Research On Photonic Neuromorphic Devices And Systems Based On Semiconductor Lasers And Silicon-Based MZI

Photonic neuromorphic computing is an innovative computing approach that draws inspiration from the workings of the biological nervous system.It fully combines photonics and neuromorphic computing.It employs photonics principles and optical devices for information processing and computing,offering significant advantages such as high speed,large bandwidth,low power consumption,and strong scalability.It aims to break the power and storage bottlenecks of traditional von Neumann computing architectures and explore new frontiers in computational science,offering exciting possibilities and opportunities for future research and applications in computational science.Photonic neuromorphic devices and photonic neuromorphic computing systems have been a research hotspot in recent years and are currently still in an early exploratory stage,full of both challenges and opportunities.This dissertation focuses on international cutting-edge hotspots and researches the photonic neuromorphic devices and systems supported by the National Key Research and Development Program and the National Natural Science Foundation.Combining theory and experiment,neuron-like properties of two-stage semiconductor lasers are studied,including the vertical-cavity surface-emitting laser with a saturable absorber(VCSEL-SA)and the Fabry–Pérot edge-emitting laser with a saturable absorber(FP-SA).Synaptic-like properties based on the vertical-cavity semiconductor optical amplifier(VCSOA)and the silicon photonic Mach-Zehnder interferometer(MZI)are also explored.Based on these two critical neuromorphic devices,photonic spiking neural networks(SNNs)are established to achieve various intelligent neuromorphic computing tasks.The research results are of great significance for promoting the development of photonic neuromorphic computing.The research content and innovations present in this paper are as outlined below:1.Photonic spiking neurons are key neuromorphic devices of the photonic neuromorphic computing system.Based on the excitability of two-stage semiconductor lasers,the twostage laser rate equation for optically injected VCSEL-SA was established by the Yamada model of two-stage semiconductor lasers.And neuron-like responses,including excitation thresholds,temporal integration,and refractory periods were realized in simulation.Based on the FP-SA spiking neuron chip,repeatedly producing neuron-like responses under the external optical injection were demonstrated.Photonic spiking neurons based on two-stage semiconductor lasers possess characteristics of rapid response and low power consumption,laying the device foundation for photonic neuromorphic computing.2.For the first time,the photonic synaptic plasticity based on a single VCSOA was experimentally confirmed.Based on the cross-gain modulation property of the VCSOA,two experiments including the injection of both single-polarization light and dual-polarization light into a single VCSOA were designed and demonstrated to achieve optical synaptic plasticity.They can generate a few nanoseconds spike-timing-dependent plasticity(STDP)learning window at lower power consumption(a few milliamps)and ultra-fast speed(on the order of nanoseconds).Considering the transmission properties and reconfigurability of MZIs,the cascaded MZI network was decomposed using the iterative decomposition method,and analytical solutions for all phase shifts were obtained by simulation calculation.Based on the gradient descent algorithm,the 4×4 MZI photonic synapse chip was experimentally tested.Its transmission matrix can simulate the target matrix and complete the linear computation function of synapses.The study of photonic synapses and plasticity provides rich computational capabilities and application potential for photonic neuromorphic computing systems.3.The fully VCSEL-SA-based photonic SNNs were proposed to implement sound azimuth detection and spike sequence learning tasks by simulation.A 2×2 photonic SNN was constructed to perform sound azimuth detection.The time difference between the spike firings of two post-synaptic neurons was used as an indicator of sound azimuth.Compared to biological sound localization systems,this network exhibits considerably lower energy consumption and higher resolution(on the order of nanoseconds).Subsequently,a multiinput single-output supervised photonic SNN was built to achieve spike sequence learning by exploiting the temporal information of spikes and combining with the improved Remote Supervised Method(Re Su Me).This network can learn and accurately reproduce any desired target spike sequence.This simulation work lays a solid model foundation for the implementation of brain-inspired and more complex spike information processing in photonic neuromorphic computing systems.4.The hardware photonic SNN systems based on the FP-SA chip were proposed.Considering the multi-longitudinal mode characteristics of the FP-SA,a hardware system scheme with non-coherent dual-wavelength lights injected into the FP-SA was designed.It was verified that the simultaneous injection of two non-coherent lights can still trigger neuron-like responses of FP-SA,revealing the potential applications of multi-mode lasers in neuromorphic computing.Then,a hardware-algorithm collaborative photonic SNN was constructed and a time-multiplexed spike coding scheme was proposed to simplify the hardware structure.It was demonstrated this network could successfully recognize “XDU”patterns.The system cascaded with two FP-SAs achieved better recognition results.The experimental demonstration based on FP-SA introduces new avenues for co-designing and optimizing both hardware and software in large-scale multi-layer photonic SNNs,enabling them to tackle intricate tasks more effectively.5.A hybrid integrated photonic SNN that combines Indium Phosphide(In P)based VCSELSA photonic spiking neurons with silicon-based MZI photonic synapses was proposed.It served as an inference framework to complete the recognition of numbers “0-3”.It was found that the phase shift error of the MZI array and the bit-width of the drive voltage could affect recognition performance to some extent.Based on the matrix blocking idea and the reconfigurability of a MZI array,a large-scale 400×10 network was simulated by timemultiplexing a 4×4 hybrid integrated small SNN,which achieved the recognition of optical characters “0-9”.Further analysis revealed that the phase shift of the internal phase shifter was more sensitive to noise.This work demonstrates the feasibility of the proposed hybrid integrated photonic SNN as an inference framework for future large-scale photonic neuromorphic applications.It also provides new possibilities and flexibility for the development of multi-mode neuromorphic chips.