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

1. Problem Statement

The field of Quantum Machine Learning (QML), particularly in the photonic domain, faces significant fragmentation and a lack of systematic benchmarking.

• Software Silos: Existing frameworks are often specialized for specific hardware paradigms (e.g., superconducting qubits vs. neutral atoms) or specific simulation types (gate-based vs. continuous-variable). Photonic QML lacks a unified framework that combines efficient simulation, integration with standard Machine Learning (ML) workflows, noise modeling, and hardware access.
• Reproducibility Crisis: A large portion of QML literature relies on isolated algorithmic proposals with low code availability (approx. 43% for photonic papers). There is a lack of rigorous, multi-dimensional benchmarking, with many studies relying on single-run metrics rather than statistical validation.
• Hardware-Algorithm Disconnect: There is a need for "co-design" tools that allow researchers to explore how algorithmic choices interact with the specific constraints and capabilities of photonic hardware (e.g., photon loss, detector types) before deploying on physical quantum processing units (QPUs).

2. Methodology: The MerLin Framework

The authors introduce MerLin, an open-source discovery engine designed to bridge the gap between photonic quantum simulation and standard ML ecosystems (PyTorch and scikit-learn).

Core Architecture

• Differentiable Strong Simulation: MerLin integrates SLOS (Strong Linear Optical Simulation), an optimized tool for simulating linear-optical circuits. Unlike methods that compute matrix permanents for every output configuration, SLOS reuses intermediate results to compute the exact quantum state with a time complexity of $O(n \binom{m+n-1}{n})$.
• PyTorch Integration: MerLin exposes SLOS as differentiable primitives within PyTorch. It uses a precomputed sparse computation graph to accelerate training. The graph structure (determined by circuit topology and input state) is built once, while only the unitary-dependent coefficients are recomputed during forward passes, enabling efficient gradient-based optimization via automatic differentiation.
• QuantumLayer Abstraction: The framework introduces a QuantumLayer (a torch.nn.Module) that decouples three core concepts:
  1. Measurement Strategy: Defines outputs (full probability distribution, photon number expectations, or complex amplitudes).
  2. Computation Space: Supports full Fock space, restricted spaces, or encoded spaces (e.g., dual-rail, QLOQ).
  3. Data Encoding: Supports Angle Encoding (mapping features to phase shifts, naturally realizing Fourier-like models) and Amplitude Encoding (direct initialization of state amplitudes).
• Hardware Awareness:
  • Detector Models: Users can simulate realistic hardware behaviors, such as Photon-Number-Resolving (PNR) or threshold detectors.
  • MerlinProcessor: An abstraction layer that connects quantum layers to cloud-accessible photonic hardware (e.g., Quandela's Belenos QPU), handling latency, shot-based sampling, and batching.
  • Hybrid Workflows: Includes a QuantumBridge to enable interoperability between gate-based qubit circuits and photonic circuits (e.g., mapping qubits to photons).

3. Key Contributions

• Unified Discovery Engine: MerLin provides the first framework to combine optimized discrete-variable (DV) photonic simulation with end-to-end differentiable training in PyTorch and scikit-learn.
• Systematic Reproduction of State-of-the-Art: The authors reproduced 18 state-of-the-art photonic and hybrid QML works covering diverse architectures:
  • Kernel methods (Fidelity kernels).
  • Recurrent architectures (QLSTM, QRNN).
  • Convolutional networks (Photonic QCNNs).
  • Generative models (QGANs).
  • Training paradigms (Self-supervised learning, Knowledge Distillation).
• Benchmarking Repository: The authors released a companion repository containing these reproductions as modular, reusable experiments, establishing a shared baseline for empirical benchmarking in QML.
• Hardware-Algorithm Co-Design: The framework allows for "forward-looking" exploration, enabling researchers to test algorithms on simulated hardware constraints and real QPUs, informing future processor design.

4. Results and Findings

The reproduction efforts yielded several critical insights:

• Performance Validation: MerLin successfully reproduced the results of the 18 selected papers, validating its correctness. In many cases, porting from previous frameworks (like Perceval) to MerLin resulted in order-of-magnitude speedups in simulation time due to optimized sparse graph execution.
• Expressivity and Photons: Reproduction of variational circuits confirmed that increasing the number of photons linearly increases the expressivity (frequency spectrum) of the model without requiring deeper circuits.
• Accuracy Improvements: Through systematic hyperparameter optimization, the MerLin implementation of Photonic QCNNs improved reported binary classification accuracies from ~92.7% to 98.2% on the Custom BAS dataset and from ~93.1% to 98.8% on MNIST (0 vs 1).
• Hybrid vs. Classical:
  • QCNNs: While quantum models showed advantages in parameter efficiency, classical models remained more parameter-efficient overall, though this metric is less relevant for photonic hardware where parameters are bounded by physical components.
  • LLM Fine-Tuning: Reproduction of a Quantum Large-Language Model fine-tuning study found no clear separation between quantum and classical models (both reached ~89% accuracy), suggesting the original task might have been too simple to demonstrate quantum utility.
  • Adversarial Robustness: The study found that amplitude encoding is highly vulnerable to small perturbations, whereas angle encoding is substantially more robust.
• Hardware Execution: The framework successfully executed reservoir computing experiments on the Belenos QPU (a real photonic processor), demonstrating the viability of the hardware-aware workflow.

5. Significance

• Standardization: MerLin addresses the fragmentation in QML software by providing a "PyTorch-native" standard for photonic models, lowering the barrier to entry for classical ML practitioners.
• Empirical Rigor: By shifting the focus from isolated proposals to systematic, reproducible benchmarking, MerLin helps the community distinguish between genuine quantum advantages and artifacts of data preprocessing or optimization strategies.
• Co-Design Tool: The framework acts as a critical feedback loop between software algorithms and hardware engineering. Insights gained from large-scale simulation can directly inform the design of future photonic quantum processors.
• Future-Proofing: By supporting hybrid workflows and emerging paradigms (like photonic memristors), MerLin positions itself as a forward-looking tool for the near-term era of quantum computing, where hybrid classical-quantual approaches are most viable.

In conclusion, MerLin is not just a simulator but a comprehensive ecosystem for the discovery, validation, and deployment of photonic quantum machine learning, bridging the gap between theoretical proposals and practical hardware implementation.