Quandela Demonstrates Photonic QPU for Quantum Machine Learning

A Quandela-led research collaboration - also including researchers from the Center for Theoretical Physics of the Polish Academy of Sciences and the University of Warsaw - experimentally demonstrated a programmable silicon-photonic QPU performing quantum reservoir processing for machine learning and single-basis quantum tomography, supported by the EU Horizon Europe QUONDENSATE Pathfinder project. The system routed non-classical multimode single-photon states from a semiconductor quantum dot through a 12-mode active demultiplexer into Quandela's 24-mode Belenos QPU, with photon-number-resolving (PNR) detectors and an electronic correlator at the output, per Quantum Computing Report and arXiv:2605.10471.

The reservoir is a non-trainable universal Bell-Walmsley interferometer mesh on a silicon chip, comprising integrated optical waveguides, mode couplers, and thermo-optic phase shifters. Single photons traverse a programmable Mach-Zehnder interferometer array performing complex unitary transformations via quantum interference. A software ridge-regression readout layer then processes 15-element PNR coincidence probability feature vectors for classification and tomography.

The QRP protocol executed quantum state tomography in a single, fixed measurement basis - contrast to standard tomography, which requires an exponential number of different bases. Mean density-matrix reconstruction fidelity reached 0.820 versus 0.747 for a direct PNR benchmark that failed to resolve off-diagonal coherences. The system additionally extracted purity, von Neumann entropy, and entanglement negativity. Feature space requirements scale quadratically with mode count, supporting a path to 3-mode (45 independent parameters) and larger characterization. A companion experiment on Quandela's legacy 12-mode Ascella QPU used hardware-aware perturbation training - injecting random unitary noise at the chip's measured transpilation error amplitude during optimization - and reached approximately 79.7% binary classification accuracy, outperforming an ideal classical simulation in the coherent-state regime.

Physical quantum reservoir computing maps nonlinear transformations natively in hardware, sidestepping barren-plateau optimization problems of parameterized quantum circuits. Photonic platforms operate at room temperature with native multi-mode interference, practical advantages over cryogenic alternatives for certain QML workloads. This is the first reported photonic QML system to successfully execute a quantum task (state tomography), not merely classical ML - that distinction is the paper's primary novelty claim.

Reproducibility at larger mode counts, integration of higher-efficiency PNR detectors, and benchmarking against classical and other quantum hardware on standard QML tasks. The quadratic feature-space scaling result is the most concrete near-term indicator of practical scalability.