This invention discloses an all-optical brain-like training
system and method based on a three-dimensional photonic
memristor array, belonging to the field of photonic brain-like computing technology. This invention constructs a three-dimensional optical
waveguide interconnect network through
femtosecond laser direct writing, and sets GST / LiNbO₃ non-volatile optical modulation units at the
optical path spatial intersection nodes to form a three-dimensional photonic
synapse array. Relying on on-
chip multi-
wavelength light sources,
pulse delay control, and an all-optical differential module, combined with
wavelength division
multiplexing and
time division multiplexing mechanisms, the same physical
optical path can carry forward
inference optical signals and backward error update optical signals in a time-division manner. The
phase transition effect of photo-induced materials is used to achieve in-situ dynamic weight updates in the pure optical domain, constructing an all-optical brain-like training
closed loop without electronic computing power intervention. This invention solves the technical problems of low computing
power density, inability to perform on-
chip autonomous training in optoelectronic
hybrid architectures, and high training
power consumption of existing two-dimensional photonic chips. It has the advantages of high computing
power density, high architectural integration, extremely low
power consumption, and hardware-native iterative learning capabilities, and is suitable for various low-power, high-speed brain-like
artificial intelligence computing scenarios.