MerLin 0.3 is an open-source framework developed by Quandela, designed to systematically explore photonic and hybrid quantum machine learning (QML). Built upon the Perceval SDK, it leverages the Strong Linear Optics Simulation (SLOS) to perform precise quantum state computations within a native PyTorch environment. Its core architecture is the QuantumLayer, a torch.nn.Module that supports end-to-end differentiable training of linear optical circuits. By precomputing sparse photon number transition graphs, the MerLin framework can accelerate gradient-based optimization of circuit parameters, such as adjustments to phase shifters and beam splitters, directly within standard classical AI workflows. 
The MerLin framework supports various data encoding methods, including angle encoding for Fourier-like feature mappings and amplitude encoding for state vector initialization. The QuantumBridge abstraction enables cross-paradigm architecture comparisons by mapping qubit-based gates to photonic dual-rail or QLOQ encodings. MerLin is also designed for hardware-aware execution via the MerlinProcessor interface, facilitating the offloading of hybrid model components to physical Quantum Processing Units (QPUs), such as Quandela's Belenos system. Furthermore, it integrates noise models and detector-specific semantics, including photon-number-resolving detectors and threshold detectors, allowing researchers to simulate hardware constraints during the training phase.
To address reproducibility challenges in quantum machine learning, the MerLin framework includes a library of 18 reproduced top-tier papers, covering quantum kernels, reservoir computing, and convolutional architectures. These modular experiments provide standardized benchmarks for comparing photonic and gate-based models under unified conditions. Technical insights gained from these reproductions suggest that the expressive power of photonic variational quantum circuits (VQCs) scales linearly with the number of input photons, without requiring increased circuit depth. This empirical approach aims to shift QML from isolated demonstrations towards a disciplined engineering framework for assessing quantum utility.
