Photonic integrated circuits (PICs) are considered as the way to make photonic systems or subsystems cheap and ubiquitous. PICs still are several orders of magnitude more expensive than their microelectronic counterparts, which has restricted their application to a few niche markets.

Integrated photonic neural networks provide a promising platform for energy-efficient, high-throughput machine learning with extensive scientific and commercial applications. Photonic neural networks efficiently transform optically encoded inputs using Mach-Zehnder interferometer mesh networks interleaved with nonlinearities. We experimentally trained a three-layer, four-port silicon photonic neural network with programmable phase shifters and optical power monitoring to solve classification tasks using “in situ backpropagation,” a photonic analog of the most popular method to train conventional neural networks. We measured backpropagated gradients for phase-shifter voltages by interfering forward- and backward-propagating light and simulated in situ backpropagation for 64-port photonic neural networks trained on MNIST image recognition given errors. All experiments performed comparably to digital simulations ( > 94% test accuracy), and energy scaling analysis indicated a route to scalable machine learning.

Propagation of light beams through scattering or multimode systems may lead to the randomization of the spatial coherence of the light. Although information is not lost, its recovery requires a coherent interferometric reconstruction of the original signals, which have been scrambled into the modes of the scattering system. Here we show that we can automatically unscramble optical beams that have been arbitrarily mixed in a multimode waveguide, undoing the scattering and mixing between the spatial modes through a mesh of silicon photonics tuneable beam splitters. Transparent light detectors integrated in a photonic chip are used to directly monitor the evolution of each mode along the mesh, allowing sequential tuning and adaptive individual feedback control of each beam splitter. The entire mesh self-configures automatically through a progressive tuning algorithm and resets itself after significantly perturbing the mixing, without turning off the beams. We demonstrate information recovery by the simultaneous unscrambling, sorting and tracking of four mixed modes, with residual cross-talk of −20 dB between the beams. Circuit partitioning assisted by transparent detectors enables scalability to meshes with a higher port count and to a higher number of modes without a proportionate increase in the control complexity. The principle of self-configuring and self-resetting in optical systems should be applicable in a wide range of optical applications.

Slow light devices such as photonic crystal waveguides (PhCW) and coupled resonator optical waveguides (CROW) have much promise for optical signal processing applications and a number of successful demonstrations underpinning this promise have already been made. Most of these applications are limited by propagation losses, especially for higher group indices. These losses are caused by technological imperfections (“extrinsic loss”) that cause scattering of light from the waveguide mode. The relationship between this loss and the group velocity is complex and until now has not been fully understood. Here, we present a comprehensive explanation of the extrinsic loss mechanisms in PhC waveguides and address some misconceptions surrounding loss and slow light that have arisen in recent years. We develop a theoretical model that accurately describes the loss spectra of PhC waveguides. One of the key insights of the model is that the entire hole contributes coherently to the scattering process, in contrast to previous models that added up the scattering from short sections incoherently. As a result, we have already realised waveguides with significantly lower losses than comparable photonic crystal waveguides as well as achieving propagation losses, in units of loss per unit time (dB/ns) that are even lower than those of state-of-the-art coupled resonator optical waveguides based on silicon photonic wires. The model will enable more advanced designs with further loss reduction within existing technological constraints.

In this paper, we report a direct comparison between coupled resonator optical waveguides (CROWs) and photonic crystal waveguides (PhCWs), which have both been exploited as tunable delay lines. The two structures were fabricated on the same silicon-on-insulator (SOI) technological platform, with the same fabrication facilities and evaluated under the same signal bit-rate conditions. We compare the frequency- and time-domain response of the two structures; the physical mechanism underlying the tuning of the delay; the main limits induced by loss, dispersion, and structural disorder; and the impact of CROW and PhCW tunable delay lines on the transmission of data stream intensity and phase modulated up to 100 Gb/s. The main result of this study is that, in the considered domain of applications, CROWs and PhCWs behave much more similarly than one would expect. At data rates around 100 Gb/s, CROWs and PhCWs can be placed in competition. Lower data rates, where longer absolute delays are required and propagation loss becomes a critical issue, are the preferred domain of CROWs fabricated with large ring resonators, while at data rates in the terabit range, PhCWs remain the leading technology.