We present an optimized Mach–Zehnder Interferometer (MZI) with phase change materials for photonic neural networks (PNNs). With 0.2 dB loss, −38 dB crosstalk, and length of 52 µm, the designed MZI significantly improves the scalability and accuracy of PNNs under loss and crosstalk.
Frequency-modulated (FM) laser combs, which offer a quasi-continuous-wave output and a flat-topped optical spectrum, are emerging as a promising solution for wavelength-division multiplexing applications, precision metrology, and ultrafast optical ranging. The generation of FM combs relies on spatial hole burning, group velocity dispersion, Kerr nonlinearity, and four-wave mixing (FWM). While FM combs have been widely observed in quantum cascade Fabry-Perot (FP) lasers, the requirement for a low-dispersion FP cavity can be a challenge in platforms where the waveguide dispersion is mainly determined by the material. Here we report a 60 GHz quantum-dot (QD) mode-locked laser in which both the amplitude-modulated (AM) and the FM comb can be generated independently. The high FWM efficiency of -5 dB allows the QD laser to generate FM comb efficiently. We also demonstrate that the Kerr nonlinearity can be practically engineered to improve the FM comb bandwidth without the need for GVD engineering. The maximum 3-dB bandwidth that our QD platform can deliver is as large as 2.2 THz. This study gives novel insights into the improvement of FM combs and paves the way for small-footprint, electrically pumped, and energy-efficient frequency combs for silicon photonic integrated circuits (PICs).
We report a high-speed (GHz rate) and hardware-friendly Neuromorphic Photonic Spiking Neural Network (pSNN) architecture formed by up to >1000 spiking nodes and built with a single off-the-shelf telecom Vertical-Cavity Surface Emitting Laser (VCSEL). We demonstrate the successful application of the proposed pSNN architecture to a complex nonlinear classification task, running at sub-nanosecond speeds per spike directly in the optical domain.
We present a statistical analysis of arrayed waveguide gratings in the presence of phase errors in the waveguides caused by fabrication process variations and characterized by the coherence length. The proposed method allows to predict yield and to optimize complex photonic integrated circuits.
Optimal hyperparameters and inference accuracy for analog-based deep learning hardware is highly dependent on system architecture and component noise. We present AnalogVNN as a fully modular framework to easily model and train arbitrary analog photonic neural networks using the simple modular layer structures of PyTorch.
