Photonic Quantum Machine Learning

• Photonic quantum machine learning (PQML) is an emerging field that uses optical circuits and photon-based encoding to implement quantum models directly on photonic hardware.
• PQML leverages diverse architectures—discrete-variable, continuous-variable, and hybrid—to perform tasks such as state tomography, classification, and kernel evaluation.
• Recent advances include variational models, benchmarking platforms, and automated design tools that integrate classical optimization with photonic quantum processing.

Searching arXiv for recent and foundational papers on photonic quantum machine learning to ground the article in published work. Photonic quantum machine learning (PQML) denotes machine learning with or through photonic quantum circuits: photons and optical circuitry are used to encode data, process information, generate feature maps or kernels, and train or evaluate quantum models. In the published literature, PQML spans discrete-variable, continuous-variable, and hybrid photonic architectures; variational quantum circuits and quantum neural networks; fixed-reservoir and random-feature models; kernel methods based on photonic state overlaps; and machine-learning-assisted design, calibration, and programming of photonic quantum hardware (Alagiyawanna et al., 10 Mar 2026, Notton et al., 29 Oct 2025, Notton et al., 11 Feb 2026).

1. Scope and conceptual boundaries

PQML is defined in two closely related senses. The first is quantum machine learning implemented on photonic hardware, in which linear-optical interferometers, bosonic states, continuous-variable circuits, or single-photon gate-model processors act as the trainable model, the feature map, or the physical reservoir. The second is machine learning for photonic quantum technology, in which classical optimization or learning methods are used to synthesize photonic states and gates, or to program and calibrate integrated interferometers whose internal structure is only partially characterized (Arrazola et al., 2018, Sabapathy et al., 2018, Stanev et al., 11 May 2026).

This boundary is broader than a qubit-centric view of quantum machine learning. Several papers emphasize that photonics is not merely “quantum hardware for qubits,” but a platform with photon-native computational primitives, including Fock-state encodings, boson-sampling-style interference, homodyne and photon-number-resolving measurements, and high-dimensional Hilbert spaces. A plausible implication is that PQML should be treated as a family of hardware-native learning paradigms rather than as a single circuit template transplanted from other quantum platforms (Notton et al., 29 Oct 2025, Gan et al., 2021).

The field also spans both Quantum for AI and AI for Quantum. On the former side are photonic classifiers, kernels, reservoirs, and hybrid neural networks; on the latter are automated photonic circuit synthesis, non-Gaussian resource-state engineering, and black-box programming of reconfigurable photonic chips. This dual usage is explicit in software platforms that support photonic quantum circuits as differentiable modules inside machine-learning workflows while also targeting photonic algorithm design and hardware co-design (He et al., 22 Dec 2025, Notton et al., 11 Feb 2026).

2. Photonic substrates, encodings, and observables

The main photonic architectures used in PQML are discrete-variable (DV), continuous-variable (CV), and hybrid photonic quantum computing. In DV photonics, information is carried by discrete degrees of freedom such as polarization or path; in CV photonics, it is encoded in continuous observables such as amplitude and phase; hybrid systems combine both and inherit the strengths and vulnerabilities of each paradigm (Alagiyawanna et al., 10 Mar 2026).

In linear-optical PQML, information is commonly encoded in Fock states of photons across optical modes, and circuits are built from beam splitters and phase shifters. The output statistics are governed by matrix permanents. A standard representation used across several works is that an input state $\ket{\phi(x)}$ is produced by a data-dependent unitary $U(x)$ acting on a reference Fock state, and kernels are defined by fidelity overlaps,

$$\ket{\phi(x)} = U(x)\ket{s}, \qquad k(x_1,x_2)=|\langle \phi(x_1)\mid\phi(x_2)\rangle|^2 .$$

The accessible bosonic feature space grows combinatorially with photon number and mode number; for $n$ photons in $m$ modes, the Fock-space dimension is

$d=\binom{n+m-1}{n}.$

This high-dimensionality is one of the central physical resources invoked in photonic data embedding and multi-photon learning-capacity arguments (Notton et al., 29 Oct 2025, Gan et al., 2021, Wang et al., 26 Nov 2025).

CV PQML uses Gaussian and non-Gaussian optical operations such as displacement, squeezing, interferometers, homodyne detection, and, in some architectures, Kerr nonlinearities. In Gaussian formulations, states are represented by covariance matrices and displacement vectors; in non-Gaussian formulations, explicit Fock-state or bosonic decompositions are required. These representational choices are not merely formal: they determine whether training is exact but small-scale, Gaussian and efficiently simulable, or expressive enough to support non-Gaussian resource engineering (He et al., 22 Dec 2025, Arrazola et al., 2018).

Platform-specific encodings illustrate the diversity of PQML substrates. A single-photon QELM encoded unknown qubit states in polarization while using orbital angular momentum (OAM) as a high-dimensional coined quantum-walk reservoir (Suprano et al., 2023). Gate-based photonic QNNs have used dual-rail qubit encoding on KLM-style single-photon hardware (McKiernan et al., 11 May 2026). Other works use coherent reference states, multi-photon Fock inputs, or quadrature displacements into qumodes, depending on whether the target task is tomography, classification, kernel estimation, or fast analog random projection (Wang et al., 2021, Maier et al., 15 Oct 2025).

3. Principal learning paradigms

PQML includes several distinct model classes that differ in where trainability is placed: in the optical circuit, in the measurement, in the classical readout, or in the hardware-design loop itself.

A major family is the variational photonic model. Here, a parameterized photonic circuit is optimized end-to-end with a classical optimizer. Representative examples include active-learning-enabled variational quantum classifiers on a programmable free-space photonic processor, hybrid continuous-variable photonic quantum-classical neural networks, and small gate-based photonic QNNs implemented on single-photon hardware (Ding et al., 2022, Austin et al., 2024, McKiernan et al., 11 May 2026). In the active-learning experiments, integrating uncertainty sampling or query-by-committee into photonic VQC training saved at most $85\%$ labeling efforts and $91.6\%$ computational efforts relative to training without active learning, while maintaining the trained model’s performance (Ding et al., 2022).

A second family is the kernel and feature-map approach, where the photonic device defines a map from classical data to quantum states and a classical learner operates on the resulting kernel matrix or feature vector. Adaptive Boson Sampling via post-selection is an explicit example: experimentally reconstructed state overlaps were used as kernels for SVM classification, and moving from qubit to qutrit outputs improved the reported 2D make_moons() task from $A=0.65$ to $A=0.90$ (Hoch et al., 27 Feb 2025). The Perceval Challenge also included photonic kernel SVMs, photonic quantum neural networks, and boson-sampling-inspired feature extractors as benchmarked methods on reduced 10-class MNIST (Notton et al., 29 Oct 2025).

A third family is the reservoir or extreme-learning machine formulation. In these models, the photonic dynamics are fixed or randomly initialized and only a simple classical readout is trained. In a continuous-variable photonic QELM, classical detector features were embedded through quadrature displacements, propagated through a fixed Gaussian substrate, and read out via homodyne or photon-number-compatible measurements; only the linear readout was trained, and the model outperformed an MLP with two hidden units for all considered training sizes on top-jet tagging and Higgs-boson identification (Maier et al., 15 Oct 2025). This suggests a distinct PQML regime in which the photonic device acts as a physical random-feature generator rather than a fully optimized quantum circuit.

A fourth family uses machine learning as a design and control tool for photonic quantum hardware. Continuous-variable quantum neural networks have been trained to synthesize single photons, GKP states, NOON states, cubic phase gates, random unitaries, and cross-Kerr interactions, routinely obtaining fidelities above $U(x)$0 with short-depth circuits (Arrazola et al., 2018). Machine learning has also been used to optimize non-Gaussian resource-state preparation for weak cubic phase gates, raising success probability to at least $U(x)$1, an increase by a factor of $U(x)$2 over standard sequential photon-subtraction techniques (Sabapathy et al., 2018). More recently, black-box stochastic training procedures have been proposed for continuously coupled reconfigurable interferometers using a limited number of single- and two-photon measurements (Stanev et al., 11 May 2026).

Paradigm	Representative role of photonics	Representative papers
Variational models	Trainable circuit or quantum hidden layer	(Ding et al., 2022, Austin et al., 2024, McKiernan et al., 11 May 2026)
Kernels and feature maps	State overlap, boson-sampling embedding, photonic annotator	(Hoch et al., 27 Feb 2025, Notton et al., 29 Oct 2025)
Reservoirs and QELMs	Fixed substrate, trained linear readout	(Suprano et al., 2023, Maier et al., 15 Oct 2025, Carreño et al., 11 May 2026)
Hardware design and programming	Learned state preparation, gate synthesis, black-box chip tuning	(Arrazola et al., 2018, Sabapathy et al., 2018, Stanev et al., 11 May 2026)

4. Quantum-state characterization and quantum-property inference

One of the most developed PQML application areas is quantum-state characterization. In a photonic quantum extreme learning machine implemented with single photons, the unknown input qubit was encoded in polarization, the reservoir was realized as a coined quantum walk in OAM, and projective measurements over a fixed OAM basis generated the classical features. The trained readout reconstructed the expectation values of the Pauli observables $U(x)$3. The paper reports that the median MSE already decreases significantly with as few as five training states, that the setup does not need precise characterization of the measurement apparatus, and that the optimized reservoir is only marginally better than a random choice of coin parameters. At the same time, the specific two-step implementation was not informationally complete, so the demonstration is best viewed as property reconstruction rather than universal tomography (Suprano et al., 2023).

A related line of work uses photonic patterns produced by simple interference circuits as machine-learning inputs. In a four-mode linear-optical network with three 50:50 beam splitters, coherent reference states interfered with an unknown photonic state, and the output photon-number distribution was processed by supervised regressors to predict state parameters or entanglement. For single-mode tomography, the reported mean fidelity reached about $U(x)$4 for $U(x)$5 without PCA and $U(x)$6 with PCA using Extremely Randomized Trees; for entanglement estimation, the best ERT+PCA setting yielded MAEs about $U(x)$7 for $U(x)$8, $U(x)$9 for $$\ket{\phi(x)} = U(x)\ket{s}, \qquad k(x_1,x_2)=|\langle \phi(x_1)\mid\phi(x_2)\rangle|^2 .$$0, and $$\ket{\phi(x)} = U(x)\ket{s}, \qquad k(x_1,x_2)=|\langle \phi(x_1)\mid\phi(x_2)\rangle|^2 .$$1 for $$\ket{\phi(x)} = U(x)\ket{s}, \qquad k(x_1,x_2)=|\langle \phi(x_1)\mid\phi(x_2)\rangle|^2 .$$2 (Wang et al., 2021). The underlying idea is that the photonic circuit performs a complicated nonlinear transformation that is analytically tractable only in small cases, making learned inverse maps practical.

Integrated photonic reservoir processing has extended this theme to mixed-state tomography and entanglement estimation on hardware. A programmable silicon photonic chip excited with single photons was used to perform quantum state tomography and to estimate negativity, purity, and von Neumann entropy. For two-photon mixed-state tomography, the reported testing fidelity was $$\ket{\phi(x)} = U(x)\ket{s}, \qquad k(x_1,x_2)=|\langle \phi(x_1)\mid\phi(x_2)\rangle|^2 .$$3, compared with a PNR benchmark of $$\ket{\phi(x)} = U(x)\ket{s}, \qquad k(x_1,x_2)=|\langle \phi(x_1)\mid\phi(x_2)\rangle|^2 .$$4. For a two-mode state of the form

$$\ket{\phi(x)} = U(x)\ket{s}, \qquad k(x_1,x_2)=|\langle \phi(x_1)\mid\phi(x_2)\rangle|^2 .$$5

the negativity satisfies

$$\ket{\phi(x)} = U(x)\ket{s}, \qquad k(x_1,x_2)=|\langle \phi(x_1)\mid\phi(x_2)\rangle|^2 .$$6

The same work also introduced a perturbation-based mitigation method for experimental imperfections and reported a significant improvement in accuracy in comparison to the same system operating in the classical regime (Carreño et al., 11 May 2026).

These works collectively establish a characteristic PQML use case: replacing large families of calibrated measurement settings or explicit inversion formulas with fixed photonic transformations plus learned estimators. This suggests that PQML is especially well aligned with tasks where the target quantity is a property of a quantum state and the measurement apparatus is difficult to model exactly.

5. Benchmarking, software infrastructure, and automated design

A central development in PQML has been the move from isolated demonstrations to shared benchmarks and reusable software. The clearest benchmark is the Perceval Challenge, an open, reproducible evaluation on a reduced but hardware-feasible 10-class MNIST task. It ran for more than three months, drew 64 teams worldwide in its first phase, selected 11 finalist teams for GPU resources and cloud photonic hardware execution, and established the first unified baseline landscape for photonic machine learning. The benchmark tracked not only accuracy but also parameter count, FLOPs, and convergence speed. Its classical baseline CNN reached about $$\ket{\phi(x)} = U(x)\ket{s}, \qquad k(x_1,x_2)=|\langle \phi(x_1)\mid\phi(x_2)\rangle|^2 .$$7 test error on the reduced dataset; photonic and hybrid submissions varied widely, and the authors explicitly stated that no clear heuristic quantum advantage was observed (Notton et al., 29 Oct 2025).

This benchmarking turn is supported by software platforms designed for differentiable, photonic-native experimentation. MerLin integrates strong simulation of linear-optical circuits into PyTorch and scikit-learn, reproduces eighteen state-of-the-art photonic and hybrid QML works, and is designed around systematic benchmarking, modular experiments, and hardware-aware execution (Notton et al., 11 Feb 2026). DeepQuantum provides a PyTorch-based platform with Fock, Gaussian, and Bosonic backends, and describes itself as the first framework to realize closed-loop integration of quantum circuits, photonic quantum circuits, and measurement-based quantum computing (He et al., 22 Dec 2025). A plausible implication is that software standardization is becoming part of PQML’s scientific methodology rather than merely an implementation convenience.

Automated architecture and compression methods have also entered the field. Q-PhotoNAS uses a genetic-algorithm neural architecture search over 19 hyperparameters across six gene groups, jointly searching classical preprocessing, learnable quantum phase encoding, photonic-layer structure, classifier head, and training hyperparameters. It reported final validation accuracies of $$\ket{\phi(x)} = U(x)\ket{s}, \qquad k(x_1,x_2)=|\langle \phi(x_1)\mid\phi(x_2)\rangle|^2 .$$8 on Digits and $$\ket{\phi(x)} = U(x)\ket{s}, \qquad k(x_1,x_2)=|\langle \phi(x_1)\mid\phi(x_2)\rangle|^2 .$$9 on MNIST, with first-principles Ascella execution-time estimates of $n$0 ms and $n$1 ms per image, respectively (Elnakhal et al., 21 May 2026). In a different hybrid direction, a distributed photonic Quantum-Train framework used photonic QNNs plus an MPS mapping to generate classical CNN parameters, reporting $n$2 test accuracy with 3,292 parameters at $n$3, compared with $n$4 for a classical baseline with 6,690 parameters; at $n$5, it achieved about a 10× compression ratio with a relative accuracy loss below $n$6 (Chen et al., 13 May 2025).

Taken together, these developments indicate that contemporary PQML research is increasingly organized around three infrastructure layers: reproducible benchmarks, trainable photonic software stacks, and automated search or compression procedures that respect photonic hardware constraints.

6. Noise, performance claims, and unresolved questions

Noise is a recurring limiting factor across PQML. The main mechanisms identified in the review literature are photon loss, mode mismatch, phase noise, and thermal noise, all of which can distort interference, bias expectation values, destabilize gradients, and lower predictive accuracy. In the formal description of noisy inference,

$n$7

the discrepancy from the ideal expectation value directly affects both training and deployment (Alagiyawanna et al., 10 Mar 2026). This has motivated hardware-level mitigation, noise-aware training, error-mitigation strategies, and classical post-processing that learns to compensate for device bias.

A recurring controversy concerns the status of quantum advantage. Benchmark studies have been cautious: on reduced MNIST, no clear heuristic quantum advantage was observed, and the principal contribution was a unified baseline rather than classical outperformance (Notton et al., 29 Oct 2025). Some photonic architectures are also explicit about not targeting asymptotic supremacy; the CV-QELM for collider data, for example, uses a Gaussian and classically simulable substrate, arguing for deterministic timing, compact random features, and rapid retraining rather than non-Gaussian quantum advantage (Maier et al., 15 Oct 2025).

At the same time, several papers report more specific forms of advantage. A multi-photon PQML study used the data quantum Fisher information matrix to define learning capacity and derived the critical dataset size

$n$8

It argued that learning capacity scales polynomially with photon number and experimentally showed that, in a unitary-learning task, two photons generalized with $n$9 examples whereas single photons required $m$0 (Wang et al., 26 Nov 2025). A gate-based photonic QNN study reported effective dimension $m$1 for a 2-parameter photonic QNN versus $m$2 for a matched 2-parameter ANN, and on XOR the photonic QNN converged to loss $m$3 and $m$4 accuracy while the matched ANN remained at random-guessing performance (McKiernan et al., 11 May 2026). These are substantial results, but they are task-specific, architecture-specific, and parameter-matched rather than broad demonstrations of universal practical superiority.

Important limitations remain explicit in the literature. Adaptive Boson Sampling currently emulates adaptivity through post-selection because chip reconfiguration is not fast enough for real-time feed-forward, and tomography-based kernel estimation is feasible only at low output dimension (Hoch et al., 27 Feb 2025). The photonic QELM for polarization reconstruction does not yet provide full tomography in its demonstrated two-step form (Suprano et al., 2023). High-dimensional or non-Gaussian photonic simulations remain computationally demanding, which is why tensor-network methods, distributed simulation, and strong-simulation backends have become central software concerns (He et al., 22 Dec 2025, Notton et al., 11 Feb 2026).

The main research directions repeatedly identified across the field are therefore not singular. They include scaling to larger mode and photon numbers, introducing fast active feed-forward, extending beyond Gaussian or post-selected regimes, improving calibration-free but noise-aware training, standardizing stronger benchmarks, and tightening the co-design loop between algorithms, software, and hardware (Hoch et al., 27 Feb 2025, Alagiyawanna et al., 10 Mar 2026, Notton et al., 11 Feb 2026). This suggests that PQML is best understood as an evolving experimental and algorithmic ecosystem in which photonic hardware is not merely a substrate for imported quantum-learning ideas, but a source of distinctive learning models, constraints, and opportunities.