MerLin - Photonic Quantum Machine Learning Framework
MerLin - Photonic Quantum Machine Learning in PyTorch
MerLin enables researchers to build, train, and benchmark photonic quantum machine learning models using familiar ML tools.
By integrating photonic circuit simulation with PyTorch and scikit-learn, MerLin allows hybrid quantum-classical models to be developed and evaluated within standard machine learning workflows.
Explore MerLin
Quickstart
Start with the essentials and build your first photonic quantum layer.
Latest Releases
Track recent updates, fixes, and newly added capabilities in MerLin.
Benchmark State-Of-The-Art Papers
Benchmark new QML ideas against 21 papers reproduced in one unified MerLin framework.
What you can do with MerLin
Start simple, then scale: MerLin lets you prototype your first quantum layer and progress to reproducible, benchmark-ready QML experiments.
With MerLin you can:
⚡ Build photonic quantum layers directly in PyTorch.
🔬 Benchmark quantum models on real machine learning datasets.
🔁 Reproduce state-of-the-art QML papers.
🧠 Combine classical neural networks with quantum circuits.
🧪 Experiment with kernels, reservoirs, QNNs, and generative models.
🖥 Run simulations locally or execute circuits on photonic hardware.
MerLin is designed as a discovery engine for photonic and hybrid quantum machine learning, with reproducible experimentation across models, datasets, and hardware constraints.
Reproducing State-of-the-art QML Papers
A core goal of MerLin is to make quantum machine learning research more reproducible, comparable, and hardware-aware. As the field grows, ensuring that results can be independently reproduced and evaluated across consistent benchmarks becomes increasingly important. While many QML papers provide valuable algorithmic insights, reproducing published results can still be challenging due to differences in hardware and software stacks, data preprocessing pipelines, experimental settings, or hardware assumptions. MerLin addresses this by providing a unified framework and a growing catalog of reproduced QML experiments, implemented within a consistent environment and designed to reflect realistic quantum hardware constraints. These reproductions help:
Validate MerLin implementations against published results.
Provide reusable baselines and reference implementations for new research.
Enable systematic benchmarking across models, datasets, and encodings.
Study how algorithms interact with hardware constraints and photonic architectures, supporting algorithm–hardware co-design.
Our Reproduced Benchmarks
Explore Reproduced Papers
Browse our catalog of 21 papers reproduced with reusable QML baselines.
Highlight: QORC
Quantum optical reservoir computing powered by boson sampling.
Highlight: QSSL
Quantum self-supervised learning with photonic and gate-model quantum backends.
Want to Contribute?
Submit examples, reproductions, or documentation improvements.
Who is MerLin For?
Choose your path based on your background and goals:
- Machine Learning Engineers
New to quantum? Start with the Quickstart: Classify a Nonlinear Dataset with the Photonic Quantum Layer for a hands-on introduction. No quantum background required—MerLin abstracts the complexity so you focus on building and training models. See Examples for common patterns.
- Quantum Researchers
Explore Quantum Expert Area for detailed documentation on photonic circuits, gate operations, and hardware integration. MerLin provides flexible APIs for designing custom quantum layers and debugging quantum behavior.
- Paper Reproducers
Start with Reproduced Papers to access reference implementations of published QML papers. Compare your results, adapt methods, and build on proven baselines. See Performance for benchmarking and hardware constraints.
- Curious Explorers
Browse notebooks/index for practical tutorials, or dive into User Guide for comprehensive guidance on workflows and patterns.
Installation
pip install merlinquantum
This installs MerLin and core dependencies including PyTorch, Perceval, NumPy, Pandas, and scikit-learn.
Verify the installation:
import merlin as ML
print(ML.__version__)
Minimal Example
Train a photonic quantum layer inside a PyTorch workflow:
import torch
import numpy as np
import merlin as ML
from sklearn.datasets import make_circles
from sklearn.model_selection import train_test_split
# Dataset
X, y = make_circles(n_samples=400)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)
# Normalize using training data statistics (prevent data leakage)
min_vals = X_train.min(axis=0, keepdims=True)
max_vals = X_train.max(axis=0, keepdims=True)
X_train = (X_train - min_vals) / np.clip(max_vals - min_vals, a_min=1e-6, a_max=None)
X_test = (X_test - min_vals) / np.clip(max_vals - min_vals, a_min=1e-6, a_max=None)
X_train = torch.tensor(X_train, dtype=torch.float32)
y_train = torch.tensor(y_train, dtype=torch.long)
# Quantum layer
quantum_layer = ML.QuantumLayer.simple(
input_size=2,
output_size=2,
)
# Training loop
optimizer = torch.optim.Adam(quantum_layer.parameters(), lr=0.01)
criterion = torch.nn.CrossEntropyLoss()
for _ in range(100):
optimizer.zero_grad()
logits = quantum_layer(X_train)
loss = criterion(logits, y_train)
loss.backward()
optimizer.step()
This example:
Builds a photonic circuit.
Wraps it as a PyTorch module.
Trains it with standard ML tooling.
Kickstart Your Journey
The MerLin Model
In MerLin, a quantum circuit behaves like a regular PyTorch layer, so you can plug it into familiar pipelines and train end-to-end with gradient descent.
model = torch.nn.Sequential(
torch.nn.Linear(10, 2),
ML.QuantumLayer.simple(input_size=2),
torch.nn.Linear(2, 2),
)
This hybrid architecture can combine:
Classical preprocessing: Encode raw features into quantum-friendly dimensions
Quantum processing: Let the photonic circuit learn quantum-specific patterns
Classical output: Map quantum results to predictions
Common patterns include quantum kernels, variational quantum classifiers (VQC), and quantum feature maps. Explore Examples for working implementations and use cases.
Next Steps
Get started: Quickstart: Classify a Nonlinear Dataset with the Photonic Quantum Layer
Understand MerLin: User Guide
Explore examples: :doc:`examples/index
Reproduce papers: Reproduced Papers
Contribute: Contribute to MerLin and Reproduced Papers
API reference: API Reference
How to Cite Us
@article{notton2026merlin,
title={MerLin: A Discovery Engine for Photonic and Hybrid Quantum Machine Learning},
author={Notton, Cassandre and Stott, Benjamin and Schoeb, Philippe and Walsh, Anthony and Leboucher, Gr{\'e}goire and Espitalier, Vincent and Apostolou, Vassilis and Vigneux, Louis-F{\'e}lix and Salavrakos, Alexia and Senellart, Jean},
journal={arXiv preprint arXiv:2602.11092},
year={2026}
}