Topological Quantum Neural Networks

• TQNNs are neural architectures that embed topological invariants and symmetry constraints into their design to recover global properties from local quantum data.
• They utilize diverse methodologies, including convolutional layers, quaternion operations, and hybrid quantum circuits, to compute invariants like winding and Chern numbers.
• The architecture aligns physical symmetries with computational models, enhancing error correction and phase recognition in complex quantum systems.

Topological Quantum Neural Networks (TQNNs) are neural architectures whose defining task is to encode, detect, or compute topological structure in quantum systems or in topological formalisms used to model computation. In current arXiv usage, the term spans several non-identical lineages: classical neural networks that infer winding or Chern numbers from local band-structure data (Zhang et al., 2017, Lin et al., 2022); quantum or hybrid quantum-classical networks that process many-body states directly to recognize symmetry-protected or intrinsically topological phases (Zapletal et al., 2023, Scarato et al., 26 Jun 2026, Hoffmann et al., 26 Jun 2026); and TQFT-based spin-network constructions in which intertwiners, cobordisms, and physical projectors constitute the computational substrate itself (Fields et al., 2022, Marciano et al., 2022, Fields et al., 2024). The literature therefore treats TQNNs less as a single standardized architecture than as a family of topology-aware models whose inductive biases are tied to symmetry, gauge structure, and global invariants.

1. Definitions, scope, and recurrent design principles

A common misconception is that a TQNN must be a neural network running on quantum hardware. The literature does not support that restriction. Some prototypical TQNNs are entirely classical: a small convolutional network can learn the winding number of a one-dimensional chiral insulator from local samples of the Bloch Hamiltonian (Zhang et al., 2017), and short-range restricted Boltzmann machines can represent the 1D cluster state and the 2D and 3D toric code states exactly and efficiently (Deng et al., 2016). Other works use quantum circuits or hybrid quantum-classical pipelines, while another branch defines TQNNs directly as TQFT boundary spin-network computations (Zapletal et al., 2023, Fields et al., 2024).

Despite this terminological divergence, several structural motifs recur. First, TQNNs are designed to recover global, nonlocal quantities from local or quasi-local input data. Second, they are typically symmetry-aware: chiral symmetry in one-dimensional band topology, gauge patching in quaternionic Chern-insulator pipelines, or gauge invariance and flatness constraints in BF-theoretic spin-network models. Third, successful designs generally align architecture with the discrete formula of the invariant to be learned. This can mean finite-difference kernels over momentum space, pooling layers that implement error correction, or intertwiner networks whose admissibility rules already encode the relevant algebraic constraints. This suggests that the distinctive feature of TQNNs is not merely the presence of “topology” in the dataset, but the embedding of topological structure into the model class itself.

2. Learning topological invariants from local quantum data

A canonical example is the learning of the winding number in one-dimensional chiral-symmetric insulators. In the two-band AIII case, one may write

$$H(k)=h_x(k)\sigma_x+h_y(k)\sigma_y,\qquad U(k)=\tilde h_x(k)+i\tilde h_y(k),$$

with winding number

$$w=-\frac{i}{2\pi}\int_0^{2\pi} U^*(k)\,\partial_k U(k)\,dk.$$

The network input is the $(L+1)\times 2$ matrix of normalized Bloch-Hamiltonian components $\tilde h_x(k),\tilde h_y(k)$ sampled uniformly on $k\in[0,2\pi]$, with results reported as insensitive once $L\ge 32$. The output is a regressed real number $\tilde w$, interpreted by rounding to the nearest integer (Zhang et al., 2017).

The crucial result is not just high in-range accuracy, but extrapolation beyond the winding numbers present in the training set. The convolutional architecture uses two convolutional layers, with a first $2\times 2$ kernel acting on adjacent $k$-points and the two input channels, then a $1\times 1$ layer, a small fully connected head, and a linear output. Its interpretability analysis shows that the first layer implements a differentiable phase-unwrapping finite difference, and the final layers sum local increments so as to recover the discrete winding formula

$$w=-\frac{i}{2\pi}\int_0^{2\pi} U^*(k)\,\partial_k U(k)\,dk.$$0

The model therefore computes a topological invariant from local samples rather than merely fitting labels (Zhang et al., 2017).

Test set	CNN	FC
Held-out $$w=-\frac{i}{2\pi}\int_0^{2\pi} U^*(k)\,\partial_k U(k)\,dk.$$1	99.6%	82.2%
Adversarial local-twist, $$w=-\frac{i}{2\pi}\int_0^{2\pi} U^*(k)\,\partial_k U(k)\,dk.$$2	100.0%	99.1%
Unseen $$w=-\frac{i}{2\pi}\int_0^{2\pi} U^*(k)\,\partial_k U(k)\,dk.$$3	98.2%	22.8%
Unseen $$w=-\frac{i}{2\pi}\int_0^{2\pi} U^*(k)\,\partial_k U(k)\,dk.$$4	99.3%	1.4%

Two additional observations are methodologically important. Training only on SSH Hamiltonians induces a shortcut based on inversion symmetry, namely reading off $$w=-\frac{i}{2\pi}\int_0^{2\pi} U^*(k)\,\partial_k U(k)\,dk.$$5, which generalizes poorly outside SSH. Regularization also behaves atypically: in the clean synthetic setting, larger $$w=-\frac{i}{2\pi}\int_0^{2\pi} U^*(k)\,\partial_k U(k)\,dk.$$6 weight regularization hurts extrapolation to larger winding numbers, whereas with 4% label noise it becomes beneficial and restores generalization (Zhang et al., 2017).

Quaternion-based models extend the same local-to-global program to two-dimensional Chern insulators. For

$$w=-\frac{i}{2\pi}\int_0^{2\pi} U^*(k)\,\partial_k U(k)\,dk.$$7

the phase diagram obeys $$w=-\frac{i}{2\pi}\int_0^{2\pi} U^*(k)\,\partial_k U(k)\,dk.$$8 for $$w=-\frac{i}{2\pi}\int_0^{2\pi} U^*(k)\,\partial_k U(k)\,dk.$$9 and $(L+1)\times 2$0 for $(L+1)\times 2$1. The lower- and upper-band eigenstates are mapped into quaternions, and a gauge-dependent quaternionic convolution feature

$(L+1)\times 2$2

is used for unsupervised PCA, while supervised learning uses a qCNN with a first quaternion-convolution layer followed by conventional convolutional layers. In the unsupervised pipeline, two principal components separate only four groups because of a modulo-4 ambiguity, whereas including components up to PC6 fully separates all seven classes. In the supervised pipeline, the qCNN achieves 99.67% accuracy across nine classes, compared with 94.12% for the real-valued CNN baseline, and it performs markedly better on out-of-distribution categories such as sign-flips, component swaps, and trivial ferromagnetic, helical, and conical textures (Lin et al., 2022).

The quaternion construction is significant because it supplies an explicit inductive bias for local solid-angle estimation. The first $(L+1)\times 2$3 quaternion convolution aggregates nearest-neighbor spins in a plaquette, and the Hamilton product together with an arctan nonlinearity is argued to align with the analytic form

$(L+1)\times 2$4

thereby approximating local Berry curvature. At the same time, the unsupervised feature $(L+1)\times 2$5 is not gauge-invariant, so the approach exposes both the power and the fragility of representation engineering in TQNN design (Lin et al., 2022).

3. Quantum and hybrid architectures for topological phase recognition

Quantum and hybrid TQNNs process quantum states directly rather than inferring invariants from precomputed classical data. A leading example is the QCNN for generalized cluster–Ising chains with $(L+1)\times 2$6 symmetry and two distinct SPT phases, “ZXZ” and “ZXXXZ”. The QCNN alternates convolutional layers with pooling layers that act as QEC-like error-correcting maps, followed by a final measurement in the $(L+1)\times 2$7 basis. Its output is

$(L+1)\times 2$8

For symmetry-preserving $(L+1)\times 2$9-noise, the effective error rate obeys

$\tilde h_x(k),\tilde h_y(k)$0

and the QCNN output is

$\tilde h_x(k),\tilde h_y(k)$1

Alternating $\tilde h_x(k),\tilde h_y(k)$2- and $\tilde h_x(k),\tilde h_y(k)$3-correcting layers yields explicit symmetry-breaking thresholds $\tilde h_x(k),\tilde h_y(k)$4 for the ZXZ cluster and $\tilde h_x(k),\tilde h_y(k)$5 for the ZXXXZ cluster, while a constant-depth implementation becomes possible by replacing most of the logarithmic-depth quantum processing with classical post-processing. The reported sample complexity reduction over direct local measurements is exponential in system size (Zapletal et al., 2023).

Hybrid quantum-classical models replace the deep QCNN by a shallow parameterized quantum circuit (PQC), full computational-basis measurement, and a classical neural head. In the surface-code setting, one such model combines hardware-efficient $\tilde h_x(k),\tilde h_y(k)$6 and $\tilde h_x(k),\tilde h_y(k)$7 gates with a small feedforward network and is trained jointly with superconducting hardware in the loop. The task is to distinguish topological ground states of the surface code in a longitudinal magnetic field from a featureless ensemble of product states formed from tensor products of Pauli eigenstates. The classifier achieves accuracies above 85% in single-shot measurements and above 99% when averaging over ten measurements, while remaining robust against any single-qubit Pauli error on the topological state (Scarato et al., 26 Jun 2026).

A closely related hybrid architecture is optimized explicitly to maximize the statistical distance between phase-dependent measurement distributions. For two classes with distributions $\tilde h_x(k),\tilde h_y(k)$8 and $\tilde h_x(k),\tilde h_y(k)$9, the optimal residual cost is

$k\in[0,2\pi]$0

with Bayes error

$k\in[0,2\pi]$1

In the $k\in[0,2\pi]$2 surface-code-versus-1-design task, randomized Pauli-6 measurements yield $k\in[0,2\pi]$3, whereas the trained PQC reaches $k\in[0,2\pi]$4. Single-shot error falls from 23% for a random circuit to $k\in[0,2\pi]$5 after training, and multi-shot inference attains false negatives $k\in[0,2\pi]$6 with $k\in[0,2\pi]$7 and false positives $k\in[0,2\pi]$8 with $k\in[0,2\pi]$9. In the surface-code-versus-symmetry-enriched-topological-phase task, comparable error rates require $L\ge 32$0, while the classical baseline trained on randomized Pauli measurements needs $L\ge 32$1 (Hoffmann et al., 26 Jun 2026).

Architecture	System	Reported result
QCNN	Generalized cluster–Ising SPT phases	$L\ge 32$2 for ZXZ, $L\ge 32$3 for ZXXXZ; constant-depth implementation; exponential sample-complexity reduction
Hybrid PQC + classical NN	Surface code in a longitudinal field	Above 85% single-shot accuracy; above 99% with ten-shot averaging
Hybrid sample-efficient recognizer	Surface code vs 1-design and symmetry-enriched topological phase	$L\ge 32$4 after training for $L\ge 32$5; $L\ge 32$6 false negatives with $L\ge 32$7 and $L\ge 32$8 false positives with $L\ge 32$9 in the topological-vs-1-design task

These models share a specific design logic. The quantum component is shallow but nonlocal: it learns a basis in which topological information becomes visible in simple measurement statistics. The classical head then performs a comparatively low-complexity discrimination problem on bit strings rather than reconstructing a nonlocal order parameter from randomized local measurements. In that sense, the quantum circuit acts less as a generic variational ansatz and more as a topology-aligned measurement preconditioner.

4. TQFT, spin networks, and error-correcting interpretations

A different and more formal branch of the literature identifies TQNNs with TQFT-based computational systems. In this usage, sequential finite measurements are represented by cone–cocone diagrams of Barwise–Seligman classifiers, and there exists a functor

$\tilde w$0

which maps measurement semantics into finite cobordisms of Hilbert spaces. Composing $\tilde w$1 with a TQFT functor yields dynamics on a measured sector, while the neural computation is realized on topological quantum neural 2-complexes (TQN2Cs). Inputs and outputs are spin-network states in $\tilde w$2, and faces impose flatness through projectors $\tilde w$3 (Fields et al., 2022).

The same general program appears in earlier work mapping quantum neural networks to spin networks. Boundary states are cylindrical functions of holonomies on a graph $\tilde w$4, edges carry SU(2) labels $\tilde w$5, nodes carry intertwiners, and the forward pass is a TQFT state sum on a 2-complex. In the semiclassical, large-spin regime, coherent-state amplitudes become Gaussian peaks around classical labels $\tilde w$6, and deep neural networks emerge as semiclassical limits of the underlying topological quantum model (Marciano et al., 2020). Later work develops this into an explicit theory of supervised learning in BF-like topological field theory, with the classifier written as

$\tilde w$7

where $\tilde w$8 is a physical projector imposing flatness. In that account, topological capacity and metric capacity are separated, and overfitting in classical DNNs is interpreted as a pathology of metric-only models that have lost direct access to topological structure (Marciano et al., 2022).

In the holographic and error-correcting literature, TQNNs become boundary spin-network computational systems supported by TQFTs. A TQFT is a symmetric monoidal functor

$\tilde w$9

and on a boundary graph $2\times 2$0 with $2\times 2$1 links and $2\times 2$2 nodes, the Hilbert space is

$2\times 2$3

Links are labeled by irreducible representations $2\times 2$4, nodes by intertwiners $2\times 2$5, and the “neurons” are invariant contractions at the vertices. The tensor-network realization defines an encoding isometry

$2\times 2$6

and the resulting code satisfies Knill–Laflamme conditions. In this picture, HaPPY networks are paradigmatic TQNNs: perfect intertwiners implement the local computation, minimal cuts control entanglement, and the redundancy structure induces effective spacetimes supporting LOCC between boundary agents (Fields et al., 2024).

A further step is the claim that TQNNs support universal quantum computation and admit a computation–scattering correspondence. In the Reshetikhin–Turaev and Turaev–Viro formulation, a TQNN is a TQFT realized on a spin-network basis, with braid generators $2\times 2$7, recoupling operators $2\times 2$8, and TV code subspaces defined as the range of a boundary projector $2\times 2$9. Two formal statements summarize this program: Theorem 1 states, “A TQNN provides quantum processes for UQC,” and Theorem 2 states, “Within a TQNN, a UQC corresponds to a scattering process with amplitudes given by an amplituhedron, and conversely” (Fields et al., 24 Sep 2025).

Taken together, these works define a distinctly nonstandard notion of “neural network.” Here the network is not primarily a trainable function approximator, but a topological computational geometry whose local maps are intertwiners, whose global dynamics are cobordism amplitudes, and whose robustness derives from topological or code-theoretic protection.

5. State representations, simplicial models, and exact evaluation

TQNN ideas also appear in exact neural representations of topological many-body states. Short-range restricted Boltzmann machines provide exact and efficient representations of the 1D symmetry-protected topological cluster state and the 2D and 3D toric-code states. For all three cases, the required number of hidden neurons is as small as the number of physical spins, and the number of parameters scales linearly with system size. In the 2D toric code, $k$0, with one hidden neuron per vertex and one per face. The same construction represents excited states with abelian anyons and their mutual statistics. Beyond exactly solvable cases, reinforcement learning with RBM ansätze applied to the non-integrable Hamiltonian

$k$1

reaches relative energy error $k$2 for $k$3 and hidden density $k$4, and identifies a phase transition around $k$5 at $k$6 (Deng et al., 2016).

Quantum simplicial networks generalize the topological deep learning perspective to quantum circuits on simplicial complexes. A QSN assigns one qubit per simplex and uses incidence matrices $k$7 and Hodge Laplacians

$k$8

to determine which qubits interact. The Quantum Simplicial Layer encodes simplex features by single-qubit rotations, applies learnable embeddings, and then performs topology-driven two-qubit interactions only on lower- or upper-adjacent simplices. This is the first quantum topological deep learning model in the cited literature. On synthetic solenoidal-component detection, the schematic QSN reaches accuracies up to 0.957 while classical GSCN performance degrades to 0.598 at depth five; on source localization, the base QSN reaches up to 0.702, compared with 0.665 for the QGNN baseline (Piperno et al., 9 Jan 2025).

For spin-network TQNNs, exact evaluation of the physical scalar product is itself a central algorithmic problem. For hexagonal honeycomb spin-networks, the Noui–Perez physical inner product factorizes as

$k$9

after gauge-fixing and Haar-integration identities decouple the two networks. The evaluation algorithm proceeds by repeated bubble moves, recoupling into $1\times 1$0-symbols, and an inductive decomposition

$1\times 1$1

The number of summations obeys

$1\times 1$2

so the summation count grows quadratically in $1\times 1$3. Even so, exhaustive storage rapidly becomes infeasible: for a $1\times 1$4 honeycomb, $1\times 1$5 cycles and $1\times 1$6 pair-of-cycle configurations already exceed 32-bit RAM capacity, so exhaustive results are reported only for $1\times 1$7 and $1\times 1$8 (Lulli et al., 2023).

These constructions broaden the TQNN landscape in two directions. One direction treats neural networks as exact ansätze for topological quantum states. The other treats topological combinatorics—simplicial structure, recoupling theory, or spin-network scalar products—as the primary computational primitive.

6. Limitations, misconceptions, and open directions

The strongest source of confusion in this area is terminological. The same label is applied to classical CNNs that regress winding numbers, hybrid PQC classifiers for topological phases, and TQFT spin-network machines whose “neurons” are intertwiners. This suggests that claims about “the TQNN architecture” are usually too coarse to be informative. Whether a model is trainable by backpropagation, protected by a code subspace, or interpretable as a TQFT amplitude depends on which branch of the literature is under discussion (Zhang et al., 2017, Fields et al., 2024).

Several technical limitations recur. Gauge dependence remains unresolved in some pipelines: the quaternion PCA feature $1\times 1$9 is not gauge-invariant and depends on carefully selected patching (Lin et al., 2022). Shortcut learning is a documented risk: SSH-only training induces classification from inversion-symmetric data at $$w=-\frac{i}{2\pi}\int_0^{2\pi} U^*(k)\,\partial_k U(k)\,dk.$$00 rather than from the winding formula itself (Zhang et al., 2017). In noise-tolerant QCNNs, the symmetry-breaking threshold decreases near phase boundaries according to

$$w=-\frac{i}{2\pi}\int_0^{2\pi} U^*(k)\,\partial_k U(k)\,dk.$$01

so error tolerance weakens as the correlation length diverges (Zapletal et al., 2023). In hybrid surface-code classifiers, the strongest results rely on shallow, problem-informed circuits derived from inverse state-preparation constructions, and the demonstrated hardware systems are finite, with 3×3 and 4×4 lattices in experiment and up to 5×5 in the companion numerical study (Scarato et al., 26 Jun 2026, Hoffmann et al., 26 Jun 2026). In spin-network TQNNs, exact scalar-product evaluation remains a bottleneck despite recoupling-based acceleration (Lulli et al., 2023).

Future directions are comparatively consistent across otherwise disparate papers. Recurrent proposals include explicit gauge-invariant layers based on link variables, Wilson loops, or discrete Berry-flux estimators; extensions from two-band to multi-band models and from 1D or 2D examples to higher-dimensional invariants; and richer algebraic structures such as Clifford or biquaternion layers for 3D topological insulators, chiral superconductors, and non-Hermitian systems (Lin et al., 2022). Another line of work moves the topological information into the parameter-space definition of the model itself. Neural network field theory with explicit discrete topological labels $$w=-\frac{i}{2\pi}\int_0^{2\pi} U^*(k)\,\partial_k U(k)\,dk.$$02 reproduces the BKT transition, the RG equations

$$w=-\frac{i}{2\pi}\int_0^{2\pi} U^*(k)\,\partial_k U(k)\,dk.$$03

and T-duality under $$w=-\frac{i}{2\pi}\int_0^{2\pi} U^*(k)\,\partial_k U(k)\,dk.$$04 with $$w=-\frac{i}{2\pi}\int_0^{2\pi} U^*(k)\,\partial_k U(k)\,dk.$$05. That construction suggests TQNNs with sector registers, duality gates, and training trajectories interpreted directly as RG flow (Ferko et al., 2 Apr 2026).

A plausible implication is that the next stage of TQNN research will be less concerned with choosing between “classical,” “quantum,” and “topological” definitions than with combining them: explicit invariant modules for gauge or symmetry protection, shallow quantum preprocessing for sample-efficient measurement, and topological or code-theoretic structure to control generalization and robustness.