Deep-Learning-Based VCSEL Transmitter Emulator

The design of photonic/opto-electronic structures is a crucial field of research due to the wide range of applications in which these devices are utilized, such as communications, sensing and imaging. The traditional approach to design, which involves numerical simulation and iterative adjustments, is both time-consuming and computationally intensive. The ability to design-emulate photonic structures quickly and efficiently could bring new technologies to market at a faster pace and can also facilitate the process of extracting critical performance metrics from complex fabricated devices. The Vertical Cavity Surface Emitting Laser (VCSEL) and its use as a transmitter broadly used in data center interconnects is an example of a rather complex nonlinear system also exhibiting bandwidth limitation mechanisms attributed to its frequency response and parasitic effects, and thus its numerical simulation based on conventional rate equations can only partially approximate its behavior. When it comes to modelling fabricated devices, such physical models are notoriously difficult to parameterize in order to fit to observed laboratory data; in contrast, the emulation of the experimentally recorded behavior with the use of nonlinear models based on neural networks provides a much more efficient path to modelling, and with potentially higher prediction accuracy [1]. In this work, two types of bidirectional recurrent neural networks (RNN), namely Long Short-Term Memory (bi-LSTM) and Vanilla (bi-VRNN) were used to simulate the dynamic behavior of an experimentally characterized VCSEL-based, PAM-4 transmitter, exhibiting a high prediction accuracy approaching 100%.