Scientists are increasingly focused on identifying the practical benefits of quantum machine learning (QML) models, necessitating a shift from isolated algorithm development to systematic empirical investigation across diverse models, datasets, and hardware limitations. In response to this need, Cassandre Notton and Benjamin Stott of Quandela Quantique Inc., Philippe Schoeb of Université de Montréal DIRO & Mila, Anthony Walsh of Quandela, Grégoire Leboucher of ENS Paris-Saclay, Vincent Espitalier and Vassilis Apostolou of Quandela, and Louis-Félix Vigneux have partnered with Alexia Salavrakos. and Jean Senellart and colleagues introduce MerLin, an open-source framework designed as a detection engine for photonic and hybrid machine learning. This collaboration integrates optimized linear optical circuits into standard machine learning workflows, enables end-to-end differentiable training of layers, and establishes a shared experimental baseline through reproduction of 18 state-of-the-art QML works. By embedding photonic models within an established ecosystem, MerLin facilitates ablation studies, cross-modality comparisons, and hybrid classical workflows, positioning it as a future-proof tool for co-designing algorithms, benchmarks, and hardware.
This new platform integrates optimized simulations of linear optical circuits directly into standard machine learning workflows using PyTorch and scikit-learn, enabling fully differentiable training of quantum layers.
MerLin addresses the critical need for systematic benchmarking and reproducibility in the rapidly evolving field of quantum machine learning, moving beyond isolated algorithm proposals toward empirical exploration across diverse models, datasets, and hardware. The framework’s architecture allows researchers to leverage existing machine learning tools for detailed analysis, comparison, and creation of hybrid classical-quantum systems.
By embedding photonic quantum models within an established artificial intelligence ecosystem, MerLin facilitates ablation studies and cross-modality comparisons previously hampered by fragmented software environments. The research team initially reproduced 18 state-of-the-art photonic and hybrid quantum machine learning works, including kernel techniques, reservoir computing, convolutional and recurrent architectures, generative models, and modern training paradigms.
These replications will be released as reusable modular experiments to establish a shared experimental baseline consistent with empirical benchmarking methodologies widely adopted in modern artificial intelligence. This approach makes results verifiable and adaptable, fosters collaboration, and accelerates progress.
MerLin’s hardware-enabled capabilities enable testing on available quantum hardware while enabling exploration beyond current capabilities, positioning it as a future-proof collaborative design tool. This framework links algorithms, benchmarks, and hardware to address the challenge of translating theoretical advances into real-world applications.
This capability is particularly relevant for photonic quantum computing, which is a promising platform due to its scalability, robustness, and compatibility with existing optical communication technologies. MerLin is committed to unlocking the full potential of optical quantum machine learning and driving innovation in the field by providing an integrated platform for implementation, training, and evaluation.
Validating the MerLin framework, computational complexity and scalability limitations
Reproduction within the MerLin framework successfully validates 18 state-of-the-art photonic and hybrid quantum machine learning applications, establishing a consistent baseline for empirical benchmarks. These replications demonstrate the correctness of MerLin and provide a basis for future contributions to this field.
The core of this framework relies on powerful linear optical simulations, achieving a time complexity of O(n m+n−1 n). Here, “n” represents the number of photons and “m” represents the number of modes. This simulation method efficiently computes quantum states, avoids redundant computations of matrix eternity, and provides significant speedups for practical applications.
The implementation of MerLin requires a memory footprint of O(n m+n−1 n), and current standard hardware limits actual simulations to approximately 20 photons. However, this limitation positions the framework as a valuable co-design tool to link algorithms, benchmarks, and hardware capabilities for future exploration.
MerLin facilitates ablation studies, cross-modality comparisons, and hybrid-classical workflows by embedding photonic models within established machine learning ecosystems such as PyTorch and scikit-learn. An independent study has confirmed that photonic contributions currently account for about 6% of all quantum machine learning publications, a figure that is consistent with the research team’s findings.
Code availability remains a challenge in the broader QML environment, hovering around 27% for gate-based approaches and 43% for photonic QML papers, highlighting the importance of MerLin’s emphasis on reproducibility. The framework’s design prioritizes not only identifying performance improvements, but also understanding their origins and disentangling improvements that result from data preprocessing, model engineering, or entirely new computational mechanisms.
Reproducing established quantum machine learning with differentiable photonic simulations
MerLin, an open source framework, powers this effort as a detection engine for photonic and hybrid quantum machine learning. MerLin’s core integrates powerful, optimized simulations of linear optical circuits directly into established machine learning workflows powered by PyTorch and scikit-learn. This integration facilitates end-to-end differentiable training of quantum layers, an important step toward seamless quantum-classical model hybrid development.
By embedding photonic models within these standard ecosystems, researchers can easily apply existing tools for detailed ablation studies, cross-modality comparisons, and creation of complex hybrid classical-quantum workflows. In this study, we systematically reproduced 18 state-of-the-art photonic and hybrid QML works covering diverse architectures including kernel methods, reservoir computing, convolutional and recurrent networks, generative models, and modern training paradigms.
These replications are intentionally designed as reusable, modular experiments to establish a shared experimental baseline that aligns with empirical benchmark methodologies common in modern artificial intelligence. This modularity enables direct extension and adaptation of existing experiments, fostering collaboration and accelerating the pace of discovery.
MerLin’s design prioritizes systematic benchmarking and reproducibility, addressing critical needs in the rapidly evolving field of quantum machine learning. The framework implements hardware-enabled features that enable testing on currently available quantum hardware while facilitating the exploration of algorithms beyond the limitations of current hardware.
This forward-looking approach positions MerLin as a co-design tool, linking algorithms, benchmarks, and hardware in a consistent and future-proof way. The choice of linear optical circuits is based on their potential scalability, robustness, and compatibility with existing optical communication technologies, offering a promising path toward practical quantum computing.
big picture
The relentless pursuit of quantum machine learning has often felt like chasing a mirage. Demonstration of potential speedups has often been overshadowed by the sheer difficulty of building practical, scalable systems. This new framework, MerLin, represents a major shift by focusing on providing tools to systematically explore a wide range of possibilities, rather than a single quantum algorithm.
This is a move away from isolated proofs of concept and toward a more engineering-driven approach, recognizing that the path to useful quantum computation is likely to be paved by hybrid classical-quantum solutions. MerLin’s strength lies in its integration with existing machine learning ecosystems. Embedding photonic models within established workflows such as PyTorch and scikit-learn allows researchers to leverage familiar tools for important tasks such as ablation studies and cross-modality comparisons.
This lowers the barrier to entry, enables the broader community to contribute to the field, and accelerates the process of identifying truly promising applications. Reproducing the 18 existing photonic and hybrid QML works is not a trivial task, and establishing a shared experimental baseline is essential for meaningful benchmarking.
However, the framework’s current focus on photonic systems has limitations. While photonics has advantages in room temperature operation and connectivity, it is not the only viable quantum hardware platform. Additionally, real tests demonstrate performance improvements on real datasets and problems, not just carefully curated benchmarks.
Next steps will require tackling more complex and messy data and evaluating the framework’s ability to handle the noise and imperfections inherent in current quantum hardware. Ultimately, MerLin’s success will depend on its ability to bridge the gap between theoretical possibility and practicality and guide the field toward a future where quantum and classical computation work in tandem.
