MerLin: A Discovery Engine for Photonic and Hybrid Quantum Machine Learning

Identifying where quantum models may offer practical benefits in near term quantum machine learning (QML) requires moving beyond isolated algorithmic proposals toward systematic and empirical exploration across models, datasets, and hardware constraints. We introduce MerLin, an open source framework designed as a discovery engine for photonic and hybrid quantum machine learning. MerLin integrates optimized strong simulation of linear optical circuits into standard PyTorch and scikit learn workflows, enabling end to end differentiable training of quantum layers. MerLin is designed around systematic benchmarking and reproducibility. As an initial contribution, we reproduce eighteen state of the art photonic and hybrid QML works spanning kernel methods, reservoir computing, convolutional and recurrent architectures, generative models, and modern training paradigms. These reproductions are released as reusable, modular experiments that can be directly extended and adapted, establishing a shared experimental baseline consistent with empirical benchmarking methodologies widely adopted in modern artificial intelligence. By embedding photonic quantum models within established machine learning ecosystems, MerLin allows practitioners to leverage existing tooling for ablation studies, cross modality comparisons, and hybrid classical quantum workflows. The framework already implements hardware aware features, allowing tests on available quantum hardware while enabling exploration beyond its current capabilities, positioning MerLin as a future proof co design tool linking algorithms, benchmarks, and hardware.

💡 Research Summary

The paper introduces Merlin, an open‑source framework designed to serve as a discovery engine for photonic and hybrid quantum machine learning (QML). Recognizing that near‑term quantum advantage must be demonstrated through systematic, reproducible experimentation rather than isolated algorithmic proposals, the authors build Merlin around three pillars: (1) strong simulation of linear‑optical circuits, (2) seamless integration with mainstream machine‑learning ecosystems (PyTorch and scikit‑learn), and (3) a benchmarking‑first workflow that enables reproducibility of existing photonic QML results.

At the technical core is the Strong Linear‑Optical Simulation (SLOS) engine, originally developed for the Perceval library. SLOS computes the exact quantum state after an arbitrary interferometer by constructing a sparse transition graph that captures all valid photon‑number transitions across circuit layers. The graph is built once for a given input photon configuration; during each forward pass only the unitary‑dependent coefficients are recomputed. This yields a time complexity of O(n·C(m+n‑1,n)) and a memory footprint proportional to the full Fock‑space dimension, allowing exact simulations of up to roughly 20 photons on commodity GPUs.

Merlin wraps SLOS in a PyTorch‑compatible QuantumLayer, a subclass of torch.nn.Module. The layer exposes trainable parameters (phase shifters and beam‑splitter reflectivities), supports batching, and can be stacked with conventional neural‑network modules. Three orthogonal concepts are made explicit: (i) measurement strategy (full probability distribution, per‑mode photon‑number expectations, or raw amplitudes), (ii) computation space (full Fock, restricted subspaces, or qubit encodings such as dual‑rail), and (iii) detector model (photon‑number‑resolving, threshold, or custom noise models). This separation decouples circuit evolution, state representation, and read‑out semantics, enabling flexible hybrid architectures.

Data encoding, a known bottleneck for QML, is handled via two photonic‑native schemes. Angle encoding maps classical features onto phase shifts of selected modes, yielding a Fourier‑like model where the spectrum of accessible frequencies grows linearly with the number of input photons. The authors demonstrate that a three‑mode, three‑photon circuit can exactly fit a third‑order Fourier series and achieve high accuracy on the Moon binary‑classification benchmark. Amplitude encoding directly loads a normalized complex vector into the amplitudes of a quantum state, useful when the state originates from upstream photonic processing rather than raw classical data.

Hardware awareness is built into both simulation and execution. Users can configure loss, detector efficiency, and mode count to reflect realistic photonic processors. At runtime, the MerlinProcessor abstraction allows transparent connection to Quandela’s cloud‑accessible photonic QPUs, enabling a “simulation‑first, hardware‑later” workflow. Parameters learned in simulation can be uploaded to hardware for experimental validation, while larger‑scale designs can be explored purely in silico.

To validate the platform, the authors reproduced eighteen recent photonic and hybrid QML works covering kernel methods, reservoir computing, convolutional and recurrent networks, generative models, and modern training paradigms. Each reproduction uses the same Merlin experiment template, ensuring identical preprocessing, optimizer settings, and evaluation metrics. The reproduced results match the original papers within statistical variance, and the systematic framework makes it possible to isolate the contribution of data preprocessing, model engineering, and optimization from genuine quantum representational advantages.

In summary, Merlin fills a critical gap in the QML software landscape by unifying efficient exact simulation, automatic differentiation, hardware‑aware modeling, and reproducible benchmarking. It lowers the barrier for researchers to prototype photonic quantum models, conduct rigorous cross‑modal comparisons, and co‑design algorithms with emerging photonic hardware, thereby accelerating the path toward practical quantum advantage in machine learning.