Photonic Information Processing
Bueno Moragues, Julián (Supervisors: Fischer, Ingo; Brunner, Daniel)
PhD Thesis (2019)
In the current thesis we experimentally study the dynamics of complex photonic systems including semiconductor laser systems with delayed feedback and spatially coupled optical systems, and employ their dynamics for information processing. Photonic delay systems offer a broad range of complex dynamics, making them excellent testbeds for the study of nonlinear dynamics. We take advantage of the complex dynamics and turn the photonic delay systems into information processing systems by applying techniques from the ﬁeld of neural networks, studying their properties, and demonstrating successful performance.
First we study the dynamics of a semiconductor laser with two delayed optical feedbacks. The temporal evolution of the intensity from the laser is represented and studied in a two dimensional pseudo-space, which conveniently allows to visualize structures in the dynamics. Two dynamical regimes are being addressed in this study, Spiral Phase Defects and Defect-mediated Turbulence. Spiral Phase Defects that were predicted for this system have not been observed, and we discuss the possible reasons. Defect-mediated Turbulence is observed in a broad range of parameters. The defect-mediated dynamics is analyzed in detail by using intensity distribution and spectral analysis. The spectral analysis shows an exponential decay of the spectra for low bias currents of the semiconductor laser. For increasing bias currents, the spectra exhibit a double power law decay that merges into a single power law as the bias currents is increased. This indicates a scale-free behavior in the dynamics for high bias currents.
Next, we study the use of a semiconductor laser with a single delay optical feedback and optical injection as a delay-based reservoir computer. We study the fundamental properties of the system, speciﬁcally the properties of injection locking, consistency, and memory. We evaluate the performance of the system in a prediction task, and link the different properties to performance. The impact of crucial experimental parameters on the properties are evaluated, and the implications discussed. Moreover, we explore further functionalities. Suitable properties and good performance are demonstrated for injection at modulation rates up to to 20 GSa/s, and also for optical injection detuned hundreds of GHz from the solitary emission of the semiconductor laser. Furthermore, we demonstrate how to create a reservoir with nodes with different sets of properties, and show that performance in a prediction task can be improved with this approach.
In the ﬁnal part, we design and build a spatially extended reservoir computer with 900 coupled nonlinear nodes. The reservoir is implemented with a Spatial Light Modulator (SLM), and the nonlinearity is realized by taking advantage of the polarization modulation of the Spatial Light Modulator (SLM) together with a Polarization Beam Splitter (PBS). The coupling in the network is experimentally implemented with a Diffractive Optical Element (DOE), resulting in coupling beyond ﬁrst neighbors. The input layer is implemented digitally employing the control computer. The output layer is experimentally implemented with a Digital Micro-mirror Display (DMD) and a lens, where the conﬁguration of the Digital Micro-mirror Display (DMD) determines the values of the output weights in the reservoir computer. We evaluate the dynamics of the nodes and demonstrate suitable nonlinear dynamics in the network. Learning rules are implemented to allow the system to autonomously change the output weights based on previous performance, and to search for an optimized conﬁguration attaining low errors in speciﬁc tasks. We demonstrate successful operation of the reservoir computer by showing that, despite some limitations, it is capable of learning reducing the error in a prediction task and ﬁnally exhibiting very low prediction errors.