Integrated Silicon Photonic Spiking Neuron Device for Neuromorphic Photonic Computing

Background

The field of neuromorphic photonics leverages photonic integrated circuits to mimic neural spiking behavior for ultrafast computing tasks. Silicon photonic chips promise high bandwidth, low latency, and energy-efficient information processing by routing optical signals through compact waveguides and modulators. As demands for AI acceleration and real-time data analysis grow, there is a pressing need for integrated photonic neurons capable of generating and routing sub-nanosecond optical spikes without resorting to bulky discrete components. Achieving such capabilities on silicon platforms could enable scalable, high-density neural networks for machine learning at unprecedented speeds.

However, conventional injection-locked laser schemes rely on discrete optical circulators and isolators to separate incoming pump signals from outgoing spikes, adding significant footprint and complexity. These magneto-optical devices are typically incompatible with standard CMOS fabrication, making integration onto silicon chips difficult and expensive. Moreover, precise polarization alignment often requires additional splitters and rotators, each introducing insertion loss and alignment sensitivity. The reliance on off-chip components and elaborate polarization control routines creates increased system losses, reliability issues, and poor scalability, preventing the realization of densely integrated, energy-efficient neuromorphic photonic processors.

Technology

This technology consists of a silicon photonic chip that tightly integrates two vertical-cavity surface-emitting lasers (VCSELs) with on-chip polarization rotator/splitters and a polarization splitter to route pump and signal light without discrete circulators or isolators. A pump VCSEL injects orthogonally polarized light into a spiking VCSEL, and the resulting optical spikes are rerouted back into the photonic integrated circuit (PIC) via the same spatial waveguide. Polarization-maintaining fibers or photonic wire bonds couple the VCSELs to the PIC, while on-chip electro-optic or microring resonator modulators control injection power and spike firing rate. A microring weight bank further adjusts spike amplitudes for neuromorphic processing, enabling a compact, single-chip photonic neuron that emits sub-nanosecond optical pulses and supports scalable synaptic weighting.

What sets this approach apart is its complete elimination of bulky magneto-optical components like circulators and isolators by leveraging orthogonal polarization states for bidirectional light routing. Direct compatibility with standard silicon photonic fabrication processes allows seamless integration of high-speed spiking VCSELs, modulators, and weight banks on a single chip. This solves the key barrier to miniaturized, large-scale neuromorphic photonic systems by offering a fully integrated, scalable platform for ultrafast optical neural networks.

Benefits

• Eliminates bulky discrete circulators and isolators by using on-chip polarization control
• Fully compatible with standard silicon photonic fabrication for seamless integration
• Enables compact, scalable neuromorphic photonic processors with tight VCSEL-PIC coupling
• Generates ultrafast sub-nanosecond optical spikes for high-speed artificial spiking neurons
• On-chip modulators and microring weight banks provide precise control over spike rates and synaptic weights
• Uses mature photonic components to reduce complexity and support large-scale neural network scaling

Applications

• Photonic neuromorphic computing chips
• Ultrafast photonic AI accelerators
• Low-latency optical neural networks
• High-speed optical signal processing
• Edge photonic AI processors

Additional info

A silicon photonic chip integrating a pump VCSEL and a spiking VCSEL uses on-chip polarization rotator/splitters and polarization splitters to inject orthogonally polarized pump light into the spiking VCSEL and route its output back without discrete circulators. TE/TM mode conversion permits a single waveguide. Polarization-maintaining fibers or photonic wire bonds ensure coupling. Electro-optic or microring modulators control firing rates, and on-chip microring weight banks adjust spike amplitudes for neuromorphic processing.