Photonic neural network
Inventors
Assignees
Publication Number
US-12340301-B2
Publication Date
2025-06-24
Expiration Date
2040-06-02
Interested in licensing this patent?
MTEC can help explore whether this patent might be available for licensing for your application.
Abstract
A photonic neural network device may include a planar waveguide; a layer having a changeable refractive index adjacent to the planar waveguide; and a plurality of electrodes. Each electrode may be electrically coupled to the layer having the changeable refractive index at a corresponding location of the layer having the changeable refractive index. Each electrode may be configured to apply a corresponding, configurable voltage to the corresponding location to affect a refractive index of the corresponding location of the layer having the changeable refractive index to induce an amplitude modulation or a phase modulation of a light waveform propagating through the photonic neural network device to configure a corresponding neuron of the photonic neural network device in order to perform a computation.
Core Innovation
The invention is a photonic neural network device that includes a planar waveguide, an adjacent layer with a changeable refractive index, and a plurality of electrodes each electrically coupled to specific locations of the layer having the changeable refractive index. Each electrode applies a configurable voltage to its corresponding location to affect the refractive index, thereby inducing amplitude or phase modulation of a light waveform propagating through the device. This modulation configures corresponding neurons in the photonic neural network to perform computations.
The background identifies limitations in existing artificial neural network hardware accelerators, which rely on traditional digital electronics constrained by speed and energy efficiency. Prior photonic neural networks suffer from issues such as a small number of neurons, lack of reconfigurability due to fixed optical connections, or dependence on free-space optical elements unsuitable for integrated chips. These challenges restrict scalability and reconfigurability needed for advanced neural network functions.
The disclosure addresses these challenges by providing a photonic neural network device that supports scalable, reconfigurable computation via voltage-controlled modulation of the changeable refractive index layer adjacent to a planar waveguide. The device can support numerous neurons and connections within a chip-scale area, enabling computations with high speed and energy efficiency surpassing state-of-the-art digital processors. The device leverages analog photonic processing combined with neural network architectures to enable flexible and programmable functionalities.
Claims Coverage
The patent includes three independent claims that cover devices and methods underlying the photonic neural network with configurable electrodes and modulation capabilities. The inventive features encompass structural elements, modulation mechanisms, neural network layering, and training and implementation methods of the device.
Photonic neural network device with voltage-controlled refractive index modulation
A device comprising a planar waveguide, a layer with changeable refractive index adjacent to the waveguide, and a plurality of electrodes each configured to apply a configurable voltage to specific locations of the layer to induce amplitude or phase modulation of a propagating light waveform, thus configuring corresponding neurons to perform computations.
Training and physical implementation method of a photonic neural network device
A method including modeling a photonic neural network device with a planar waveguide, a changeable refractive index layer, and multiple electrodes configured to apply voltages affecting refractive index; training the modeled device by selecting configurable voltages to produce desired output waveforms in response to input waveforms; and physically implementing the trained photonic neural network device based on the training.
Photonic neural network device with multiple neural network layers including columns of neurons and refractive or diffractive elements
A device comprising multiple neural network layers, each including a column of neurons with electrodes adjacent to a changeable refractive index layer adjacent to a planar optical waveguide, and a column of refractive or diffractive elements adjacent to the waveguide; the device is configured to induce amplitude or phase modulation of a light waveform to configure neurons for computation.
The independent claims collectively define a photonic neural network system with configurable neurons achieved by voltage-controlled refractive index modulation adjacent to planar waveguides, methods for modeling and training such a system to achieve desired computational outputs, and a device architecture involving multiple layers of neurons and refractive or diffractive elements supporting complex neural network functionalities.
Stated Advantages
The photonic neural network device can support tens, hundreds, or thousands of neurons in a small chip-scale area, achieving orders of magnitude improvement over previous photonic integrated circuit designs.
It offers energy efficiency with power consumption limited mainly to input electro-optic conversion and low-power electrode voltage control, enabling processing efficiency greater than 5000 TOPS/watt, surpassing digital FPGA and ASIC technologies.
The device enables reconfigurability and programmability through voltage-controlled phase and amplitude modulation, allowing flexible implementation of various processing algorithms and neural network functions.
The approach allows high-speed processing potentially up to 10,000 TOPS with low power consumption below 5 watts, greatly improving performance per watt compared to current AI accelerators.
Documented Applications
High-performance real-time signal and sensor processing systems where data throughput is limited by traditional digital processing power.
Machine learning and artificial intelligence applications requiring analog photonic implementations of neural networks.
Radio frequency (RF) signal classification and maximum likelihood estimation using neural network architectures implemented in photonic processors.
Interested in licensing this patent?