Light-based technologies represent an essential component in countless applications in our everyday life. In the last few decades, optical technologies have provided a key platform also for several areas of research, from fundamental tests of quantum mechanics to quantum simulation and communication. More recently, optical circuits of modest size have been applied to various tasks, for instance to provide evidence of a quantum computational advantage. The potential inherent to these technologies is rooted in the properties of the individual photons, the elementary particles of light, such as their mobility, speed, high bandwidth and ease of manipulation. At the same time, machine learning (ML) has established itself as a powerful approach to enable problem solving. Motivated by the outstanding success in their respective domains, first steps have been made to bring ML and photonics together. Results, here, fall into two main categories. On the one hand, (i) photonic devices serve as efficient platforms to implement ML: proof-of-principle demonstrations include optical neural networks and neuromorphic processors, which promise higher performance than conventional architectures. On the other hand, (ii) ML can be used to gain insights into single-photon quantum processes. Also, ML offers a promising toolbox to optimize modern-day photonic devices and achieve even stronger demonstrations. In this case, major obstacles are generally represented by imperfect single-photon sources and detectors, as well as by an imperfect control over the parameters (phases) that govern the dynamics in such circuits. While practical solutions can be engineered for the two former issues, the latter represents a real challenge when compact and dense integration of several optical components is desired. Specifically, the adoption of densely integrated circuits with several components requires hardware and software solutions to mitigate the effects of crosstalk noise and biased errors in the phase settings. An effective solution to this problem is necessary to ensure that future protocols and infrastructures – believed to offer great benefits with respect to nowadays classical technologies – can be implemented on hardware affected by some unavoidable level of noise.
The research project MAZINGER took up this challenge by bringing together analytical and numerical tools (both from standard optimization techniques and ML), in order to enhance state-of-the-art optical applications. To this end, MAZINGER explored well-established optimization algorithms to cope with changing, noisy environments and non-ideal reconfigurable components, as well as novel frameworks to enable scientific discovery with optical circuits. The project, carried out in one of the leading groups in theoretical quantum ML, also involves a collaboration with a leading experimental group in photonics, with the goal of testing any findings on an actual, high-precision quantum experiment. The employed techniques have been developed within the general framework of single- and multi-photon interference, to ensure that any results are readily transferable to related lines of research.
