merlin-photonic-qml

name	merlin-photonic-qml
description	MerLin discovery engine for photonic and hybrid quantum machine learning. Embeds linear optical circuit simulation into PyTorch/scikit-learn for end-to-end differentiable training of quantum layers. Use when: (1) building hybrid quantum-classical ML models, (2) reproducing photonic QML benchmarks, (3) designing quantum layer architectures, (4) benchmarking QML against classical baselines, (5) hardware-aware quantum ML testing. Activation: merlin, photonic qml, hybrid quantum machine learning, quantum benchmarking, photonic simulation, quantum layer, end-to-end quantum training.
metadata	{"arxiv_id":"2602.11092","published":"2026-02-11","authors":"Cassandre Notton, Benjamin Stott, Philippe Schoeb","tags":["quantum","machine-learning","photonic","benchmarking","pytorch","hybrid"]}

MerLin: Photonic & Hybrid QML Discovery Engine

Core Concept

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 within established ML ecosystems.

Architecture Patterns

1. Quantum Layer in PyTorch Pipeline

Classical Preprocessing → Quantum Layer (linear optical) → Classical Postprocessing → Loss

• Quantum layers are differentiable via strong simulation
• Gradients flow through quantum circuits back to classical layers
• Use PennyLane-compatible interfaces for circuit definition

2. Hybrid Transfer Learning

• Pre-train classical feature extractor (ResNet, ViT, etc.)
• Freeze early layers, replace final layer with quantum circuit
• Fine-tune only quantum layer parameters + final classical layer
• Benefits from classical pretraining + quantum expressivity

3. Photonic QML Reproduction

MerLin reproduces 18 SOTA photonic/hybrid QML works across:

• Kernel methods
• Reservoir computing
• Convolutional and recurrent architectures
• Generative models
• Modern training paradigms

Methodology

Step 1: Define Quantum Circuit

• Use parameterized linear optical circuits
• Define encoding scheme (amplitude, phase, etc.)
• Specify measurement operators

Step 2: Embed in ML Pipeline

• Wrap quantum circuit as differentiable module
• Connect to PyTorch autograd
• Enable gradient flow through simulation

Step 3: Train End-to-End

• Standard PyTorch training loop
• Adam/SGD optimizer
• Mini-batch compatible

Step 4: Benchmark

• Compare against classical baselines
• Test on same datasets
• Report parameter efficiency, accuracy, training time

Hardware-Aware Features

• Simulate beyond current hardware capabilities
• Test on available quantum hardware
• Co-design: algorithms ↔ benchmarks ↔ hardware

Pitfalls

• Simulation cost: Strong simulation scales exponentially with qubit count; limit to ~20-25 qubits
• Gradient noise: Quantum layer gradients can be noisy; use larger batch sizes
• Barren plateaus: Deep quantum circuits suffer from vanishing gradients; keep circuits shallow
• Encoding bottleneck: Poor data encoding limits quantum advantage; experiment with multiple encoding schemes

References

• arXiv: 2602.11092v2
• Framework: MerLin (open-source)
• Reproduced works: 18 SOTA photonic/hybrid QML papers