Coherent modulation based complex-valued photonic convolutional computing system

Abstract

This application provides a complex-valued photonic convolution computation system based on coherent modulation. Its light source module generates a multi-wavelength optical carrier signal loaded with a weight matrix; a signal processing module converts the input matrix into two time-varying electrical signals; an electro-optic modulation module divides the optical carrier into two equal paths: one path is directly modulated by the first electrical signal, and the other path is modulated by the second electrical signal after adding a 90° phase difference, and the two paths are combined to obtain a complex-valued optical field; an optical signal transmission module delays the dispersion of the complex-valued optical field, creating a fixed delay difference between different wavelength channels; a coherent detection module performs zero-difference coherent detection on the delayed optical signal to recover amplitude and phase information, obtaining the convolution computation result. This application solves the problems of low spectral efficiency and phase information loss in existing photonic computations by employing IQ complex-valued modulation and zero-difference coherent detection, simultaneously utilizing both amplitude and phase dimensions of the optical field to carry information, thus improving spectral efficiency and noise resistance; combined with parallel processing of multiple wavelength channels, it increases computational throughput.

Description

Technical Field

[0001] This application relates to the field of optoelectronic neural network technology, and in particular to a complex-valued photonic convolution computation system based on coherent modulation. Background Technology

[0002] As the core carrier of information transmission and computation, the precise control of optical signals directly determines the parallel processing efficiency and computational accuracy of the system. Light waves contain multiple physical dimensions such as amplitude, phase, and wavelength. By combining different modulation formats, data streams from multiple dimensions can be processed simultaneously. Photonic computing, with its inherent advantages of high bandwidth, low latency, and low power consumption, has become a core technological direction for breaking through the performance bottleneck of electronic computing. Electro-optic modulation is a key step in realizing the mapping of electrical domain input vectors to optical domain signals and completing parallel computation in the optical domain. Multi-dimensional electro-optic modulation can fully exploit the free-dimensional resources of the optical field, improve the information capacity per unit bandwidth and the parallel processing efficiency of the system, and is an important supporting technology for photonic computing architecture.

[0003] Currently, mainstream electro-optic modulation schemes mainly include two categories: phase modulation and intensity modulation. Phase modulation linearly modulates the electrical signal onto the phase of the optical carrier, using a Mach-Zehnder interferometer (MZI) to convert phase changes into intensity changes, which are then directly detected by a photodetector (PD). This scheme has advantages such as high linearity and no need for bias control. Intensity modulation, on the other hand, uses a Mach-Zehnder modulator (MZM) or an electroabsorption modulator to directly load the electrical signal onto the intensity (optical power) of the optical carrier, employing direct detection for reception. It has advantages such as simple structure and low cost.

[0004] However, for phase modulation schemes, ordinary photodiodes (PDs) only respond to optical power and cannot directly detect phase changes, resulting in the complete loss of phase information after direct detection. If coherent digital demodulation is used, a highly stable local oscillator laser and a complex optical mixer are required, significantly increasing the cost and complexity of the receiver. For intensity modulation schemes, direct detection also results in the loss of phase information of the optical field, leading to low system spectral efficiency, limited transmission rate, and susceptibility to power-selective fading. Therefore, there is an urgent need for a photonic computing system that can simultaneously utilize optical field amplitude and phase information, improve spectral efficiency, and not significantly increase receiver complexity. Summary of the Invention

[0005] In view of this, embodiments of this application provide a complex-valued photonic convolution computation system based on coherent modulation to eliminate or improve one or more defects existing in the prior art.

[0006] One aspect of this application provides a complex-valued photonic convolution computation system based on coherent modulation, comprising: The light source module is used to obtain a multi-wavelength optical carrier signal loaded with a weight matrix based on a continuous light source; The signal processing module is used to convert the input matrix into a first time-varying electrical signal and a second time-varying electrical signal; An electro-optic modulation module, connected to the light source module and the signal processing module respectively, is used to divide the multi-wavelength optical carrier signal loaded with a weight matrix into two optical signals. One optical signal is directly modulated by the first time-varying electrical signal, and the other optical signal is modulated by the second time-varying electrical signal after adding a 90° phase difference. The two modulated optical signals are then combined to obtain a complex optical field signal. An optical signal transmission module, connected to the electro-optic modulation module, is used to perform dispersive delay on the complex-valued optical field signal so that a fixed delay difference is generated between different wavelength channels of the complex-valued optical field signal to obtain a delayed optical signal. The coherent detection module, connected to the optical signal transmission module, is used to perform zero-difference coherent detection on the delayed optical signal to recover the amplitude and phase information of the delayed optical signal, and to obtain the convolution calculation result of the input matrix and the weight matrix.

[0007] In some embodiments of this application, in the optical signal transmission module, the fixed delay difference is the delay step size between adjacent wavelength channels of the complex-valued optical field signal; The delay step is set to be equal to the symbol period of the first time-varying electrical signal and the second time-varying electrical signal, and satisfies the following quantitative relationship: in, This indicates the delay step size; Indicates the length of the tunable dispersive fiber; Represents the tunable dispersion coefficient; Indicates the wavelength interval between adjacent wavelength channels; This indicates the symbol period.

[0008] In some embodiments of this application, the light source module includes: a continuous light source, an EDFA amplifier, a first polarization controller, a micro-ring resonator, and a waveform shaper connected in sequence; The single-wavelength laser output from the continuous light source is injected into the micro-ring resonant cavity after passing through the EDFA amplifier and the first polarization controller. A Kerr optical frequency comb is generated through the four-wave mixing effect to serve as a multi-wavelength optical carrier signal. The waveform shaper independently controls the optical power of each tooth of the Kerr optical frequency comb, and maps each weight value in the weight matrix to a single wavelength channel to obtain a multi-wavelength optical carrier signal loaded with the weight matrix.

[0009] In some embodiments of this application, the first time-varying electrical signal and the second time-varying electrical signal generated by the signal processing module are obtained by directly splitting the input matrix after flattening it.

[0010] In some embodiments of this application, the signal processing module includes: First digital signal processor; An arbitrary waveform generator, connected to the first digital signal processor, is used to flatten the input matrix into a one-dimensional vector under the control of the digital signal processor, map the one-dimensional vector into a time-varying electrical signal waveform through digital-to-analog conversion, and evenly split the time-varying electrical signal waveform into a first time-varying electrical signal and a second time-varying electrical signal.

[0011] In some embodiments of this application, the electro-optic modulation module includes: A one-to-two optical splitter is connected to the light source module and is used to receive the multi-wavelength optical carrier signal loaded with a weight matrix and to divide the multi-wavelength optical carrier signal loaded with a weight matrix into two optical signals. The first Mach-Zehnder modulator is connected to the one-to-two optical splitter and the signal processing module, and the first Mach-Zehnder modulator receives the first time-varying electrical signal and is driven by the first time-varying electrical signal to modulate one optical signal output by the one-to-two optical splitter. The second Mach-Zehnder modulator is connected to the one-to-two optical splitter and the signal processing module respectively. The second Mach-Zehnder modulator receives the second time-varying electrical signal and is driven by the second time-varying electrical signal to modulate the other optical signal output by the one-to-two optical splitter by adding a 90° phase difference. An output coupler is connected to the first Mach-Zehnder modulator and the second Mach-Zehnder modulator respectively, and is used to combine the modulated optical signals output by the first Mach-Zehnder modulator and the second Mach-Zehnder modulator respectively to obtain a complex-valued optical field signal. Both the first and second Mach-Zehnder modulators operate in push-pull mode and have their DC bias set at the quadrature bias point.

[0012] In some embodiments of this application, the coherent detection module includes: The second polarization controller, connected to the optical signal transmission module, is used to adjust the polarization state of the delayed optical signal to obtain a polarization-adjusted delayed optical signal. A local oscillator laser is used to generate local oscillator light; wherein the frequency of the local oscillator light is the same as the frequency of the continuous light source in the light source module; A 90° optical mixer, connected to the second polarization controller and the local oscillator laser, is used to coherently mix the polarization-adjusted delayed optical signal with the local oscillator light and output four mixed optical signals. A balanced receiver is connected to the 90° optical mixer. The balanced receiver includes two photodetector groups, each consisting of two photodetectors. Each photodetector group differentially detects two optical signals with a phase difference of 180° in the four mixed optical signals to convert the four mixed optical signals into two electrical signals.

[0013] In some embodiments of this application, the coherent detection module further includes: The second digital signal processor, connected to the balanced receiver, is used to sequentially perform orthogonalization and normalization, clock recovery, dispersion compensation, polarization equalization, carrier frequency offset estimation and phase recovery, and time demultiplexing on the two electrical signals output by the balanced receiver, so as to obtain the convolution calculation result of the input matrix and the weight matrix.

[0014] In some embodiments of this application, the dispersion delay in the optical signal transmission module also satisfies the following full calculation model: in, This indicates the delay step size, which is the length of the dispersive fiber. and wavelength The function; This represents the difference in path length between different wavelength channels; Indicates group velocity; Represents the speed of light; Indicates the difference in refractive index; Indicates the dispersion coefficient; This indicates the wavelength interval between adjacent wavelength channels.

[0015] In some embodiments of this application, the system is used for convolutional layer calculation in an opto-neural network, wherein the number of different wavelength channels corresponds to the number of weights of the convolutional kernel, the input matrix is ​​the input feature map of the opto-neural network, and the weight matrix is ​​the convolutional kernel weights of the opto-neural network; The system performs a sliding window multiplication and accumulation operation on the input feature map and the convolution kernel weights through time-delayed interleaving of multiple wavelength channels to obtain the convolution calculation result of the input matrix and the weight matrix as the output of the convolution layer of the photoelectric neural network.

[0016] The complex-valued photonic convolution computation system based on coherent modulation provided in this application includes: a light source module for obtaining a multi-wavelength optical carrier signal loaded with a weight matrix from a continuous light source; a signal processing module for converting the input matrix into a first time-varying electrical signal and a second time-varying electrical signal; an electro-optic modulation module connected to the light source module and the signal processing module, for dividing the multi-wavelength optical carrier signal loaded with the weight matrix into two optical signals, one optical signal being directly modulated by the first time-varying electrical signal, and the other optical signal being modulated by the second time-varying electrical signal after adding a 90° phase difference, and then combining the two modulated optical signals to obtain a complex-valued optical field signal; an optical signal transmission module connected to the electro-optic modulation module for dispersively delaying the complex-valued optical field signal to generate a fixed delay difference between different wavelength channels of the complex-valued optical field signal, to obtain a delayed optical signal; and a coherent detection module connected to the optical signal transmission module for performing zero-difference coherent detection on the delayed optical signal to recover the amplitude and phase information of the delayed optical signal, to obtain... The convolution calculation result of the input matrix and the weight matrix, in this coherent modulation-based complex-valued photonic convolution calculation system, can perform IQ complex-valued modulation on multi-wavelength optical carrier signals through an electro-optic modulation module. This loads the two time-varying electrical signals of the input matrix onto the in-phase and quadrature components of the optical carrier, respectively, thereby simultaneously utilizing both amplitude and phase dimensions to carry information on the same optical carrier. Under the same bandwidth conditions, this doubles the data carrying capacity per unit symbol compared to single-dimensional modulation, thus improving the system's spectral efficiency. The optical signal transmission module performs dispersive delay on the complex-valued optical field signal, creating a fixed delay difference between different wavelength channels. This enables pipelined data processing using the parallelism of multiple wavelength channels and the delay difference, improving the system's parallel processing capability. The coherent detection module performs zero-difference coherent detection on the delayed optical signal, completely recovering the amplitude and phase information of the optical field and avoiding phase information loss due to direct detection. This allows for accurate demodulation of the convolution calculation result of the input matrix and the weight matrix at the receiving end, improving the system's computational accuracy and noise immunity.

[0017] Additional advantages, objectives, and features of this application will be set forth in part in the description which follows, and will in part become apparent to those skilled in the art upon review of the following description, or may be learned by practice of the application. The objectives and other advantages of this application can be realized and obtained by means of the structures specifically pointed out in the specification and drawings.

[0018] Those skilled in the art will understand that the purposes and advantages that can be achieved with this application are not limited to those specifically described above, and that the above and other purposes that this application can achieve will be more clearly understood from the following detailed description. Attached Figure Description

[0019] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, do not constitute a limitation thereof. The components in the drawings are not drawn to scale but are merely for illustrating the principles of this application. For ease of illustration and description of certain parts of this application, corresponding portions in the drawings may be enlarged, i.e., may appear larger relative to other components in an exemplary device actually manufactured according to this application. In the drawings: Figure 1 This is a schematic diagram of the structure of a complex-valued photonic convolution computation system based on coherent modulation according to an embodiment of this application.

[0020] Figure 2 This is a schematic diagram of the signal flow of the light source module in one embodiment of this application.

[0021] Figure 3 This is a schematic diagram of the signal processing module in one embodiment of this application.

[0022] Figure 4 This is a schematic diagram of the electro-optic modulation module in one embodiment of this application.

[0023] Figure 5 This is a schematic diagram of the signal flow of the optical signal transmission module in one embodiment of this application.

[0024] Figure 6 This is a schematic diagram of the signal flow of the coherent detection module in one embodiment of this application.

[0025] Figure 7 This is a schematic diagram of the structure of a complex-valued photonic convolution computation system based on coherent modulation in an application example of this application.

[0026] Figure 8 This is a schematic diagram illustrating the principle of sliding window multiplication and accumulation operation in an application example of this application, where the input data and weights are staggered and intertwined on the time axis after the multi-wavelength channels are dispersed and delayed.

[0027] Figure 9 This is a flowchart of the processing of the second digital signal processor in the coherent detection module of an application example of this application. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and their descriptions are used to explain this application, but are not intended to limit it.

[0029] It should also be noted that, in order to avoid obscuring this application with unnecessary details, only the structures and / or processing steps closely related to the solution according to this application are shown in the accompanying drawings, while other details that are not closely related to this application are omitted.

[0030] It should be emphasized that the term "including / comprises" as used herein refers to the presence of a feature, element, step, or component, but does not exclude the presence or addition of one or more other features, elements, steps, or components.

[0031] It should also be noted that, unless otherwise specified, the term "connection" in this article can refer not only to a direct connection, but also to an indirect connection involving an intermediary.

[0032] In the following description, embodiments of the present application will be illustrated with reference to the accompanying drawings. In the drawings, the same reference numerals represent the same or similar parts, or the same or similar steps.

[0033] It should be noted that IQ modulation is a high-order electro-optic modulation technique. Its core principle is to utilize the in-phase and quadrature vertical components of an optical carrier to achieve parallel signal modulation. Its principle is based on quadrature carrier modulation: the electrical signal to be transmitted is split into two baseband signals, I and Q, which are respectively loaded onto two optical carriers of the same frequency but 90° out of phase. The amplitude and phase of the two optical carriers are independently controlled by an electro-optic modulator. After the two modulated optical signals are combined, a single optical domain signal is formed. This signal can carry twice the information of a single modulation, achieving multi-dimensional signal multiplexing. Demodulation only requires separating the I and Q paths using a phase shifter to recover the original signal, combining high bandwidth utilization with the ability to carry complex information.

[0034] Convolution is the core operation in optoelectronic convolution computing systems. The convolution operation originated in the field of signal processing, such as in signal processing where a signal is input at each time step m. The formula is obtained as follows: In the formula This refers to the weight vector (or convolution kernel function). In offset The weight value at that location, For signal Input vector at time step, In order to be in The output signal at time t. This corresponds to a one-dimensional convolution operation, which involves flipping the convolution kernel vector and performing a dot product summation with the input signal vector. However, in optoelectronic neural networks, since the convolution kernel is learned through training rather than pre-designed, the "flipping" step is redundant. Omitting the flipping step allows the hardware to perform the calculation more directly and efficiently. Therefore, the optoelectronic convolution calculation system can achieve the convolution operation by mainly handling the sliding dot product summation operation.

[0035] Multidimensional electro-optic modulation can fully exploit the free-dimensional resources of optical field amplitude, phase, and wavelength, and improve the information capacity per unit bandwidth and the parallel processing efficiency of the system. It is a key supporting technology for photonic computing architecture. However, the current mainstream phase modulation and intensity modulation schemes have inherent technical defects that are difficult to overcome, which seriously restrict the performance implementation of photonic computing systems.

[0036] To address the issues of low spectral efficiency and loss of phase information in existing photonic computation systems, this application provides a complex-valued photonic convolution computation system based on coherent modulation. It employs IQ complex-valued modulation and zero-difference coherent detection, while utilizing both the amplitude and phase dimensions of the optical field to carry information, thereby improving spectral efficiency and noise immunity. Furthermore, it combines multi-wavelength channel parallel processing to enhance computational throughput.

[0037] The following examples will provide a detailed description.

[0038] Based on this, embodiments of this application provide a complex-valued photonic convolution computation system based on coherent modulation, see [link to relevant documentation]. Figure 1 The complex-valued photonic convolution computation system based on coherent modulation specifically includes the following: (1) Light source module 10, used to obtain a multi-wavelength optical carrier signal loaded with a weight matrix based on a continuous light source.

[0039] Continuous optical sources refer to single-wavelength lasers with stable output power and narrow linewidth, such as distributed feedback laser diodes (DFB-LDs). The weight matrix is ​​a set of convolution kernel parameters pre-trained in an optoelectronic neural network. Multi-wavelength optical carrier signals refer to optical signals containing multiple discrete wavelength channels, each of which can independently carry a weight value.

[0040] In practice, a single-wavelength continuous laser beam is first generated by a continuous light source. After optical power amplification and polarization state adjustment, this laser beam is injected into a microring resonator (MRR). The nonlinear effect of four-wave mixing (FWM) within the cavity generates a Kerr optical frequency comb, resulting in multiple equally spaced wavelength channels. Subsequently, a waveform shaper independently controls the optical power of each wavelength channel: each weight value in the weight matrix is ​​assigned to the corresponding wavelength channel according to a preset mapping rule; for example, the weight value is normalized and converted into the optical power attenuation of that channel. After this processing, the output of the light source module receives a multi-wavelength optical carrier signal loaded with the weight matrix.

[0041] (2) Signal processing module 20, used to convert the input matrix into a first time-varying electrical signal and a second time-varying electrical signal.

[0042] The input matrix is ​​the data to be used in the convolution operation, such as the output feature map of the previous layer in a photoelectric neural network or the original input image data. The time-varying electrical signal refers to an analog voltage waveform whose amplitude changes over time, and its variation is determined by the numerical sequence of the input matrix. The first and second time-varying electrical signals have the same symbol period and are mathematically orthogonal.

[0043] In practice, the signal processing module first flattens the input matrix into a one-dimensional vector using either row-major or column-major order. Then, a digital-to-analog converter (DAC) sequentially maps each value in the one-dimensional vector to an analog voltage amplitude within the corresponding time interval, forming a continuous time-varying electrical signal waveform. Finally, this time-varying electrical signal waveform is evenly split into two parallel branches: the first branch directly outputs as the first time-varying electrical signal, and the second branch outputs as the second time-varying electrical signal. Both signals have the same symbol period τ and amplitude range, but are independent of each other.

[0044] (3) Electro-optic modulation module 30, which is connected to the light source module and the signal processing module respectively, is used to divide the multi-wavelength optical carrier signal loaded with weight matrix into two optical signals. One optical signal is directly modulated by the first time-varying electrical signal, and the other optical signal is modulated by the second time-varying electrical signal after adding a 90° phase difference. The two modulated optical signals are then combined to obtain a complex optical field signal.

[0045] It should be noted that "equal division" refers to distributing the input optical signal equally into two beams in terms of power, typically achieved using fiber optic couplers or waveguide couplers. A 90° phase difference refers to a relative phase offset of π / 2 radians between the two optical signals, which can be achieved using phase shifters or unequal-length optical paths. A complex-valued optical field signal refers to an optical signal that simultaneously carries information about in-phase and quadrature components.

[0046] In practical implementation, the electro-optic modulation module receives a multi-wavelength optical carrier signal from the light source module. This signal enters a 1-to-2 optical splitter (e.g., a 50:50 fiber coupler) and is split into two equal paths: a first optical signal and a second optical signal. The first optical signal is directly fed into a first Mach-Zehnder modulator (MZM), while a first time-varying electrical signal I(t) is applied as a driving voltage to the electrodes of the MZM, thereby linearly converting the amplitude change of I(t) into the intensity change of the first optical signal. The second optical signal first passes through a 90° phase shifter (e.g., by adding a precisely long optical waveguide) to lag its phase by 90°, and then is fed into a second Mach-Zehnder modulator. Simultaneously, a second time-varying electrical signal Q(t) drives the MZM, modulating Q(t) onto the phase-shifted second optical signal. Both MZMs operate in push-pull mode with their DC bias set at the minimum point of their power transfer curves (i.e., the quadrature bias point) to ensure modulation linearity. After modulation, the two optical signals are combined through an output coupler, and the combined optical signal is the complex-valued optical field signal. Since each wavelength channel in the multi-wavelength optical carrier signal undergoes the above-mentioned equalization, modulation, and combining processes simultaneously, each wavelength channel in the complex-valued optical field signal carries the same I(t) and Q(t) information, but the optical field amplitude of each channel differs due to the different pre-loaded weight values.

[0047] (4) Optical signal transmission module 40, connected to the electro-optic modulation module, is used to perform dispersion delay on the complex optical field signal so that a fixed delay difference is generated between different wavelength channels of the complex optical field signal to obtain the delayed optical signal.

[0048] As can be understood, dispersion delay refers to the physical phenomenon that when an optical signal propagates in a dispersive medium (such as single-mode fiber), different wavelength components experience different propagation times due to their different group velocities. A fixed delay difference means that the relative time delay between adjacent wavelength channels is a constant value that does not change over time. The delayed optical signal refers to the optical signal in which, after transmission through the dispersive medium, the wavelength channels are sequentially staggered along the time axis.

[0049] During implementation, the complex-valued optical field signal output from the electro-optic modulation module is fed into a section of dispersive optical fiber (e.g., standard single-mode fiber or dispersion-compensating fiber). Because the group velocity of light from different wavelength channels differs within the fiber, longer wavelength components propagate faster, while shorter wavelength components propagate slower (in the normal dispersion region), and vice versa. After passing through a dispersive fiber of length L, a fixed time difference is generated between adjacent wavelength channels. By selecting appropriate fiber length and dispersion coefficients, this fixed delay difference can be made to reach the designed value (e.g., equal to the symbol period of the input signal). Simultaneously, a polarization controller can be placed before or after the dispersive fiber to adjust the polarization state of the optical signal, compensating for random polarization changes caused by fiber bending or environmental stress during transmission, and ensuring polarization matching for subsequent coherent detection. The output of the optical signal transmission module is the delayed optical signal, in which the data carried by different wavelength channels are sequentially staggered along the time axis according to wavelength.

[0050] (5) Coherent detection module 50, connected to the optical signal transmission module, is used to perform zero-difference coherent detection on the delayed optical signal to recover the amplitude and phase information of the delayed optical signal and obtain the convolution calculation result of the input matrix and the weight matrix.

[0051] In this context, zero-difference coherent detection refers to mixing the signal light with a local oscillator light of the exact same frequency as the signal light, directly obtaining the baseband signal after photoelectric conversion without intermediate frequency processing. Recovering amplitude and phase information involves converting the complex amplitude (including real and imaginary parts) carried in the optical signal into two electrical signals through optical mixing and balanced detection, thereby obtaining the original modulated I(t) and Q(t) components. The convolution calculation result refers to the output feature map data obtained after the input matrix and weight matrix undergo a sliding window multiplication and accumulation operation.

[0052] In practice, after the delay, the optical signal first enters the polarization controller inside the coherent detection module. This polarization controller adjusts the polarization state of the signal light to match that of the local oscillator laser to eliminate the reduction in mixing efficiency caused by polarization mismatch. The local oscillator laser generates a local oscillator beam with the same frequency as the continuous light source in the light source module (i.e., zero-difference condition). The polarization-adjusted signal light and the local oscillator beam are simultaneously fed into a 90° optical mixer. The 90° optical mixer contains multiple couplers and phase shifters, and can output four mixed optical signals, corresponding to in-phase addition, in-phase subtraction, quadrature addition, and quadrature subtraction of the signal light and the local oscillator beam, respectively. These four optical signals are received by two pairs of balanced photodetectors (BPDs): each pair of balanced receivers consists of two matched photodetectors (PDs), which differentially detect two optical signals with a phase difference of 180°. Differential detection can suppress local oscillator intensity noise and environmental interference in common mode, significantly improving the signal-to-noise ratio.

[0053] The two electrical signals output by the balanced receiver carry linear combinations of I(t) and Q(t), respectively. Since the local oscillator frequency is exactly the same as the signal light frequency, these two electrical signals are the baseband signals and do not require further frequency down-conversion. By performing subsequent processing on these two baseband signals (such as analog-to-digital conversion and digital signal processing), the original I(t) and Q(t) waveforms can be recovered, and the convolution result of the input matrix and the weight matrix can be obtained. This convolution result can be directly used as the input or final output of the next layer of the opto-neural network.

[0054] As described above, the complex-valued photonic convolution calculation system based on coherent modulation provided in this application can perform IQ complex-valued modulation on multi-wavelength optical carrier signals through an electro-optic modulation module, loading the two time-varying electrical signals of the input matrix onto the in-phase and quadrature components of the optical carrier respectively. This allows information to be carried on the same optical carrier simultaneously using both amplitude and phase dimensions, doubling the data carrying capacity per unit symbol compared to single-dimensional modulation under the same bandwidth conditions, thereby improving the system's spectral efficiency. The optical signal transmission module performs dispersive delay on the complex-valued optical field signal, creating a fixed delay difference between different wavelength channels. This enables pipelined data processing using the parallelism of multiple wavelength channels and the delay difference, improving the system's parallel processing capability. The coherent detection module performs zero-difference coherent detection on the delayed optical signal, completely recovering the amplitude and phase information of the optical field and avoiding phase information loss due to direct detection. This allows for accurate demodulation of the convolution calculation result of the input matrix and weight matrix at the receiving end, improving the system's computational accuracy and noise immunity.

[0055] To further address the limitations of existing light sources in efficiently generating dense multi-wavelength channels and flexibly allocating weights, this application provides a complex-valued photonic convolution computation system based on coherent modulation, see [link to relevant documentation]. Figure 2 The light source module 10 specifically includes a continuous light source 11, an EDFA amplifier 12, a first polarization controller 13, a micro-ring resonator 14, and a waveform shaper 15 connected in sequence.

[0056] The single-wavelength laser output from the continuous light source 11 is injected into the micro-ring resonant cavity 14 after passing through the EDFA amplifier 12 and the first polarization controller 13. A Kerr optical frequency comb is generated through the four-wave mixing effect to serve as a multi-wavelength optical carrier signal. The waveform shaper 15 independently controls the optical power of each tooth of the Kerr optical frequency comb, and maps each weight value in the weight matrix to a single wavelength channel to obtain a multi-wavelength optical carrier signal loaded with the weight matrix.

[0057] The EDFA (Erbium-Doped Fiber Amplifier) ​​is used to amplify the single-wavelength laser output from a continuous-wavelength optical source, bringing its power to a threshold sufficient to drive nonlinear effects. The polarization controller (PC) adjusts the polarization state of the input light to match the resonant mode of the microring resonator, maximizing the enhancement of the optical field within the cavity. A microring resonator (MRR) is a ring-shaped optical waveguide structure with dimensions on the micrometer scale. When the light wavelength meets the resonance condition, the light resonates within the ring, significantly enhancing the optical field and thus exciting nonlinear effects. Four-wave mixing (FWM) is a third-order nonlinear optical process. When a strong pump light propagates in a nonlinear medium, it generates new frequency components, forming equally spaced frequency combs. A Kerr optical frequency comb is based on the Kerr nonlinear effect, generating multi-wavelength optical signals with equal frequency spacing. Its spectrum exhibits a comb-like shape, with each comb tooth corresponding to an independent wavelength channel. A waveform shaper is a programmable optical filter that can independently adjust the amplitude (optical power) and phase of each wavelength channel. A comb tooth is a discrete wavelength component in a Kerr optical frequency comb, similar to the teeth on a comb, with equal frequency spacing between adjacent teeth.

[0058] In one example, assume the continuous light source is a distributed feedback semiconductor laser (DFB-LD) outputting a continuous laser with a center wavelength of 1550 nm and a power of 10 mW. This laser first enters an EDFA amplifier, where the power is boosted to 500 mW to meet the threshold requirements for four-wave mixing. The amplified light then passes through a first polarization controller, adjusting its polarization state to TE mode (transverse electric mode) to match the resonant polarization direction of the microring resonator. Subsequently, the light is injected into a silicon-based microring resonator with a radius of 50 μm (free spectral range FSR = 200 GHz). When the pump wavelength is precisely at the resonant peak of the MRR, the optical field within the cavity is sharply enhanced, exciting a significant four-wave mixing effect, generating a Kerr frequency comb centered on the pump wavelength and spaced 200 GHz apart. For example, this comb contains five teeth with wavelengths of 1549.2 nm, 1549.6 nm, 1550.0 nm, 1550.4 nm, and 1550.8 nm. The optical frequency comb outputs a multi-wavelength optical carrier signal to a waveform shaper. The waveform shaper independently measures the optical power of each comb tooth and attenuates the power of the corresponding tooth according to a preset weight matrix (e.g., [0.2, 0.5, 0.8, 0.3, 0.6]): attenuating the power of the first tooth to 20%, the second to 50%, and so on. After passing through the waveform shaper, the power of each wavelength channel represents a weight value, thus obtaining a multi-wavelength optical carrier signal loaded with the weight matrix.

[0059] As can be seen from the above description, the complex-valued photonic convolution computing system based on coherent modulation provided in this application generates a Kerr optical frequency comb through a micro-ring resonant cavity and independently controls the power of each comb tooth in combination with a waveform shaper, which can realize a one-to-one mapping between a single wavelength and a single weight, thereby improving the utilization rate of wavelength resources and the parallel processing capability of the system.

[0060] To further address the problem in existing technologies that require preprocessing of input data, such as Hilbert transform or carrier coding, to generate two orthogonal signals, in a complex-valued photonic convolution computation system based on coherent modulation provided in this application embodiment, the first and second time-varying electrical signals generated by the signal processing module 20 are obtained by directly splitting the input matrix after flattening.

[0061] Flattening is the operation of rearranging a multidimensional input matrix into a one-dimensional vector in row-major or column-major order. For example, a 2×2 matrix [[a,b],[c,d]] flattens to obtain a one-dimensional vector [a,b,c,d]. Direct splitting involves sequentially assigning the elements of the flattened one-dimensional vector to the first and second time-varying electrical signals without undergoing additional mathematical transformations such as Hilbert transform or carrier coding.

[0062] As can be seen from the above description, the complex-valued photonic convolution computing system based on coherent modulation provided in this application generates two time-varying electrical signals by directly splitting them after flattening, which can avoid complex preprocessing steps and thus reduce system latency and hardware overhead.

[0063] To further address the issues of requiring multiple independent DACs and complex timing synchronization in existing solutions for generating orthogonal analog signals, this application provides a complex-valued photonic convolution computation system based on coherent modulation, see [link to relevant documentation]. Figure 3 The signal processing module 20 specifically includes the following components: (1) First digital signal processor 21; (2) Arbitrary waveform generator 22, connected to the first digital signal processor 21, is used to flatten the input matrix into a one-dimensional vector under the control of the digital signal processor, map the one-dimensional vector into a time-varying electrical signal waveform through digital-to-analog conversion, and split the time-varying electrical signal waveform into a first time-varying electrical signal and a second time-varying electrical signal.

[0064] The first digital signal processor, which can be abbreviated as the first DSP, is a programmable digital processing chip used to perform digital domain operations such as matrix flattening and waveform parameter configuration, and to send control commands to the arbitrary waveform generator. An arbitrary waveform generator (AWG) is an electronic instrument capable of generating arbitrary analog voltage waveforms from digital waveform data, integrating a digital-to-analog converter and waveform storage memory. Digital-to-analog conversion (DAC) is the process of converting discrete digital values ​​into continuous-time, continuous-amplitude analog voltage signals.

[0065] In one example, assume the input matrix is ​​a 2×2 matrix [[1,2],[3,4]]. The first digital signal processor (DSP) first flattens this matrix into a one-dimensional vector [1,2,3,4] in row-major order and calculates the waveform sample sequence for each signal: the sample corresponding to the first time-varying signal is [1,3], and the sample corresponding to the second time-varying signal is [2,4]. The first DSP sends these sample data and parameter configurations such as the symbol period τ=100ps to the arbitrary waveform generator (AWG). According to the received configuration, the AWG stores the two sample sequences in its internal memory and generates two analog voltage waveforms through a digital-to-analog converter (DAC): the first waveform outputs 1V in the first symbol period and 3V in the second symbol period; the second waveform outputs 2V in the first symbol period and 4V in the second symbol period. The two waveforms are completely synchronized, and the voltage value remains constant within each symbol period (zero-order hold). The AWG outputs these two analog electrical signals directly as the first and second time-varying signals. The entire process requires no external filters or additional modulators, enabling direct conversion from a digital input matrix to two analog quadrature signals.

[0066] As can be seen from the above description, the complex-valued photonic convolution calculation system based on coherent modulation provided in this application embodiment can generate two precise time-varying electrical signals synchronously within one instrument by controlling an arbitrary waveform generator through a digital signal processor, thereby simplifying the circuit structure and improving the signal synchronization accuracy.

[0067] To further address the issues of traditional single-MZM intensity modulation's inability to simultaneously process two orthogonal signals and its susceptibility to chirping, this application provides a complex-valued photonic convolution computation system based on coherent modulation, see [link to relevant documentation]. Figure 4 The electro-optic modulation module 30 specifically includes the following components: (1) A one-to-two optical splitter 31 is connected to the light source module 10 and is used to receive the multi-wavelength optical carrier signal loaded with a weight matrix and to divide the multi-wavelength optical carrier signal loaded with a weight matrix into two optical signals. (2) The first Mach-Zehnder modulator 32 is connected to the one-to-two optical splitter 31 and the signal processing module 20 respectively. The first Mach-Zehnder modulator 32 receives the first time-varying electrical signal and is driven by the first time-varying electrical signal to modulate one optical signal output by the one-to-two optical splitter 31. (3) The second Mach-Zehnder modulator 33 is connected to the one-to-two optical splitter 31 and the signal processing module 20 respectively. The second Mach-Zehnder modulator 33 receives the second time-varying electrical signal and is driven by the second time-varying electrical signal to modulate the other optical signal output by the one-to-two optical splitter 31 after adding a 90° phase difference. (4) Output coupler 34 is connected to the first Mach-Zehnder modulator 32 and the second Mach-Zehnder modulator 33 respectively, and is used to combine the modulated optical signals output by the first Mach-Zehnder modulator 32 and the second Mach-Zehnder modulator 33 respectively to obtain a complex optical field signal. The first Mach-Zehnder modulator 32 and the second Mach-Zehnder modulator 33 both operate in push-pull mode and the DC bias is set at the quadrature bias point.

[0068] Assume the multi-wavelength optical carrier signal output by the light source module, loaded with a weighted matrix, contains 5 wavelength channels. The optical power of each channel has been set according to the weight values ​​[0.2, 0.5, 0.8, 0.3, 0.6]. The signal enters a 1-to-2 optical splitter (e.g., a 50:50 fiber coupler) and is split into two optical signals: a first optical signal and a second optical signal. The first optical signal is directly fed into the first Mach-Zehnder modulator (MZM1), while the first time-varying electrical signal I(t) (e.g., an analog waveform with an amplitude of 0~1V) provided by the signal processing module is applied to the RF electrodes of MZM1 as a driving voltage. MZM1 adopts a push-pull mode (the polarities of the two arms are opposite) and is DC biased at the quadrature bias point (i.e., the lowest point of the power transfer curve), so that the intensity of the output optical field is proportional to I(t), and the introduced chirp is approximately zero. The second optical signal first passes through a 90° phase shifter (which can be integrated into the input waveguide of MZM2) to lag the phase by 90°, and then is fed into the second Mach-Zehnder modulator (MZM2). The MZM2 also operates in push-pull mode and is biased at a quadrature bias point. It is driven by a second time-varying electrical signal Q(t) to linearly modulate the phase-delayed second optical signal. The two modulated optical signals are combined through an output coupler (e.g., a 2×1 waveguide coupler), and the resulting optical field is a complex-valued optical field signal. Each wavelength channel in this complex-valued optical field signal carries the same I(t) and Q(t) information, but the amplitudes differ due to different weights.

[0069] As can be seen from the above description, the complex-valued photonic convolution calculation system based on coherent modulation provided in this application embodiment can achieve high linearity and low chirp IQ complex-valued modulation by combining dual MZM push-pull mode with orthogonal bias, thereby improving modulation accuracy and system stability.

[0070] To further address the problem in existing photonic convolution calculations where the delay difference and symbol period mismatch between different wavelength channels prevent precise stepping of the convolution sliding window, this application provides a complex-valued photonic convolution calculation system based on coherent modulation. (See [link to relevant documentation]). Figure 5 The optical signal transmission module 40 specifically includes the following components: The dispersion delay unit 41, connected to the electro-optic modulation module 30, is a tunable dispersion fiber (e.g., a standard single-mode fiber or a dispersion-compensating fiber) used to generate a wavelength-related group delay difference for the complex-valued optical field signal. This group delay difference is a fixed delay difference between adjacent wavelength channels of the complex-valued optical field signal (i.e., the delay step between adjacent wavelength channels).

[0071] The third polarization controller 42 is connected to the dispersion delay unit 41 and is used to adjust the polarization state of the optical signal, compensate for polarization drift during transmission, and ensure that the polarization state of the output light matches the local oscillator light of the subsequent coherent detection module.

[0072] The complex-valued optical field signal first enters the dispersion delay unit 41, where different wavelength channels have a fixed delay difference due to their different group velocities; then it enters the third polarization controller 42, which adjusts the polarization state of the optical signal to match the local oscillator light of the subsequent coherent detection module.

[0073] It should be noted that the fixed delay difference is the delay step between adjacent wavelength channels of the complex optical field signal; The delay step is set to be equal to the symbol period of the first time-varying electrical signal and the second time-varying electrical signal, and satisfies the following quantitative relationship: in, This indicates the delay step size; Indicates the length of the tunable dispersive fiber; Represents the tunable dispersion coefficient; Indicates the wavelength interval between adjacent wavelength channels; This indicates the symbol period.

[0074] It should be noted that the delay step size This is the fixed time difference between adjacent wavelength channels due to dispersion. The tunable dispersion fiber length L is the physical length of the dispersion fiber, which can be adjusted by replacing the fiber or using a variable optical delay line. The tunable dispersion coefficient D is the dispersion parameter of the fiber, which can be adjusted by selecting different types of fiber (such as standard single-mode fiber, dispersion-compensated fiber) or using tunable dispersion devices. The wavelength spacing between adjacent wavelength channels... It is the wavelength difference between adjacent comb teeth in the Kerr optical frequency comb, which is determined by the free spectral range (FSR) of the microring resonator.

[0075] In one example, assuming the symbol period of the input signal is τ = 100 ps (picoseconds), and the adjacent wavelength spacing of the Kerr frequency comb... =0.8nm, using standard single-mode fiber with a dispersion coefficient D=17ps / (nm·km). To increase the delay step size... equal Based on the above formula, the required fiber optic length can be calculated as follows: L= 100ps / (17ps / (nm·km)×0.8nm)≈7.35km.

[0076] In a practical system, a 7.35km long single-mode fiber can be selected, and the equivalent dispersion coefficient can be finely adjusted using a tunable dispersion compensator to make the delay step between adjacent wavelength channels exactly equal to 100ps. At this time, the same data symbols carried by different wavelength channels are staggered by exactly one symbol period on the time axis, thereby achieving precise alignment of the subsequent convolution sliding window.

[0077] As can be seen from the above description, the complex-valued photonic convolution calculation system based on coherent modulation provided in this application can ensure that the convolution sliding window steps without deviation in time by precisely setting the delay step size to the symbol period, thereby improving the accuracy and reliability of multiply-accumulate operations.

[0078] To further address the problem of calculation errors caused by neglecting group velocity differences and refractive index variations in simplified delay models, in a complex-valued photonic convolution calculation system based on coherent modulation provided in this application embodiment, the dispersion delay in the optical signal transmission module 40 also satisfies the following full calculation model: in, This indicates the delay step size, which is the length of the dispersive fiber. and wavelength The function; This represents the difference in path length between different wavelength channels; Indicates group velocity; Represents the speed of light; Indicates the difference in refractive index; Indicates the dispersion coefficient; This indicates the wavelength interval between adjacent wavelength channels.

[0079] The path length difference is the additional optical path difference caused by small differences in the propagation paths of different wavelength channels in a dispersive medium (e.g., due to waveguide dispersion or unequal geometric paths), and is usually approximated as zero or determined by design. The group velocity is the propagation speed of the light wave envelope (i.e., energy or information), related to the wavelength λ, and its reciprocal determines the delay per unit length. The refractive index difference is the difference in effective refractive index between different wavelength channels, caused by both material dispersion and waveguide dispersion. The dispersion coefficient is the delay difference per unit wavelength interval and per unit length of optical fiber.

[0080] In one example, assume the adjacent wavelength spacing of the Kerr frequency comb is... =0.8nm, center wavelength =1550nm. Using a length of... =10km of standard single-mode fiber, its group velocity (Corresponding to a refractive index of approximately 1.465), refractive index difference (Estimated from material dispersion), dispersion coefficient ≈17ps / (nm·km). Path length difference is ignored. (Right now =0). Therefore, according to the full calculation model, the delay step size is: .

[0081] Calculate the second term: Multiply by 1.5e-5 to get .

[0082] The third item: 10 × 17 × 0.8 = 136 ps.

[0083] The total delay step Δτ ≈ 186 ps. This value needs to be equal to the symbol period. Further fine-tuning can be achieved by adjusting the fiber length L or using dispersion compensation devices. To achieve precise matching.

[0084] This full-scale model, compared to the aforementioned simplified formula... More precise, suitable for high-precision requirements or long-distance transmission scenarios.

[0085] As can be seen from the above description, the complex-valued photon convolution calculation system based on coherent modulation provided in this application comprehensively considers group velocity, refractive index difference and dispersion coefficient through a full calculation model, which can more accurately determine the matching relationship between delay step and symbol period, thereby improving the accuracy and robustness of system design.

[0086] To further address the issues of direct detection's inability to recover phase information and its susceptibility to local oscillator noise, this application provides a complex-valued photonic convolution calculation system based on coherent modulation. (See also...) Figure 6 The coherent detection module 50 specifically includes the following components: (1) A second polarization controller 51 is connected to the optical signal transmission module 40 and is used to adjust the polarization state of the delayed optical signal to obtain a polarization-adjusted delayed optical signal; (2) Local oscillator laser 52, used to generate local oscillator light; wherein the frequency of the local oscillator light is the same as the frequency of the continuous light source 11 in the light source module 10; (3) 90° optical mixer 53, connected to the second polarization controller 51 and the local oscillator laser 52, is used to coherently mix the polarization-adjusted delayed optical signal with the local oscillator light and output four mixed optical signals; (4) Balanced receiver 54, connected to the 90° optical mixer 53. The balanced receiver includes two photodetector groups, each consisting of two photodetectors. Each photodetector group performs differential detection on two optical signals with a phase difference of 180° in the four mixed optical signals to convert the four mixed optical signals into two electrical signals.

[0087] The 90° optical mixer is a passive optical device that integrates multiple couplers and phase shifters. It interferes with the input signal light and the local oscillator light, outputting four optical signals with phase relationships of 0°, 90°, 180°, and 270° (or 0°, 180°, 90°, and 270°), representing the addition and subtraction of in-phase and quadrature components. The balanced photodetector (BPD) consists of two matched photodetectors (PDs) and a differential amplifier. It differentially detects two optical signals with a 180° phase difference, effectively suppressing common-mode noise (such as local oscillator intensity noise). The photodetector group comprises the two photodetectors that make up the balanced receiver, requiring highly matched responsivity. Differential detection involves subtracting the photocurrents output from the two photodetectors to obtain a differential current proportional to the signal light field intensity, thereby eliminating common-mode interference and improving the signal-to-noise ratio.

[0088] In one example: Suppose the delayed optical signal output from the optical signal transmission module contains five wavelength channels, each carrying I(t) and Q(t) information after dispersion delay. This signal first enters the second polarization controller within the coherent detection module. This second polarization controller adjusts the polarization state of the signal light to match the polarization state of the local oscillator light output from the local oscillator laser (e.g., both in TE mode). The local oscillator laser is a distributed feedback semiconductor laser, and its output frequency is exactly the same as the continuous light source in the light source module (e.g., both at 193.1 THz, corresponding to a wavelength of 1550 nm), thus achieving zero-difference detection. The polarization-adjusted signal light and the local oscillator light are simultaneously fed into a 90° optical mixer. The 90° optical mixer internally outputs four optical signals through multiple 3dB couplers and a 90° phase shifter: (1) Port 1: Signal light + local oscillator light (in-phase addition, 0° phase difference); (2) Port 2: Signal light - local oscillator light (in-phase subtraction, 180° phase difference); (3) Port 3: Signal light + j·local oscillator light (orthogonal addition, 90° phase difference); (4) Port 4: Signal light - j·Local oscillator light (orthogonal subtraction, 270° phase difference).

[0089] These four optical signals enter two balanced receivers respectively: the first balanced receiver is connected to port 1 and port 2, and its two photodetectors receive these two optical signals respectively, outputting a differential current proportional to... The second balanced receiver connects to ports 3 and 4, and outputs a differential current proportional to... ,in, The carrier phase difference between the signal light and the local oscillator light can be compensated by a DSP. The output of the balanced receiver is two baseband electrical signals, which can be directly sent to a digital signal processor for processing.

[0090] As can be seen from the above description, the complex-valued photonic convolution calculation system based on coherent modulation provided in this application embodiment can perform zero-difference coherent detection by combining a 90° optical mixer with a balanced receiver, which can completely extract the amplitude and phase information of the optical field and suppress common-mode noise, thereby improving detection sensitivity and signal-to-noise ratio.

[0091] To further address issues such as IQ imbalance, clock mismatch, dispersion, polarization mode dispersion, frequency offset, and phase noise in the signal after coherent detection, this application provides a complex-valued photonic convolution calculation system based on coherent modulation. (See also...) Figure 6 The coherent detection module 50 also specifically includes the following components: The second digital signal processor 55 is connected to the balanced receiver 54 and is used to sequentially perform orthogonalization and normalization processing, clock recovery, dispersion compensation, polarization equalization, carrier frequency offset estimation and phase recovery, and time demultiplexing on the two electrical signals output by the balanced receiver, so as to obtain the convolution calculation result of the input matrix and the weight matrix.

[0092] Orthogonalization and normalization are used to compensate for amplitude and phase mismatch between the I and Q signals caused by modulator bias errors, coupler splitting ratio asymmetry, etc. Orthogonalization restores the two signals to 90° orthogonality through linear transformation, while normalization adjusts the signal amplitude to the standard range. Clock recovery extracts the synchronization clock from the received electrical signal to eliminate the frequency and phase differences between the sampling clock of the receiver's analog-to-digital converter (ADC) and the transmitter's clock. Dispersion compensation performs inverse filtering on the static channel damage caused by fiber dispersion (such as different time delays for different wavelength components) in the digital domain to restore the signal waveform. Polarization equalization compensates for polarization mode dispersion (PMD) caused by birefringence in fiber transmission by decoupling the two polarization signals through an adaptive filter. Carrier frequency offset estimation and phase recovery estimate and compensate for the frequency offset between the local oscillator and the transmitter laser, as well as the random phase noise introduced by the laser linewidth, to restore the signal phase to an ideal state. Time-demultiplexing separates a high-speed serial data stream into multiple parallel data streams according to time slice order, and recovers the convolution calculation results corresponding to different wavelength channels.

[0093] In one example, suppose the two electrical signals output by the balanced receiver are respectively and These signals are affected by IQ imbalance, clock skew, dispersion, polarization mode dispersion, frequency offset, and phase noise. The second digital signal processor first performs orthogonalization and normalization processing on the two signals: by calculating the covariance matrix of the two signals, a compensation matrix is ​​constructed to make... and The clock is then restored to orthogonality and uniform amplitude. Next, clock recovery is performed: the Gardner or Mueller-Muller algorithm is used to extract the clock error from the signal, and the sampling phase of the ADC is adjusted to align the sampling time with the symbol center. Dispersion compensation is then performed: a frequency domain equalizer (FDE) is used to filter the signal based on the inverse function of the fiber dispersion transfer function, eliminating the time delay difference between different wavelength components. Then, polarization equalization is performed: a 2×2 adaptive butterfly finite impulse response (FIR) filter (such as the constant mode algorithm CMA or the decision-guided LMS algorithm DD-LMS) is used to separate and compensate for crosstalk between the two polarization states. Finally, carrier frequency offset estimation and phase recovery are performed: first, the residual frequency offset is estimated through fourth power calculation or pilot-based methods, and frequency rotation is performed in the digital domain; then, the Viterbi-Viterbi algorithm or blind phase search (BPS) algorithm is used to estimate and compensate for the phase noise caused by the laser linewidth. Finally, time decomposition and multiplexing are performed: the high-speed serial data stream processed above is sequentially allocated to different output ports according to the order of each symbol period, and the parallel convolution results corresponding to each wavelength channel are recovered, that is, the convolution calculation result of the input matrix and the weight matrix is ​​obtained.

[0094] As can be seen from the above description, the complex-valued photonic convolution calculation system based on coherent modulation provided in this application can eliminate the influence of various non-ideal factors on the calculation results by performing a series of digital compensation and recovery operations through a second digital signal processor, thereby improving the accuracy of convolution calculation and the robustness of the system.

[0095] To further address the issues of high computational load and low energy efficiency when traditional electronic processors perform convolution operations, this application provides a complex-valued photonic convolution calculation system based on coherent modulation. The system is used for convolutional layer calculation in an opto-neural network, wherein the number of different wavelength channels corresponds to the number of weights in the convolution kernel, the input matrix is ​​the input feature map of the opto-neural network, and the weight matrix is ​​the convolution kernel weights of the opto-neural network. The system performs a sliding window multiplication and accumulation operation on the input feature map and the convolution kernel weights through time-delayed interleaving of multiple wavelength channels to obtain the convolution calculation result of the input matrix and the weight matrix as the output of the convolution layer of the photoelectric neural network.

[0096] It should be noted that the input feature map is the input data of the convolutional layer in the photoelectric neural network, typically the output of the previous layer or the original input image, and has three dimensions: height, width, and number of channels. In this system, the input feature map is flattened into a one-dimensional vector and then converted into two time-varying electrical signals. The convolution kernel weights are convolution kernel parameters learned in the photoelectric neural network and are used to perform convolution operations with the input feature map. In this system, each weight value is mapped to an independent wavelength channel. The sliding window multiplication-accumulation operation is the core operation of the convolution operation. That is, the convolution kernel slides on the input feature map with a fixed stride, and at each window position, the convolution kernel weight is multiplied point by point with the corresponding input value and then summed to obtain a pixel value of the output feature map. The convolutional layer output is the output feature map obtained after the convolution operation, which can be used as the input of the next layer or the final output of the entire network.

[0097] In one example, suppose a convolutional layer of a photoelectric neural network has the following parameters: input feature map size of 32×32 (single channel), convolution kernel size of 3×3, number of kernel weights of 9, number of output channels of 1, and stride of 1. In this system, the light source module generates 9 wavelength channels (corresponding to 9 convolution kernel weights), and the waveform shaper loads the weight values ​​[0.1, 0.3, 0.5, 0.7, 0.9, 0.2, 0.4, 0.6, 0.8] onto the 9 wavelength channels respectively. The signal processing module flattens the input feature map (32×32=1024 pixels) into a one-dimensional vector and splits it into two time-varying electrical signals I(t) and Q(t), outputting one pixel value sequentially within each symbol period. The electro-optic modulation module and the optical signal transmission module stagger the same pixel data of the 9 wavelength channels by one symbol period in time (the delay step is equal to the symbol period). After processing by the coherent detection module and DSP, within each symbol period, the output of the balanced receiver corresponds to the weighted sum of 9 weights and 9 input pixels within the current window, which is one pixel of the convolution result. Since the sliding window stride is 1, the system continuously outputs 32×32 convolution results within 1024 symbol periods, forming an output feature map. This output feature map is the output of the convolutional layer and can be directly fed into the next layer of the optoelectronic neural network for processing.

[0098] As can be seen from the above description, the complex-valued photonic convolution computing system based on coherent modulation provided in this application embodiment, when applied to the convolutional layer of an optoelectronic neural network, can significantly improve the speed and energy efficiency of convolution operations by utilizing the parallelism of multiple wavelength channels and the multiply-accumulate operation in the optical domain.

[0099] In other words, the complex-valued photonic convolution calculation system based on coherent modulation provided in this application can solve the following problems existing in the mainstream electro-optic modulation schemes in the prior art: (1) Phase modulation based on phase-intensity conversion of Mach-Zehnder Interferometer (MZI) combined with direct detection: Since the information is hidden in the phase of the optical carrier, the ordinary photodetector (PD) at the end of the link only responds to the optical power and cannot directly "read" the phase change, resulting in the complete loss of signal information after direct detection. Theoretically, coherent digital demodulation can indeed demodulate the complete phase modulation information linearly, but this requires the introduction of a highly stable local oscillator laser and a complex optical mixer and high-speed digital signal processing unit, which leads to a sharp increase in receiver cost, power consumption and complexity. Its core PM-IM conversion device—MZI delay interferometer—is essentially a high-precision passive optical device that interferes with the phase of the current symbol with the phase of the preceding symbol delayed by one symbol period, thereby mapping the phase difference to the intensity of optical power. The fatal weakness of this mechanism is its extreme operating point sensitivity. The photodiode used in direct detection is a square-law detection device that only responds to the optical power and completely loses the phase information of the optical field. When signal light carrying severe dispersion (CD) damage undergoes long-distance fiber optic transmission, it passes through a nonlinear optical device called MZI (conversion from field to power) and is detected by PD. The dispersion distortion, which was originally in the linear optical domain, interacts with the square-law detection process and is "alienated" in the electrical domain into a highly complex and difficult-to-handle nonlinear distortion.

[0100] (2) The intensity modulation scheme uses MZM or electroabsorption modulator to directly load the electrical signal onto the intensity of the optical carrier and adopts direct detection reception (IM / DD), which has the advantages of simple structure and low cost. However, since direct detection also loses the phase information of the optical field, it leads to low system spectral efficiency, limited transmission rate, and susceptibility to power selective fading. At the same time, it cannot adopt advanced digital post-compensation technology.

[0101] The complex-valued photonic convolution computation system based on coherent modulation provided in this application abandons the traditional MZI phase-intensity conversion plus direct detection scheme. It adopts IQ complex-valued modulation and zero-difference coherent detection, which can recover the amplitude and phase information of the optical field without loss without the need for a highly stable and complex local oscillator laser and a complex demodulation unit. This reduces the cost, power consumption and complexity of the receiver, and completely avoids the loss of phase information. IQ complex plane dual-dimensional modulation is used to replace the traditional single-dimensional amplitude / phase modulation, which doubles the data carrying capacity per unit bandwidth and improves spectral efficiency and computation speed. At the same time, relying on the high sensitivity of coherent detection, the system's resistance to power selective fading and noise is enhanced. By integrating Kerr optical frequency comb multi-wavelength channels, time-division multiplexing, IQ parallel modulation and dispersive delay sliding accumulation technology, the hardware parallel execution of complex-valued convolution operations is realized, improving the system's parallel processing degree and data throughput, and adapting to the large-scale, high-speed computation requirements of optoelectronic neural networks.

[0102] To further illustrate the above embodiments, this application also provides a specific application example of a complex-valued photonic convolution computation system based on coherent modulation, see [link to relevant documentation]. Figure 7 The complex-valued photonic convolution computation system based on coherent modulation provided in this application example includes: a light source module 10, a signal processing module 20, an electro-optic modulation module 30, an optical signal transmission module 40, and a coherent detection module 50. The system mainly consists of a hardware optical path part (light source, modulation, and transmission modules) and a software processing part (digital signal processing algorithms in a digital signal processor).

[0103] The output of the light source module 10 is an optical carrier signal loaded with a weight matrix. In this embodiment, the light source module 101 first generates a Kerr optical frequency comb. By inputting a single-wavelength continuous light source and passing through devices such as a micro-ring resonator, a Kerr optical frequency comb is generated. As an optical carrier, the Kerr optical frequency comb can provide a large number of dense wavelength channels, offering advantages such as wide spectrum, high integration capability, small size, and low cost. The generated Kerr optical frequency comb, after passing through a waveform shaper, can be used to map the weights onto the optical carrier by changing the power of each optical comb, thus generating an optical carrier signal loaded with a weight matrix.

[0104] The signal processing module 20 is used to convert the input matrix into a one-dimensional vector, and then map the vector into a time-varying electrical signal waveform through a digital-to-analog converter, which is then divided into two parts and input into two MZMs respectively.

[0105] The electro-optic modulation module 30 has two input interfaces: an optical input interface for receiving the optical carrier signal loaded with a weight matrix from the light source module, and an electrical input interface for receiving electrical signals. The electrical signal is a time-division multiplexed mapping of the input matrix, specifically converting the input matrix into a one-dimensional vector. This vector is then mapped to a time-varying electrical signal waveform via a digital-to-analog converter. Finally, the time-varying electrical waveform is modulated onto the optical carrier signal using electro-optic modulation before being output. The module's output is an optical carrier signal loaded with both the input electrical signal and the weight matrix.

[0106] The optical signal transmission module 40 is used for optical dispersion compensation of module 102. It transmits the optical carrier signal loaded with the input electrical signal and weight matrix through a coefficient-tunable dispersion fiber module, and then sends it to the photoelectric detection module. The input of this module is the optical carrier signal loaded with the input electrical signal and weight matrix. The output of this module is the delayed optical carrier signal loaded with the input electrical signal and weight matrix.

[0107] The coherent detection module 50 is used to demodulate the signal. It utilizes local oscillator light to perform zero-difference coherent detection on the signal to be detected. The resulting signal is then processed by digital signal processing to separate the time-varying electrical signal waveform converted from the input matrix and the convolution result with the weights. The input to this module is a delayed optical carrier signal loaded with the input electrical signal and weight matrix. The output of this module is the convolution result of the time-varying electrical signal converted from the input matrix and the weights.

[0108] Reference Figure 2 The light source module 10 includes: a continuous light source 11, an EDFA amplifier 12, a first polarization controller 13, a micro-ring resonator 14, and a waveform shaper 15.

[0109] A continuous light source 11 provides a highly stable, narrow-linewidth continuous single-wavelength laser beam. An EDFA amplifier 12 amplifies the power of the pump laser to a level sufficient to drive the nonlinear processes within the microring resonator (the nonlinear effects required to generate the Kerr frequency comb typically require a high optical power threshold). A first polarization controller 13 finely adjusts the polarization state of the pump light before it enters the microring resonator 14, ensuring a perfect match with the resonant mode of the microring resonator 14, thereby achieving maximum intracavity optical field enhancement and the most efficient nonlinear conversion. Specifically, the first polarization controller 13 utilizes the principle of stress birefringence, like a tunable waveplate, to finely rotate the polarization state of the incident pump light, making its vibration direction completely parallel to the resonant mode of the microring resonator 14. This eliminates reflections and losses caused by polarization misalignment, ensuring that the pump energy is coupled into the microring resonator 14 most efficiently and forms resonant accumulation, achieving maximum intracavity optical field enhancement, and thus exciting the strongest nonlinear optical effects such as the Kerr frequency comb.

[0110] The micro-ring resonator 14 is the core component for generating the optical frequency comb. The pump light undergoes a four-wave mixing effect within it, ultimately producing the Kerr optical frequency comb. As a multi-wavelength optical carrier signal, the Kerr optical frequency comb can provide a large number of dense wavelength channels, and has the advantages of wide spectrum, high integration capability, small size, and low cost.

[0111] The waveform shaper 15 is used to adjust the optical power of each Kerr optical frequency comb tooth to achieve the mapping of the weight matrix. Each comb tooth represents a discrete, equally spaced wavelength channel, with different wavelengths at equal intervals. Specifically, the weight matrix is ​​expanded into a one-dimensional weight vector in a specific manner. The waveform shaper 15 independently adjusts the optical power of each Kerr optical frequency comb tooth, mapping each weight value in the weight matrix to a different comb tooth, thus achieving a one-to-one mapping between single wavelengths and single weights. Consequently, the light source module 10 outputs a multi-wavelength optical carrier signal loaded with the weight matrix.

[0112] It should be noted that this application example employs a preloading mechanism of single-wavelength-single-weight one-to-one mapping: a weight value only requires one optical frequency comb tooth, and the weight mapping can be completed by directly adjusting the optical power of a single comb tooth through a waveform shaper. The real and imaginary parts of the weight are combined through the combination of orthogonal components of the IQ modulated optical field, without needing to be split into different wavelength channels.

[0113] Reference Figure 3 The signal processing module 20 includes a first digital signal processor 21 and an arbitrary waveform generator 22. The first digital signal processor 21 configures the waveform parameters of the arbitrary waveform generator 22. The arbitrary waveform generator 22 is connected to the first digital signal processor 21 and, under the control of the first digital signal processor 21, flattens the input matrix into a one-dimensional vector, maps the one-dimensional vector into a time-varying electrical signal waveform through digital-to-analog conversion, and evenly splits the time-varying electrical signal waveform into a first time-varying electrical signal and a second time-varying electrical signal. Neither the first nor the second time-varying electrical signal undergoes electrical domain Hilbert transform or carrier coding preprocessing.

[0114] This application employs a standardized IQ modulation input mapping: after flattening the input matrix, it is directly split into a first time-varying electrical signal and a second time-varying electrical signal, which are then fed into two parallel Mach-Zehnder modulators for linear modulation. The second signal introduces a fixed 90° phase difference, and after beam combining, a complex optical field is directly generated. This process requires no preprocessing steps such as electrical domain Hilbert transform or carrier coding.

[0115] Reference Figure 4 The electro-optic modulation module 30 includes: a one-to-two optical splitter 31, a first Mach-Zehnder modulator 32 (MZM1), and a second Mach-Zehnder modulator 33 (MZM2). The one-to-two optical splitter 31 divides the input optical signal equally in terms of power and sends it to two parallel Mach-Zehnder modulators respectively. The first Mach-Zehnder modulator 32 modulates the first time-varying electrical signal (I-channel), and the second Mach-Zehnder modulator 33 modulates the second time-varying electrical signal (Q-channel) and adds a 90° phase difference to achieve orthogonal amplitude and phase multiplexing.

[0116] This module has two inputs: an optical input interface receives a multi-wavelength optical carrier signal with a weighted matrix output from the light source module 10; and an electrical input interface receives a first time-varying electrical signal and a second time-varying electrical signal output from the signal processing module 20. The input signal is generated by flattening the input matrix into a one-dimensional input data vector in a specific manner, and then converting it into a time-varying electrical signal waveform via a digital-to-analog converter. Each data symbol has a predetermined, precisely controlled symbol period τ. Since the IQ modulator has two parallel Mach-Zehnder modulators, two parallel modulations can be implemented simultaneously, achieving time-division multiplexing. The parallel operation of the first and second paths decomposes the convolution operation of a large matrix into the parallel convolution of two sub-vectors.

[0117] Assuming the input matrix is ​​divided into four equal parts, after flattening and digital-to-analog conversion, it is quantized in parallel into two time-varying electrical signals with equal symbol periods and frequencies, namely the first time-varying electrical signal I(t) and the second time-varying electrical signal Q(t). The input optical signal is equally divided into in-phase and quadrature paths, propagating along different paths with a 90° relative phase difference between the two optical signals. The Mach-Zehnder modulators of the in-phase and quadrature paths operate in push-pull mode, with the DC bias set at the quadrature bias point (i.e., the lowest point of the power transfer function). Finally, the two signals are interferometrically combined at the output coupler 34 into a single complex-valued optical field signal.

[0118] Let the input single optical comb be... The electrical signal I(t) is modulated by a modulator and applied to the optical signal. The output can be expressed as: Under small-signal conditions, the output electric field is approximately proportional to... , can be further expressed as (Small signal approximation); similarly, the other output is... The modulated signals are eventually combined into a single useful signal: Reference Figure 5 The optical signal transmission module 40 includes a dispersion delay unit 41 and a third polarization controller 42. The dispersion delay unit 41 is used to perform dispersion delay on the complex-valued optical field signal, creating a fixed delay difference between different wavelength channels. In other words, the dispersion delay section is used to compensate for dispersion, transmitting the optical carrier signal loaded with the input signal and weight matrix through a coefficient-tunable dispersion fiber module. Since different wavelengths of light propagate at different speeds in the dispersion waveguide medium, the propagation delay difference and wavelength... Fiber length L, dispersion coefficient D, refractive index difference Group speed The following conditions must be met: in This represents the difference in path length between different wavelength channels. This is due to the difference in refractive index. After transmission through a dispersive medium, the optical signals carried by light waves of different wavelengths experience time delays, interleaving, and misalignment. By tuning the dispersion module, the delay step size between adjacent wavelength channels can be adjusted. Equal to a single symbol period of the input signal That is, satisfying: Due to the effect of the dispersive waveguide medium, the beat frequency results of multiple wavelength channels will exhibit sequential delay and misalignment between symbols, facilitating subsequent accumulation operations. Figure 8 For example, h1, h2, h3, and h4 are the four weights of a 2×2 weight matrix, which are sequentially expanded into a one-dimensional weight vector in a specific manner, resulting in a misalignment after a delay. a1 to a9 are the input data sequence, where a1 to a9 are the elements (e.g., image pixel values ​​or feature map values) of the one-dimensional vector obtained after flattening the input matrix, input sequentially. y1 to y12 are the output sequence of the convolution operation, where y1 to y12 correspond to the multiplication and accumulation results at a sliding window position for each output. T is the time axis, representing the temporal order of the signal. Figure 8 The physical process of time division multiplexing and dispersion delay is demonstrated: the input data a1 and a2 are serial, but because different wavelengths have different speeds after passing through the dispersion fiber, the a1 carrying h1 and the a1 carrying h2 are misaligned in time.

[0119] During the transmission of optical signals in the optical signal transmission module 40, the polarization state of the light undergoes random rotation and distortion due to the physical bending or stress of the optical fiber. The third polarization controller 42 is used to compensate for polarization drift during transmission and avoid crosstalk. Its core function is to actively regulate the polarization state of the received optical signal to precisely match the polarization state of the local oscillator (LO), thereby maximizing coherent mixing efficiency, eliminating polarization-dependent losses and noise, and ensuring the demodulation performance of the system.

[0120] It should be noted that this application example, for the precise stepping requirements of the convolution sliding window, presents a full quantitative design system for dispersion delay: wavelength interval The free spectral range (FSR) of the Kerr frequency comb in the light source module is strictly bound to ensure that each wavelength channel corresponds to a weight value, matching the weighting process one by one; symbol period IQ modulation baud rate with electro-optic modulation module Strict synchronization is ensured to guarantee that the delay step size is perfectly aligned with the symbol period of the input signal, achieving unbiased stepping of the convolution sliding window. The dispersion coefficient D and fiber length L are tunable, allowing for flexible adjustment based on the convolution kernel size M (number of wavelength channels), thus enabling continuous adjustment of the convolution window length. A comprehensive calculation model for the delay difference of different wavelength channels is defined, taking group velocity into account in the equation. The resulting transmission delay difference and refractive index difference The resulting phase delay difference and dispersion coefficient The resulting wavelength-dependent delay difference is explained, and a precise matching formula for the delay step size of adjacent wavelength channels and a full-link collaborative constraint are given.

[0121] Reference Figure 6 The coherent detection module 50 includes: a second polarization controller 51, a local oscillator laser 52, a 90° optical mixer 53, and a balanced receiver 54.

[0122] The second polarization controller 51 is connected to the optical signal transmission module 40 and is used to adjust the polarization state of the delayed optical signal to obtain a polarization-adjusted delayed optical signal. The local oscillator laser 52 is used to generate local oscillator light, the frequency of which is the same as the frequency of the continuous light source 11 in the light source module 10 (zero-difference coherent detection). The 90° optical mixer 53 is connected to the second polarization controller 51 and the local oscillator laser 52 and is used to coherently mix the polarization-adjusted delayed optical signal with the local oscillator light, outputting four mixed optical signals. The balanced receiver 54 is connected to the 90° optical mixer 53. The balanced receiver 54 includes two photodetector groups, each consisting of two photodetectors. Each photodetector group differentially detects two optical signals with a phase difference of 180° in the four mixed optical signals to convert the four mixed optical signals into two electrical signals.

[0123] In other words, the local oscillator laser is responsible for generating local oscillator light with the same frequency as the laser source, also called the LO signal. Zero-difference coherent detection ensures that the frequency of the LO signal is exactly the same as the frequency of the received signal; after balanced detection, the signal becomes the baseband signal. A 90° optical mixer receives the matched polarization signal. and Through an optical phase-shifting and coupling structure, coherent mixing with a 90° phase difference is achieved, decomposing the complex amplitude of the optical signal (including amplitude and phase information of I / Q modulation) into four optical signals: corresponding to the mixing results of "signal light I component - local oscillator in phase", "signal light I component - local oscillator orthogonal", "signal light Q component - local oscillator in phase", and "signal light Q component - local oscillator orthogonal". The balanced receiver's function is to suppress common-mode interference such as local oscillator intensity noise and ambient light, while improving receiving sensitivity. Each balanced receiver consists of two photodetectors (PDs) that perform differential photoelectric conversion on the output optical signal of the 90° mixer. During this process, the local oscillation signal... The signal is coherently mixed with the received optical signal S. This process can be expressed by the following formula: The four optical signals output from the mixer are received by a balanced receiver 54 and converted into two electrical signals. In the digital signal processor, the original I(t) and Q(t) components can be perfectly demodulated through frequency difference estimation and carrier phase recovery. Finally, the demodulated high-speed serial data stream is time-demultiplexed to recover the parallel computation results, completing the output of the entire photonic convolutional network. The output of the balanced receiver satisfies: Finally, the calculation results are converted by photoelectric conversion and digital reconstruction to restore the output corresponding to different time slices, thus completing the inference task of the entire neural network.

[0124] It should be noted that this application example employs a zero-difference coherent detection and DSP-coordinated demodulation scheme: using a local oscillator of the same frequency, a 90° optical mixer, and a balanced receiver, it fully captures the amplitude and phase information of the optical field across all dimensions. The optical domain complex-valued convolution result is directly down-converted to baseband, and then the convolution results of the first and second time-varying electrical signals are losslessly recovered using a DSP algorithm. Specifically: Firstly, during the process of achieving coherent optical reception using a 90° optical mixer and a balanced receiver, amplitude and phase mismatch (i.e., IQ imbalance) between the I and Q signals may occur due to factors such as errors in the setting of the bias points of the I and Q paths, asymmetry in the splitting ratio of the 3dB coupler inside the mixer, or mismatch in the responsivity of the photodetector. Therefore, the DSP processing flow first performs orthogonalization and normalization on the received signal to compensate for the aforementioned imbalance.

[0125] Secondly, the two orthogonal continuous electrical signals of each polarized light path need to be sampled and quantized by an analog-to-digital converter (ADC) before being input to the DSP processing unit. Since the ADC sampling clock is usually provided by a local clock, which is independent of the transmitting clock, and the clock oscillator itself may have imperfect characteristics, there are significant differences in frequency and phase between the transmitting and receiving clocks. Therefore, a clock recovery module is introduced into the DSP processing flow to eliminate clock misalignment caused by ADC sampling clock mismatch.

[0126] Furthermore, based on the transmission function of optical fiber, even ignoring fiber nonlinear effects, the static channel impairment caused by fiber dispersion and the polarization mode dispersion effect will affect the received signal. Therefore, a dispersion compensation module and a polarization equalization module are introduced into the DSP processing flow to eliminate these effects.

[0127] Finally, due to various factors such as the fabrication process of optical devices, it is difficult to maintain perfect frequency synchronization between the local oscillator and the transmitting laser. Simultaneously, the laser linewidth introduces a corresponding phase shift, resulting in a significant amount of additional phase noise in the coherently detected signal. Therefore, the carrier recovery module in the DSP processing flow first compensates for the frequency offset using a carrier frequency offset estimation algorithm, and then further corrects the phase offset caused by the local oscillator linewidth and residual frequency offset using a carrier phase recovery algorithm, thereby achieving correct judgment and demodulation of the received signal. Figure 9 As shown, the I and Q signals of the two polarization states (X-polarization and Y-polarization) are orthogonalized and normalized respectively to compensate for IQ imbalance; then clock recovery and dispersion compensation are performed sequentially; finally, they are passed through a polarization equalizer (coefficient). The polarization mode dispersion is decoupled; finally, after frequency difference estimation and phase recovery, the decision result is output. This process completely realizes the digital domain reconstruction from the electrical signal output by the balanced receiver to the convolution calculation result. and These represent the in-phase component and quadrature component in the X-polarization direction, respectively (i.e., the corresponding signals of the first and second time-varying electrical signals in the X-polarization state). and represents the in-phase component and the quadrature component in the Y polarization direction, respectively; j is the imaginary unit.

[0128] In other words, this application presents an application example of a multidimensional multiplexed optoelectronic convolutional computing architecture based on IQ modulation. It utilizes an IQ modulator as the core computational unit and constructs a parallel mapping mechanism. By loading a weight matrix onto different wavelengths of a Kerr optical frequency comb and resolving the input matrix into two parallel time-varying electrical signals, the time-division multiplexed real-valued electrical signals are converted into a complex optical field. This not only doubles the data capacity per unit symbol but, more importantly, introduces a new degree of freedom for phase control, thereby constructing a high-parallel processing and high-data-throughput optoelectronic neural network. Through coherent reception and advanced digital signal processing techniques, the common problems of accuracy loss and error accumulation in optical computing systems are solved. Based on the provided mathematical model, using a 90° optical mixer and balanced probe, the original I and Q components can be recovered losslessly in the electrical domain, thus achieving accurate demodulation of the computation results.

[0129] Based on this, the system provided in the application example of this application has the following beneficial effects: 1. Traditional modulation is mostly "single-dimensional" (using only one dimension—amplitude, frequency, or phase—to carry information), while IQ modulation is "complex-plane two-dimensional" (using both amplitude and phase to carry information simultaneously). This offers significant advantages; under the same bandwidth, the data rate can be doubled or even multiplied, perfectly meeting the bandwidth requirements of high-speed optical communication. Furthermore, traditional amplitude modulation (such as OOK) is susceptible to light intensity fluctuations, while IQ modulation, combined with phase information, results in a lower bit error rate than simple amplitude / frequency modulation at the same received power, and exhibits superior resistance to channel fading and noise.

[0130] 2. Direct detection can only extract light intensity and cannot obtain phase information, therefore it can only demodulate simple formats such as OOK and ASK; while coherent detection can capture both amplitude and phase simultaneously, perfectly adapting to higher-order formats such as IQ modulation QPSK and 16QAM. Through the "optical domain amplification" effect of the local oscillator, coherent detection can receive weaker signals (sensitivity is 10-20 dB higher than direct detection), extending the optical signal transmission distance from tens of kilometers in direct detection to hundreds of kilometers.

[0131] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application. When implemented in hardware, it can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this application are programs or code segments used to perform the required tasks. The programs or code segments can be stored on a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried on a carrier wave.

[0132] It should be clarified that this application is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of this application is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of this application.

[0133] In this application, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or in place of features of other embodiments.

[0134] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to the embodiments of this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A complex-valued photonic convolutional computing system based on coherent modulation, characterized in that, include: The light source module is used to obtain a multi-wavelength optical carrier signal loaded with a weight matrix based on a continuous light source; The signal processing module is used to convert the input matrix into a first time-varying electrical signal and a second time-varying electrical signal; An electro-optic modulation module, connected to the light source module and the signal processing module respectively, is used to divide the multi-wavelength optical carrier signal loaded with a weight matrix into two optical signals. One optical signal is directly modulated by the first time-varying electrical signal, and the other optical signal is modulated by the second time-varying electrical signal after adding a 90° phase difference. The two modulated optical signals are then combined to obtain a complex optical field signal. An optical signal transmission module, connected to the electro-optic modulation module, is used to perform dispersive delay on the complex-valued optical field signal so that a fixed delay difference is generated between different wavelength channels of the complex-valued optical field signal to obtain a delayed optical signal. The coherent detection module, connected to the optical signal transmission module, is used to perform zero-difference coherent detection on the delayed optical signal to recover the amplitude and phase information of the delayed optical signal, and to obtain the convolution calculation result of the input matrix and the weight matrix.

2. The coherent-modulation-based complex-valued photonic convolutional computing system of claim 1, wherein, In the optical signal transmission module, the fixed delay difference is the delay step size between adjacent wavelength channels of the complex optical field signal; The delay step is set to be equal to the symbol period of the first time-varying electrical signal and the second time-varying electrical signal, and satisfies the following quantitative relationship: in, This indicates the delay step size; Indicates the length of the tunable dispersive fiber; Represents the tunable dispersion coefficient; Indicates the wavelength spacing between adjacent wavelength channels; This indicates the symbol period.

3. The complex-valued photonic convolution computation system based on coherent modulation according to claim 1, characterized in that, The light source module includes: a continuous light source, an EDFA amplifier, a first polarization controller, a micro-ring resonator, and a waveform shaper connected in sequence. The single-wavelength laser output from the continuous light source is injected into the micro-ring resonant cavity after passing through the EDFA amplifier and the first polarization controller. A Kerr optical frequency comb is generated through the four-wave mixing effect to serve as a multi-wavelength optical carrier signal. The waveform shaper independently controls the optical power of each tooth of the Kerr optical frequency comb, and maps each weight value in the weight matrix to a single wavelength channel to obtain a multi-wavelength optical carrier signal loaded with the weight matrix.

4. The complex-valued photonic convolution computation system based on coherent modulation according to claim 1, characterized in that, The first and second time-varying electrical signals generated by the signal processing module are obtained by directly splitting the input matrix after flattening it.

5. The complex-valued photonic convolution computation system based on coherent modulation according to claim 4, characterized in that, The signal processing module includes: First digital signal processor; An arbitrary waveform generator, connected to the first digital signal processor, is used to flatten the input matrix into a one-dimensional vector under the control of the digital signal processor, map the one-dimensional vector into a time-varying electrical signal waveform through digital-to-analog conversion, and evenly split the time-varying electrical signal waveform into a first time-varying electrical signal and a second time-varying electrical signal.

6. The complex-valued photonic convolution computation system based on coherent modulation according to claim 1, characterized in that, The electro-optic modulation module includes: A one-to-two optical splitter is connected to the light source module and is used to receive the multi-wavelength optical carrier signal loaded with a weight matrix and to divide the multi-wavelength optical carrier signal loaded with a weight matrix into two optical signals. The first Mach-Zehnder modulator is connected to the one-to-two optical splitter and the signal processing module, and the first Mach-Zehnder modulator receives the first time-varying electrical signal and is driven by the first time-varying electrical signal to modulate one optical signal output by the one-to-two optical splitter. The second Mach-Zehnder modulator is connected to the one-to-two optical splitter and the signal processing module respectively. The second Mach-Zehnder modulator receives the second time-varying electrical signal and is driven by the second time-varying electrical signal to modulate the other optical signal output by the one-to-two optical splitter by adding a 90° phase difference. An output coupler is connected to the first Mach-Zehnder modulator and the second Mach-Zehnder modulator respectively, and is used to combine the modulated optical signals output by the first Mach-Zehnder modulator and the second Mach-Zehnder modulator respectively to obtain a complex-valued optical field signal. Both the first and second Mach-Zehnder modulators operate in push-pull mode and have their DC bias set at the quadrature bias point.

7. The complex-valued photonic convolution computation system based on coherent modulation according to claim 1, characterized in that, The coherent detection module includes: The second polarization controller, connected to the optical signal transmission module, is used to adjust the polarization state of the delayed optical signal to obtain a polarization-adjusted delayed optical signal. A local oscillator laser is used to generate local oscillator light; wherein the frequency of the local oscillator light is the same as the frequency of the continuous light source in the light source module; A 90° optical mixer, connected to the second polarization controller and the local oscillator laser, is used to coherently mix the polarization-adjusted delayed optical signal with the local oscillator light and output four mixed optical signals. A balanced receiver is connected to the 90° optical mixer. The balanced receiver includes two photodetector groups, each consisting of two photodetectors. Each photodetector group differentially detects two optical signals with a phase difference of 180° in the four mixed optical signals to convert the four mixed optical signals into two electrical signals.

8. The complex-valued photonic convolution computation system based on coherent modulation according to claim 7, characterized in that, The coherent detection module also includes: The second digital signal processor, connected to the balanced receiver, is used to sequentially perform orthogonalization and normalization, clock recovery, dispersion compensation, polarization equalization, carrier frequency offset estimation and phase recovery, and time demultiplexing on the two electrical signals output by the balanced receiver, so as to obtain the convolution calculation result of the input matrix and the weight matrix.

9. The complex-valued photonic convolution computation system based on coherent modulation according to claim 2, characterized in that, The dispersion delay in the optical signal transmission module also satisfies the following full calculation model: in, This indicates the delay step size, which is the length of the dispersive fiber. and wavelength The function; This represents the difference in path length between different wavelength channels; Indicates group velocity; Represents the speed of light; Indicates the difference in refractive index; Indicates the dispersion coefficient; This indicates the wavelength interval between adjacent wavelength channels.

10. The complex-valued photonic convolution computation system based on coherent modulation according to any one of claims 1 to 9, characterized in that, The system is used for convolutional layer calculation in an opto-neural network, wherein the number of different wavelength channels corresponds to the number of weights in the convolutional kernel, the input matrix is ​​the input feature map of the opto-neural network, and the weight matrix is ​​the convolutional kernel weights of the opto-neural network. The system performs a sliding window multiplication and accumulation operation on the input feature map and the convolution kernel weights through time-delayed interleaving of multiple wavelength channels to obtain the convolution calculation result of the input matrix and the weight matrix as the output of the convolution layer of the photoelectric neural network.

Images