trainable operators on a given domain
Topological deep learning	Studies higher-order structures such as simplicial, cellular, or combinatorial complexes; overlaps with but is not synonymous with GDL
Scientific machine learning	Broad application and methodology area that uses GDL among many other approaches

The boundaries are partly conventional. Some authors use geometric deep learning narrowly for graphs and manifolds, following the 2017 survey, while others use the broader 2021 blueprint centered on symmetry, stability, and scale. Clear descriptions should therefore name the actual domain, group, feature type, and guarantee rather than rely on the GDL label alone.[1][2][3]

References

Bronstein, Michael M., Joan Bruna, Yann LeCun, Arthur Szlam, and Pierre Vandergheynst. "Geometric Deep Learning: Going Beyond Euclidean Data." IEEE Signal Processing Magazine 34, no. 4 (2017): 18-42. https://doi.org/10.1109/MSP.2017.2693418
Bronstein, Michael M., Joan Bruna, Taco Cohen, and Petar Veličković. "Geometric Deep Learning: Grids, Groups, Graphs, Geodesics, and Gauges." arXiv:2104.13478, 2021. https://arxiv.org/abs/2104.13478
Gerken, Jan E., et al. "Geometric Deep Learning and Equivariant Neural Networks." Artificial Intelligence Review 56 (2023): 14605-14662. https://doi.org/10.1007/s10462-023-10502-7
Bruna, Joan, Wojciech Zaremba, Arthur Szlam, and Yann LeCun. "Spectral Networks and Locally Connected Networks on Graphs." ICLR 2014. https://arxiv.org/abs/1312.6203
Cohen, Taco S., and Max Welling. "Group Equivariant Convolutional Networks." ICML 2016. https://proceedings.mlr.press/v48/cohenc16.html
Monti, Federico, Davide Boscaini, Jonathan Masci, Emanuele Rodola, Jan Svoboda, and Michael M. Bronstein. "Geometric Deep Learning on Graphs and Manifolds Using Mixture Model CNNs." CVPR 2017. https://openaccess.thecvf.com/content_cvpr_2017/html/Monti_Geometric_Deep_Learning_CVPR_2017_paper.html
Kondor, Risi, and Shubhendu Trivedi. "On the Generalization of Equivariance and Convolution in Neural Networks to the Action of Compact Groups." ICML 2018. https://proceedings.mlr.press/v80/kondor18a.html
Gilmer, Justin, Samuel S. Schoenholz, Patrick F. Riley, Oriol Vinyals, and George E. Dahl. "Neural Message Passing for Quantum Chemistry." ICML 2017. https://proceedings.mlr.press/v70/gilmer17a.html
Cohen, Taco S., Maurice Weiler, Berkay Kicanaoglu, and Max Welling. "Gauge Equivariant Convolutional Networks and the Icosahedral CNN." ICML 2019. https://proceedings.mlr.press/v97/cohen19d.html
Fuchs, Fabian B., Daniel E. Worrall, Volker Fischer, and Max Welling. "SE(3)-Transformers: 3D Roto-Translation Equivariant Attention Networks." NeurIPS 2020. https://proceedings.neurips.cc/paper/2020/hash/15231a7ce4ba789d13b722cc5c955834-Abstract.html
Satorras, Victor Garcia, Emiel Hoogeboom, and Max Welling. "E(n) Equivariant Graph Neural Networks." ICML 2021. https://proceedings.mlr.press/v139/satorras21a.html
Batzner, Simon, Albert Musaelian, Lixin Sun, Mario Geiger, Jonathan P. Mailoa, Mordechai Kornbluth, Nicola Molinari, Tess E. Smidt, and Boris Kozinsky. "E(3)-Equivariant Graph Neural Networks for Data-Efficient and Accurate Interatomic Potentials." Nature Communications 13 (2022): 2453. https://www.nature.com/articles/s41467-022-29939-5
Gainza, Pablo, Freyr Sverrisson, Federico Monti, Emanuele Rodola, Davide Boscaini, Michael M. Bronstein, and Bruno E. Correia. "Deciphering Interaction Fingerprints from Protein Molecular Surfaces Using Geometric Deep Learning." Nature Methods 17 (2020): 184-192. https://www.nature.com/articles/s41592-019-0666-6
Xu, Keyulu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. "How Powerful Are Graph Neural Networks?" ICLR 2019. https://openreview.net/forum?id=ryGs6iA5Km
Joshi, Chaitanya K., Cristian Bodnar, Simon V. Mathis, Taco Cohen, and Pietro Lio. "On the Expressive Power of Geometric Graph Neural Networks." ICML 2023. https://proceedings.mlr.press/v202/joshi23a.html
Di Giovanni, Francesco, Lorenzo Giusti, Federico Barbero, Giulia Luise, Pietro Lio, and Michael M. Bronstein. "On Over-Squashing in Message Passing Neural Networks: The Impact of Width, Depth, and Topology." ICML 2023. https://proceedings.mlr.press/v202/di-giovanni23a.html
Wang, Rui, Robin Walters, and Rose Yu. "Approximately Equivariant Networks for Imperfectly Symmetric Dynamics." ICML 2022. https://proceedings.mlr.press/v162/wang22aa.html
Zou, Deyu, Shikun Liu, Siqi Miao, Victor Fung, Shiyu Chang, and Pan Li. "GeSS: Benchmarking Geometric Deep Learning under Scientific Applications with Distribution Shifts." NeurIPS 2024 Datasets and Benchmarks Track. https://proceedings.neurips.cc/paper_files/paper/2024/hash/a8063075b00168dc39bc81683619f1a8-Abstract-Datasets_and_Benchmarks_Track.html

Editing: Geometric Deep Learning