Optical signal compensator, optical receiver, optical module, and optical communication system

Abstract

Provided are an optical signal compensator, an optical receiver, an optical module, and an optical communication system. The optical signal compensator comprises an optical signal compensation neural network, wherein the optical signal compensation neural network is deployed on a silicon-based photonic signal processing chip and used for performing compensation and recovery on a received distorted optical signal. The optical signal compensation neural network comprises a reservoir computing neural network layer and a reservoir computing readout layer, wherein the reservoir computing neural network layer is connected to an output end of an optical fiber and used for receiving a distorted optical signal transmitted by the optical fiber and processing the distorted optical signal to obtain an output optical signal sequence; and the reservoir computing readout layer is connected to the reservoir computing neural network layer and used for processing the output optical signal sequence to obtain the optical signal having undergone compensation and recovery. The optical signal compensator provided by the present application directly performs compensation and recovery on the distorted signal in an optical domain, thereby avoiding high power consumption and high latency caused by electrical domain compensation after photoelectric conversion and digital-to-analog conversion.

Description

An optical signal compensator, optical receiver, optical module and optical communication system

[0001] The present application claims priority to the Chinese patent application No. 202410918496.4, filed on July 5, 2024, entitled "An optical signal compensator, optical receiver, optical module and optical communication system", the whole content of which is incorporated herein by reference. TECHNICAL FIELD

[0002] The present application relates to the field of optical communication technology, and in particular to an optical signal compensator, optical receiver, optical module and optical communication system. BACKGROUND

[0003] With the development of communication technology, people's demand and requirement for information transmission are constantly improving. Optical communication technology has developed rapidly in recent years and achieved remarkable achievements and applications because of its high transmission efficiency, convenient hardware implementation and resistance to external interference. Optical signals are transmitted through optical fibers, and the light waves transmitted in the optical fibers have a certain spectral width, i.e., the light waves include many different frequency components. Different frequency components of the light waves propagate at different speeds in the optical fibers, and after a certain distance, the color dispersion phenomenon occurs, which leads to signal distortion. Signal distortion can cause an increase in bit error rate and a decrease in signal-to-noise ratio, which seriously affects the quality of optical communication. Therefore, it is necessary to compensate and recover the distorted signal.

[0004] However, the existing signal compensation and recovery is usually in the electrical domain, and digital signal processing (DSP) is used to realize the compensation and recovery of optical signals, which causes high power consumption of the electrical chip in the optical module and high delay of optical signal transmission. SUMMARY

[0005] Embodiments of the present application provide an optical signal compensator, optical receiver, optical module and optical communication system. The optical signal compensator deploys an optical signal compensation neural network on a silicon-based photonic signal processing chip to compensate and recover the distorted signal in the optical domain, thereby avoiding high power consumption and high delay caused by compensation in the electrical domain after photoelectric conversion and digital-to-analog conversion.

[0006] In a first aspect, the present application provides an optical signal compensator, comprising an optical signal compensation neural network deployed on a silicon-based photonic signal processing chip, for compensating and recovering a received distorted optical signal; the optical signal compensation neural network comprises a reservoir computing neural network layer and a reservoir computing readout layer, wherein the reservoir computing neural network layer is connected to an output end of an optical fiber, for receiving a distorted optical signal transmitted by the optical fiber and processing the distorted optical signal to obtain an output optical signal sequence; the reservoir computing readout layer is connected to the reservoir computing neural network layer (for example, the output end of the reservoir computing neural network layer is connected to the input end of the reservoir computing readout layer through a waveguide), for processing the output optical signal sequence to obtain a compensated and recovered optical signal.

[0007] The optical signal compensator provided by the present application realizes compensation and recovery of a distorted signal directly in the optical domain through the optical signal compensation neural network disposed on the silicon-based photonic signal processing chip (which can be referred to as a silicon optical chip), thereby avoiding high power consumption and high delay caused by compensation in the electrical domain after photoelectric conversion and digital-to-analog conversion.

[0008] In one possible implementation, the reservoir computing neural network layer comprises N reservoir layers and an output layer (at this time, the reservoir computing neural network layer can be referred to as a cascaded reservoir computing neural network layer), the output layer is connected to the N reservoir layers respectively, for performing weighted processing on the output optical signals of the respective reservoir layers in the N reservoir layers to obtain the output optical signal sequence, N is a positive integer greater than 1; the reservoir computing readout layer comprises N optical processing circuits, for separating the output optical signal sequence into the N optical processing circuits in the time dimension; and a readout weight layer is configured to perform weighted processing on the optical signals output by the N optical processing circuits to obtain the compensated and recovered optical signal.

[0009] The reservoir computing neural network can effectively reduce the number of parameters of the reservoir computing readout layer, reduce the deployment difficulty of the reservoir computing readout layer on the silicon optical chip, and reduce the optical signal compensation processing delay by disposing multiple reservoir layers.

[0010] In another possible implementation, the weighted weights of the output layer and the weighted weights of the readout weight layer are obtained based on training data, and the training data comprises initial optical signal data and distorted optical signal data, the initial optical signal data is optical signal data before transmission by the optical fiber, and the distorted optical signal data is optical signal data after transmission by the optical fiber.

[0011] An exemplary training process of the optical signal compensation neural network is as follows: obtaining a training data set, the training data set including initial optical signal data and distorted optical signal data, taking the distorted optical signal data as the input of the reservoir computing neural network layer, outputting an optical signal sequence, taking the output (i.e., the output optical signal sequence) of the reservoir computing neural network layer as the input of the reservoir computing readout layer, outputting a compensated and recovered optical signal, comparing the compensated and recovered optical signal with the initial optical signal data, and adjusting the linear weighting parameters of the output layer in the reservoir computing neural network layer and the weight parameters of the reservoir computing readout layer according to the comparison result. When a preset training condition is reached (e.g., the accuracy reaches a preset condition, or the training round reaches a preset number, or the optical signal neural network converges), the trained optical signal compensation neural network is obtained.

[0012] By using the relationship between the distorted optical signal and the initial optical signal in fiber communication, the optical signal compensation neural network is trained, the parameters involved in the optical signal compensation neural network (including the linear weighting parameters of the output layer in the reservoir computing neural network layer and the weight parameters of the reservoir computing readout layer) are optimized, and the compensation accuracy of the optical signal compensation neural network is improved.

[0013] In another possible implementation, the output layer and the readout weight layer are both provided with first programmable units. The first programmable units provided on the output layer are used to adjust the weighting weights of the output layer, and the first programmable units provided on the readout weight layer are used to adjust the weighting weights of the readout weight layer.

[0014] The parameters of the optical signal compensation neural network deployed on the silicon optical chip can be adjusted through the programmable units. For example, the parameters of the trained optical signal compensation neural network obtained through simulation training are deployed on the silicon optical chip, that is, the parameters of the optical signal compensation neural network deployed on the silicon optical chip are adjusted through the programmable units to be consistent with the parameters of the trained optical signal compensation neural network obtained through simulation training. The compensation and recovery of optical signals with different propagation speeds, transmission distances, and transmission wavelengths can be realized through the programmable units.

[0015] Optionally, the first programmable unit includes a feedback intensity modulation unit and a phase modulation unit. The feedback intensity modulation unit is used to modulate the feedback intensity of the optical signal, and the phase modulation unit is used to modulate the phase of the optical signal.

[0016] For example, the feedback intensity modulation unit can be a mach-zehnder interferometer (MZI) and a first micro-heater, the parameters of the MZI are configured by controlling the temperature of the first micro-heater, and then the feedback intensity of the hardware to the optical signal is configured. The phase adjustment unit can be a phase shifter (PS) and a second micro-heater, the parameters of the phase shifter are configured by controlling the temperature of the second micro-heater, and then the phase of the optical signal in the hardware is configured. The optical weight is configured by the MZI, and the phase error caused by thermal crosstalk and optical path difference is avoided by the phase shifter providing appropriate phase compensation.

[0017] In another possible implementation, each of the N reservoir layers includes a reservoir feedback loop, the reservoir feedback loop of each of the reservoir layers includes a feedback loop length parameter, and the feedback loop length parameter is determined based on a frequency interval of the wavelength division multiplexing system. The feedback loop length parameter is determined based on the frequency interval of the wavelength division multiplexing system, and then the free spectral range of the reservoir layer is matched to the frequency interval of the wavelength division multiplexing system.

[0018] In other words, by designing the free spectral range of the reservoir to match the frequency interval of the wavelength division multiplexing system, the length of the reservoir feedback loop is optimized to be compatible with the wavelength division multiplexing link transmission, thereby achieving simultaneous compensation and recovery of multiple wavelengths.

[0019] In another possible implementation, the reservoir feedback loop of each of the reservoir layers further includes a feedback intensity parameter, and the feedback intensity parameter is trained based on training data, the training data including initial optical signal data and distorted optical signal data, the initial optical signal data being optical signal data before the optical fiber transmission, and the distorted optical signal data being optical signal data after the optical fiber transmission.

[0020] Optionally, the feedback intensity parameter of the reservoir layer can be trained together with the output layer and the reservoir computing readout layer of the reservoir layer, or can be trained separately.

[0021] In another possible implementation, each of the reservoir layers is provided with a second programmable unit, and the second programmable unit is used to adjust the feedback intensity parameter of each of the reservoir layers. For example, the second programmable unit can include a MZI and a micro-heater, the parameters of the MZI are configured by controlling the micro-heater, and then the feedback intensity parameter of each of the reservoir layers is configured. For another example, the second programmable unit includes a MZI and a PS, and a micro-heater for heating the MZI and the PS, respectively, the MZI and the PS are configured by controlling the micro-heater, and then the feedback intensity and the phase of each of the reservoir layers are configured.

[0022] In another possible implementation, the reservoir computing readout layer is used to perform a one-dimensional convolution operation on the output optical signal to output the compensated and recovered optical signal.

[0023] In another possible implementation, a photodetector is further arranged on the silicon-based photonic signal processing chip, and the optical signal compensator is connected with the photodetector. That is, the optical signal compensator provided in the present application can be integrated with the photodetector in the optical module on a silicon optical chip as a whole, so that the optical signal is compensated and recovered before direct photoelectric conversion, thereby avoiding the use of the DSP compensation algorithm in the existing optical module and reducing the power consumption overhead in the optical module.

[0024] In a second aspect, the present application provides an optical receiver comprising the optical signal compensator as described in the first aspect or any possible implementation manner of the first aspect.

[0025] In a third aspect, the present application provides an optical module comprising the optical receiver as described in the second aspect.

[0026] In a fourth aspect, the present application provides an optical communication system comprising the optical module as described in the third aspect. BRIEF DESCRIPTION OF DRAWINGS

[0027] FIG. 1 shows an architecture schematic diagram of a reservoir computing neural network;

[0028] FIG. 2 shows an architecture schematic diagram of a cascaded reservoir computing neural network;

[0029] FIG. 3 shows a scheme implementation architecture schematic diagram of the related art 1;

[0030] FIG. 4 shows an implementation architecture schematic diagram of a dispersion compensation scheme;

[0031] FIG. 5 shows an implementation architecture schematic diagram of another dispersion compensation scheme;

[0032] FIG. 6 shows an application scenario of the optical signal compensator provided in the embodiments of the present application;

[0033] FIG. 7 shows an implementation architecture schematic diagram of the optical signal compensator provided in the embodiments of the present application;

[0034] FIG. 8 shows a structure schematic diagram of a receiving end of an existing optical module;

[0035] FIG. 9 shows a structure schematic diagram of a receiving end of an optical module to which the optical signal compensator provided in the embodiments of the present application is deployed;

[0036] FIG. 10 shows a configuration relationship schematic diagram of corresponding voltages of an MZI and a PS;

[0037] FIG. 11 shows a comparison schematic diagram of an initial signal, a fiber distortion signal and a signal compensated and recovered by using the optical signal compensator provided in the embodiments of the present application in fiber communication;

[0038] FIG. 12 shows a diagram of the relationship between BER and readout layer size, and the relationship between BER and the number of cascaded reservoirs, respectively;

[0039] FIG. 13 shows a diagram of the signal compensation effect of the optical signal compensation scheme provided by the embodiments of the present application for a WDM communication scenario with a frequency interval of 400 GHz when the fixed reservoir readout size is 5, and the signal compensation effect for a LAN WDM communication scenario when the fixed reservoir readout size is 10, respectively.

[0040] FIG. 14 shows a diagram of the relationship between BER and transmission distance, and the relationship between Q-factor and transmission distance for a C-band communication system under a communication scenario with a communication rate of 40 Gbaud, an OOK modulation format, and a transmission distance of 30-50 km, respectively. DETAILED DESCRIPTION

[0041] The term “and / or” mentioned in the present document is a description of the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B can represent the three cases of A alone, A and B together, and B alone. The symbol “ / ” in the present document represents an or relationship of the associated objects, for example, A / B represents A or B.

[0042] The terms “first” and “second” and the like in the description and claims of the present document are used to distinguish different objects, rather than to describe a specific order of the objects. For example, first memory chain data and second memory chain data are used to distinguish different memory chain data, rather than to describe a specific order of the memory chain data.

[0043] In the embodiments of the present application, the words “exemplary” or “for example” are used to mean serving as an example, instance, or illustration. Any embodiment or design scheme described as “exemplary” or “for example” in the embodiments of the present application should not be construed as being more preferred or advantageous than other embodiments or design schemes. Rather, the use of the words “exemplary” or “for example” is intended to present relevant concepts in a concrete manner.

[0044] In the description of the embodiments of the present application, unless otherwise specified, the meaning of “a plurality of” is two or more, for example, a plurality of processing units means two or more processing units, and the like; a plurality of elements means two or more elements, and the like.

[0045] In the description of the present application, the orientation or positional relationship indicated by the terms "center", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer" and the like is based on the orientation or positional relationship shown in the drawings, and is only for the convenience of describing the present application and simplifying the description, and does not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application.

[0046] In the description of the present application, it should be noted that unless otherwise explicitly specified and limited, the terms "mounting", "connection", "connection" should be understood broadly, for example, it can be fixed connection, or detachable connection, or abutment connection or integral connection; for those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0047] In order to facilitate the understanding of the scheme of the embodiments of the present application, the technical terms involved in the present application will be explained first.

[0048] On-off keying (OOK): a signal modulation format, i.e. amplitude modulation, the generated signal has only two values of 0 and 1.

[0049] Wavelength division multiplexing: that is, transmitting multiple wavelengths of signals in one optical fiber, each wavelength carries its own data. Ideally, we hope that each wavelength does not affect each other and transmits each other, but is coupled in the same optical fiber. However, in actual transmission process, the waves of each wavelength will affect each other, resulting in distortion of the signals carried on each wave.

[0050] Baud: usually called baud rate in the industry, also called symbol rate.

[0051] With the rapid maturity of generative AI, from text to video and machine perception, data volume will grow 10 times in the next few years, releasing huge bandwidth demand; the computing power required by large models will be concentrated in the cloud, giving birth to a new mode of intelligent computing cloud services, including the generation and delivery of computing power, and the provision of computing services tends to be centralized. In order to achieve more intelligent judgment and more real-time response, the training of large models needs to constantly exchange new data, and these data may span hundreds or even thousands of kilometers, requiring efficient transmission between data centers and between data centers and enterprise users. This transmission process relies heavily on the development of optical fiber communication. According to forecasts, by 2025, the monthly data transmission volume will reach 800 Exabyte (8x10 20 byte). Thanks to the development of high-speed optical devices, packaging technology and high-performance optical compensation algorithms, the networking capability of 400G optical modules has reached the level of 100G / 200G, covering long-distance and ultra-long trunk transmission scenarios.

[0052] In the past few decades, DSP technology based on Moore's Law node size expansion has continuously enabled the development of fiber transmission capacity, but the limitations of physical rules and manufacturing processes will soon bring the failure of future Moore's Law. At the same time, the limited power budget (usually 15-20W) in optical modules limits the deployment of high-power DSP technology in high-speed long dispersion scenarios, further hindering the deployment of 800G / 1.6T optical modules. To address this challenge, new solutions are needed. Low latency, low power silicon-based photonic signal processing chips are considered a key technology that can enable and continue the development of fiber capacity. This technology takes advantage of silicon-based materials and combines the latest technologies in photonics and electronics to achieve efficient signal processing. Compared to traditional DSP technology, silicon-based photonic signal processing chips have lower power consumption, greater bandwidth, and higher integration, better adapting to the power budget limitations of optical modules and providing higher performance. As the demand for global fiber network capacity continues to grow, the demand for low latency, low power silicon-based photonic signal processing chips is also increasing. This chip is expected to replace traditional DSP technology in high-speed long dispersion scenarios and become a key component of future fiber networks, and is expected to further enhance and develop fiber network capacity.

[0053] Currently, researchers have implemented various neural network architectures (silicon optical neural networks) on silicon optical chips using micro-ring array and Mach-Zehnder interferometer network structures, with a frequency response capability of more than GHz. Reservoir computing neural networks, as a variant of recurrent neural networks (RNN), have both linear memory and nonlinear processing capabilities, and only require training of the output layer, with low complexity, making them well-suited for implementation on silicon photonic chip platforms. The specific implementation principle is shown in Figure 1, which consists of an input layer, a reservoir layer, and an output layer. The input layer multiplies the input signal by the input weight W in (vector or matrix) and inputs it to the reservoir layer, which consists of one or more interconnected nonlinear nodes that perform nonlinear transformation on the input signal and nonlinear processing on the input signal in higher dimensions (in other words, the input signal with a dimension of one is projected to a higher dimension by multiplying the weight. For example, the input signal has a dimension of 1*M vector, and the reservoir layer has N nonlinear nodes (usually N>M), so the input weight W in is an M*N matrix, so the dimension of the signal becomes 1*N after multiplying the input weight matrix, and since N>M, the dimension of the signal is higher). The reservoir layer outputs the signal through the output weight W out to the final output layer. The biggest feature of the reservoir neural network that is different from traditional neural networks is that its input weight Win The connection weights W of each node in the pool layer are randomly generated and remain fixed in each round of experiments, requiring no training. The only weights W of the output layer need to be trained. out This architecture significantly reduces the number of system parameters and also makes hardware implementation possible. During the experiment, W can be randomly generated in each round. in And W, only train W. out After completing the training, W will... in W and W out It can be fixed. At the same time, the cascaded reservoir (see Figure 2) can enhance the memory length of the network architecture by linearly combining the output of each reservoir, thereby processing optical communication signals with higher speed and longer dispersion. Moreover, cascading on a silicon-based platform only requires simple waveguide connections, making it highly feasible.

[0054] Figure 1 shows a schematic diagram of the architecture of a reservoir computational neural network.

[0055] Figure 2 shows a schematic diagram of the architecture of a cascaded reservoir computing neural network.

[0056] Figure 1 shows a time-delay-based reservoir computing architecture, and Figure 2 shows a time-delay-based cascaded reservoir computing architecture. In this architecture, the reservoirs are connected sequentially, the input signal passes through each reservoir sequentially, and the outputs of all reservoirs are linearly combined to form the overall output signal.

[0057] Regarding optical signal compensation schemes in related technologies, there are still some areas that need to be explored and improved in the commercialization and industrialization of all-optical signal compensation solutions:

[0058] (1) High-speed optoelectronic devices with operating frequencies of 130 GHz and above have matured by 2023, but there is still a lack of clear scenarios and directions on how to coordinate the use of these devices or even jointly optimize them to adapt to all-optical processing solutions. In particular, for high-performance optical devices with operating frequencies of 130 GHz and above, such as modulators, ADCs, DACs and ICRs, their maturity makes 400 GHz QPSK high-performance coding possible. However, at present, academia and industry have not provided clear guidance on the route from independent devices to multi-device coordinated optimization, so there is a lack of clear development direction.

[0059] (2)Secondly, current all-optical compensation schemes mainly rely on numerical simulation, low-speed discrete device experiments and on-chip partial experiments, and lack of on-chip compensation experiments for high-speed signals. Most of the current all-optical compensation schemes focus on conceptual verification through numerical simulation or low-speed discrete device experiments. For on-chip all-optical experiments, the whole problem from design simulation to flow experiment needs to be considered, and the iteration period is long. At the same time, high-speed on-chip experiments need to accurately characterize the on-chip structure and accurately control the experimental conditions, so the experiment is difficult. In addition, the cost of high-bandwidth test instruments also limits the progress of this research.

[0060] (3)Finally, the indicators of optoelectronic devices are too discrete, resulting in a lack of clear definition of how to adopt optoelectronic hybrid packaging and the proportion of optical compensation to electrical compensation. At present, high-performance DSP algorithms have approached the Shannon limit of optical fiber, and the introduction of all-optical signal compensation systems can improve the performance of signal compensation but may introduce additional end-to-end damage. Therefore, how to define a set of indicator parameters to determine the ability of all-optical signal compensation is a problem that needs to be discussed. Balancing various indicators can promote the ability of optical compensation to electrical compensation, thereby achieving the maximum compensation effect, the lowest power consumption and delay.

[0061] In the past decade, photonic reservoir computing neural networks have rapidly emerged due to their computational versatility and low physical implementation complexity. Although the reservoir computing architecture has relatively few parameters, a large amount of exploration is needed from numerical simulation to accurate modeling of parameters in the hardware architecture. A representative example is the system framework of self-feedback injection lasers. Although researchers can numerically simulate the laser coupling equation in the dynamics of semiconductor lasers, they need a large number of physical experiments to determine the feedback length, feedback strength, laser emission power and even physical noise control required in the reservoir computing neural network. Similar situations also occur in micro-ring-based reservoir computing architectures, such as the bias size of the operating wavelength and the micro-ring resonance wavelength, which greatly affects the performance of the reservoir. Therefore, in order to optimize the performance of the reservoir computing neural network in experiments, researchers usually need to accumulate a large number of experiments and numerical simulations. To date, reservoir computing neural networks are still in the exploratory stage of processing benchmark tasks and have not been applied to real commercial scenarios.

[0062] For example, related technology one is a reservoir computing neural network architecture scheme based on semiconductor laser injection locking. FIG. 3 shows a scheme implementation architecture diagram of related technology one. As shown in FIG. 3, the core component of the scheme is a semiconductor laser with feedback, the field intensity and carrier density in the laser cavity are coupled with each other, the output signal is controlled by the Lang-Kobayashi rate equation with time delay feedback, and an optical injection dynamic term is additionally added. The reservoir layer is composed of a semiconductor laser and an optical loop feedback formed by an optical fiber, and the feedback is controlled by adjusting the attenuator in the loop. Here, the input signal of the reservoir layer is an electrical signal weighted by input weights, and an electro-optical modulator modulates the electrical signal into an optical signal output by the injection laser into the reservoir layer. After the input of the reservoir layer passes through a photodetector, the output weights are trained in a computer or a DSP. After the output layer is trained, it is fixed and no longer changed.

[0063] The implementation process of this architecture is as follows: the input signal x(t) is a received end distorted signal collected in an optical fiber link, and the input weights W in Directly weighted and upgraded, the output signal of the injection laser modulated by an arbitrary waveform generator (AWG) and an electro-optical modulator is an optical signal. Then, the input signal goes into the laser in the reservoir layer to produce a nonlinear change, and reservoir computing is performed in the feedback loop. Finally, the reservoir output passes through a photodetector to the output layer in the electrical domain, and the output weights are trained. The target signal of the training is the initial signal of the transmitting end in the optical fiber link, and the output weights are trained by linear regression algorithm. After the training is completed, the distorted signal passes through the reservoir and the trained output layer for inference, and the undistorted initial signal can be obtained, thereby achieving the effect of compensation and recovery.

[0064] However, the scheme has the following problems:

[0065] (1) The optical fiber distorted signal needs to be collected into the electrical domain for weighting of the output weights, so that the input of the system is an electrical signal instead of an initial optical signal in optical fiber communication, resulting in additional digital-to-analog conversion and mismatching with the existing communication scene.

[0066] (2) The core components in the architecture are discrete devices, which cannot be integrated into existing optical communication modules. The scheme is only a theoretical verification experiment, and it is difficult to deploy in practice. Neither the laser nor the ring shunt can be integrated into a chip.

[0067] (3) The use of multiple lasers results in high system power consumption, and the performance of the reservoir computing architecture is highly dependent on the output power of the laser, which is difficult to adjust.

[0068] In the past few decades, a variety of high-performance optical compensation DSP algorithms have been proposed to compensate for fiber dispersion and nonlinear loss, approaching the Shannon limit. In the high-speed short-distance communication scenario, signal distortion is mainly caused by dispersion. The existing electrical dispersion compensation algorithm is widely used in the receiving end, which uses the dispersion characteristics in the optical transmission network to process the transmitted data, thereby reducing the inter-symbol interference and improving the transmission quality of the optical transmission network.

[0069] The dispersion compensation schemes in the related art mainly include the following two kinds:

[0070] FIG. 4 shows a schematic diagram of an implementation architecture of a dispersion compensation scheme, which uses a dispersion compensation algorithm based on forward equalization technology. The forward equalization technology is the most widely used dispersion compensation algorithm at present, which regards the inter-symbol interference signal caused by the pulse spreading due to dispersion as a linear interference signal, and uses a cross filter to increase the reverse numerical signal to offset the interference signal, thereby achieving compensation for dispersion. This algorithm is mainly a combination filter composed of taps with different weights and different time delays. By adjusting the weight of the tap according to the estimation of the inter-symbol interference signal, linear equalization is achieved, and the effect of dispersion compensation is achieved. The linear equalizer has a simple principle and a light structure, and has been widely applied.

[0071] FIG. 5 shows a schematic diagram of an implementation architecture of another dispersion compensation scheme, which uses a dispersion compensation algorithm based on equalization technology of a least mean square filter. The least mean square filter equalization technology has an adaptive decision equalization function. The core part of the least mean square filter equalization technology is consistent with the combination filter in the forward equalization technology. The introduction of the least mean square filter can effectively suppress the interference of noise and improve the equalization effect of the inter-symbol interference caused by dispersion, and has strong adaptability.

[0072] However, the above two dispersion compensation schemes still have some problems, for example: with the increase of the number of taps, redundant data can cause the equalizer to amplify noise, which is not conducive to the transmission of signals in the optical fiber; and the scheme needs to be deployed in the DSP, which can introduce a large delay for high-speed signals and a large power consumption for long dispersion signals.

[0073] Therefore, the embodiments of the present application provide an optical signal compensator, which uses a full-optical on-chip readout layer full-optical integrated cascade reservoir computing architecture (i.e., an optical signal compensation neural network deployed on a silicon-based photonic signal processing chip, including a reservoir computing neural network and a reservoir computing readout layer). The optical signal compensator can directly compensate and recover the optical fiber signal in the optical domain before the optical fiber signal is detected by a photoelectric detector, so that the optical fiber signal does not need to be compensated and recovered by a DSP after being detected, thereby reducing the power consumption and delay of the electrical chip in the optical module.

[0074] The specific application scenarios and specific structures of the optical signal compensator provided by the embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0075] With the rapid development of short-distance applications represented by data centers, supercomputing clusters, and interconnected networks, optical fiber communication has an increasing demand for high-capacity data transmission. This scenario usually uses wavelength division multiplexing multi-channel direct modulation and direct detection signals, such as communication signals using OOK and PAM4 modulation formats.

[0076] FIG. 6 shows an application scenario of the optical signal compensator provided by the embodiments of the present application. As shown in FIG. 6, the optical signal compensator provided by the embodiments of the present application can be compatible with the current mainstream wavelength division multiplexing optical module, and can be applied to the receiving end (also referred to as an optical receiver) of the optical module to compensate and recover the distorted optical signal transmitted by the optical fiber in the optical domain.

[0077] FIG. 7 shows an implementation architecture diagram of the optical signal compensator provided by the embodiments of the present application. As shown in FIG. 7, the implementation architecture of the optical signal compensator provided by the embodiments of the present application is composed of a cascaded reservoir computing neural network layer and a reservoir computing readout layer. The cascaded reservoir computing neural network layer architecture outputs in parallel into the reservoir computing readout layer architecture. The optical signal is input and output through the waveguide (line segment in FIG. 7), and the parameters of the overall architecture can be optimized by controlling through the programmable unit (rectangular block in FIG. 7) to compensate and recover the distorted optical signal. The programmable unit includes an intensity modulation unit and a phase modulation unit, which can be realized by adjusting the micro-heaters on the MZI and the phase shifter, respectively.

[0078] As can be seen from FIG. 7, the optical signal compensation neural network of the optical signal compensator provided by the embodiments of the present application includes a cascaded reservoir computing neural network layer and a reservoir computing readout layer. The output end of the cascaded reservoir computing neural network layer is connected with the output end of the optical fiber, for receiving the distorted optical signal transmitted by the optical fiber. The distorted optical signal in the optical fiber link first enters the cascaded reservoir architecture and passes through each reservoir in turn. The feedback intensity γ and the feedback length τ of each reservoir are designed and optimized correspondingly, so that the performance of each reservoir reaches the best. The output of each reservoir is parallel output and weighted by the programmable unit of MZI+PS, to obtain the output of the cascaded reservoir architecture. The output signal is then input into the reservoir computing readout layer. The weight parameters in this structure are determined by training the reservoir output signal and the initial signal, and then configured by accurate hardware parameter modeling. Once deployed, the output signal of the cascaded reservoir is input into the reservoir readout layer, and the compensated and recovered optical signal is obtained.

[0079] In the embodiments of the present application, the cascaded reservoir computing neural network layer architecture realizes an on-chip all-optical integrated cascaded reservoir computing architecture. In combination with the characteristics of the silicon optical chip platform (such as refractive index, waveguide size, delay line length and temperature, etc.), the reservoir structure is comprehensively analyzed and optimized from the chip design simulation to the whole process of the flow experiment to obtain the best compensation effect. By designing the free spectral range of the reservoir to match the frequency interval of the wavelength division multiplexing system, the length of the feedback loop of the reservoir can be optimized to be compatible with the wavelength division multiplexing link transmission, so as to simultaneously compensate and recover multiple wavelengths.

[0080] The length τ of the feedback loop of each reservoir in the cascaded reservoir computing neural network layer cannot be adjusted after the flow experiment, so the length τ of the feedback loop of each reservoir in the cascaded reservoir computing neural network layer needs to be determined during the silicon optical chip design simulation. The length τ of the feedback loop of each reservoir determines the free spectral range of each reservoir. In actual application, the free spectral range of each reservoir needs to match the frequency interval of the wavelength division multiplexing system in the actual application optical communication system, and then be compatible with the wavelength division multiplexing link transmission, so as to realize the simultaneous compensation and recovery of multiple wavelength optical signals. Therefore, during the design simulation, the frequency interval of the wavelength division multiplexing system of the optical communication system to be applied can be analyzed, the free spectral range of each reservoir is determined according to the frequency interval of the wavelength division multiplexing system, and the length τ of the feedback loop of each reservoir is determined based on the free spectral range of each reservoir, so that the cascaded reservoir computing neural network layer is compatible with the wavelength division multiplexing link transmission, thereby realizing the simultaneous compensation and recovery of multiple wavelength optical signals.

[0081] For example, the cascaded reservoir computing neural network layer includes three reservoirs, namely reservoir 1, reservoir 2 and reservoir 3, and the length of the feedback loop of each reservoir in the reservoir 1, the reservoir 2 and the reservoir 3 needs to be determined during the silicon optical chip design simulation stage. Through analysis, it is known that the frequency interval of the wavelength division multiplexing system of the optical communication system to be applied is 100GHz, then the free spectral range of the reservoir 1 is designed to be 10GHz, the free spectral range of the reservoir 2 is designed to be 20GHz, and the free spectral range of the reservoir 3 is designed to be 25GHz, that is, the free spectral range of each reservoir is designed to be different, but the free spectral range of each reservoir can be divided by the frequency interval 100GHz of the wavelength division multiplexing system, forming an optical vernier effect. At this time, it is called that the free spectral range of each reservoir matches the frequency interval of the wavelength division multiplexing system, so that the simultaneous compensation and recovery of multiple wavelength optical signals can be realized.

[0082] When the free spectral range of each reservoir is determined, the length of the feedback loop of each reservoir is also determined. For example, if the free spectral range of reservoir 1 is designed to be 10 GHz, then the feedback loop length of reservoir 1 is y1; if the free spectral range of reservoir 2 is designed to be 20 GHz, then the feedback loop length of reservoir 2 is y2; and if the free spectral range of reservoir 3 is designed to be 25 GHz, then the feedback loop length of reservoir 3 is y3.

[0083] As can be seen from FIG. 7, the parameters of each reservoir in the cascaded reservoir computing neural network layer also include a feedback strength parameter γ (the weight parameter of each reservoir can be adjusted by adjusting the feedback strength parameter γ), and each reservoir is provided with a programmable unit, and the feedback strength parameter γ of each reservoir can be adjusted through the programmable unit. The feedback strength γ is different from the feedback length τ, and the feedback strength γ is adjustable. The optimal feedback strength γ of each reservoir can be obtained by training the training data, and then the feedback strength parameter γ of each reservoir is adjusted and optimized through the programmable unit to obtain the best compensation effect.

[0084] It is easy to understand that the feedback strength γ means the intensity lost by light after propagating in a medium, and can also be understood as the amplitude change range of light after propagating in a medium. For example, the feedback strength γ of each reservoir can be understood as an intensity loss parameter of the light signal propagating in the feedback loop of each reservoir, or the amplitude of the light signal after propagating in the feedback loop of each reservoir.

[0085] Optionally, the feedback strength γ of each reservoir can be trained and adjusted together with the linear weighting parameter of the output layer of the cascaded reservoir computing neural network layer and the weight parameter of the reservoir computing readout layer. For example, the training process of the optical signal compensation neural network is as follows: obtaining a training data set, the training data set including initial optical signal data and distorted optical signal data, taking the distorted optical signal data as the input of the reservoir computing neural network layer, outputting an optical signal sequence, taking the output (i.e., the output optical signal sequence) of the reservoir computing neural network layer as the input of the reservoir computing readout layer, outputting a compensated and recovered optical signal, comparing the compensated and recovered optical signal with the initial optical signal data, and adjusting the feedback strength parameter γ of each reservoir in the cascaded reservoir computing neural network layer, the linear weighting parameter of the output layer, and the weight parameter of the reservoir computing readout layer according to the comparison result. When a preset training condition is reached (for example, the compensation and recovery accuracy of the optical signal reaches a preset condition, or the training round reaches a preset number, or the optical signal compensation neural network converges), the trained optical signal compensation neural network is obtained.

[0086] After the training of the optical signal compensation neural network is completed, according to the parameters obtained after the training (including the intensity parameters γ of each reservoir, the linear weighting parameters of the output layer, and the weight parameters of the reservoir calculation readout layer), the programmable unit is controlled to adjust the parameters of the optical signal compensation neural network deployed on the silicon optical chip (including the intensity parameters γ of each reservoir, the linear weighting parameters of the output layer, and the weight parameters of the reservoir calculation readout layer), so as to realize the deployment of the trained optical signal compensation neural network on the silicon optical chip.

[0087] In another example, the feedback intensity γ of each reservoir in the cascaded reservoir calculation neural network layer can also be trained separately, for example, the linear weighting parameters of the output layer in the cascaded reservoir calculation neural network layer are set to default values, and the reservoir calculation readout layer is not needed. The training process of each reservoir in the cascaded reservoir calculation neural network layer is as follows: obtaining a training data set, the training data set including initial optical signal data and distorted optical signal data, taking the distorted optical signal data as the input of the reservoir calculation neural network layer, outputting the compensated and recovered optical signal, comparing the compensated and recovered optical signal with the initial optical signal data, and adjusting the feedback intensity parameters γ of each reservoir in the cascaded reservoir calculation neural network layer according to the comparison result. When a preset training condition is reached (for example, the compensation and recovery accuracy of the optical signal reaches a preset condition, or the training round reaches a preset number, or the optical signal neural network converges), the trained feedback intensity parameters γ of each reservoir are obtained.

[0088] After the feedback intensity parameters γ of each reservoir in the cascaded reservoir calculation neural network layer are trained, the feedback intensity parameters γ of each reservoir are frozen, and then the linear weighting parameters of the output layer and the weight parameters of the reservoir calculation readout layer in the cascaded reservoir calculation neural network layer are trained using the training data.

[0089] After the deployment of the trained optical signal compensation neural network, the compensation and recovery of the distorted optical signal can be realized. For example, the input end of the optical signal compensation neural network deployed on the silicon optical chip is connected with the output end of the optical fiber. The distorted optical signal transmitted through the optical fiber first enters the cascaded reservoir calculation neural network, and then the features of the distorted optical signal are extracted after the cascaded reservoir calculation neural network processing, for example, including the timing features and correlation features between each optical signal in the distorted optical signal sequence. Then, the features of the distorted optical signal are output as the input of the reservoir calculation neural network layer. After the 1D convolution operation of the reservoir calculation neural network layer, the compensated and recovered optical signal, i.e., the initial optical signal before the light transmission, is output, thereby solving the distortion problem of the optical signal after the high-speed transmission through the optical fiber, reducing the BER, and further improving the optical communication quality.

[0090] Optionally, the cascaded reservoir calculation neural network is also used for storing and memorizing the information carried by the optical signal.

[0091] It should be noted that the implementation architecture shown in FIG. 7 is only one example that can be implemented, and does not constitute a limitation on the embodiments of the present application. For example, the cascaded reservoir computing neural network layer can also be a reservoir computing neural network layer, that is, the reservoir computing neural network layer only includes one reservoir.

[0092] In one example, the programmable unit can adjust the adjustable parameters of the optical signal compensation neural network (for example, including the intensity parameters γ of each reservoir, the linear weighting parameters of the output layer, and the weight parameters of the reservoir computing readout layer) in a way of thermal tuning (for example, adjusting the temperature of the MZI and the PS by a micro-heater to change the intensity and phase of the two, respectively), or in other ways, for example, by adjusting the electric field, or by a MEMS unit or a phase change material unit. The embodiments of the present application do not limit the parameter adjustment mode of the programmable unit, and a suitable adjustment mode can be selected as needed.

[0093] Through the setting of the programmable unit, the reprogramming of each adjustable parameter in the optical signal compensation neural network can be realized, and the compensation and recovery of optical signals with different propagation speeds, transmission distances, and transmission wavelengths can be realized. In other words, through the programmable unit, the adjustable weight parameters of the optical signal compensation neural network can be adjusted to be applicable to the compensation and recovery of distorted optical signals in different communication scenarios.

[0094] The core implementation architecture of the optical signal compensator provided by the embodiments of the present application includes a cascaded reservoir computing neural network architecture and a reservoir computing readout layer architecture, and the two architectures are connected through a waveguide, which belongs to a serial unit. The overall structure is equivalent to directly implementing the reservoir and the readout layer of reservoir computing in the optical domain, so that the compensation and recovery task of the distorted optical signal can be directly completed in the optical domain.

[0095] FIG. 8 shows a structure diagram of an existing optical module receiving end.

[0096] FIG. 9 shows a structure diagram of an optical module receiving end in which the optical signal compensator provided by the embodiments of the present application is deployed.

[0097] As shown in FIG. 9, the optical signal compensator provided by the embodiments of the present application can be integrated on the same silicon optical chip as the photodetector, to realize the compensation and recovery of the optical signal directly before photoelectric conversion, thereby avoiding the use of the DSP compensation algorithm in the existing optical module (see FIG. 8), and greatly reducing the power consumption overhead in the optical module.

[0098] The optical signal compensator provided in the embodiments of the present application can be particularly applied to compensation and recovery of a distorted optical signal of direct modulation and direct detection, and the distorted optical signal after fiber transmission directly enters an optical signal compensation neural network deployed on a silicon optical chip, sequentially passes through a cascade reservoir computing neural network layer and a reservoir computing readout layer, and finally outputs a compensated and recovered optical signal, thereby achieving compensation and recovery of the distorted optical signal in the optical domain.

[0099] The number of reservoirs of the cascade reservoir computing neural network layer and the number of size of the reservoir readout in the optical signal compensation neural network can be optimized according to the condition of the optical signal to be compensated.

[0100] In the programmable unit, the MZI and the PS are configured to feedback intensity and phase modulation by adjusting the micro-heater thereof. Since the compensated optical signal is a direct modulation and direct detection signal, the actual optical weight is configured by the MZI, and the function of the phase shifter is to provide appropriate phase compensation to avoid phase errors caused by thermal crosstalk and optical path difference.

[0101] The relationship between the distorted signal and the initial signal using fiber communication is used for training, and each parameter involved in the optical signal compensation neural network architecture is optimized. Through accurate parameter modeling, accurate parameter configuration (including adjustable parameters in the optical signal compensation neural network, such as weight parameters of each reservoir in the cascade reservoir computing neural network layer (adjusted by adjusting the feedback intensity parameter γ), linear weighting parameters of the output layer in the cascade reservoir computing neural network layer, and weight parameters in the reservoir computing readout layer) can be achieved. In actual application, the micro-heater needs to be applied with a voltage value through the corresponding circuit structure on the silicon optical chip, so as to realize the configuration of the MZI and the phase shifter.

[0102] FIG. 10 shows a schematic diagram of a configuration relationship between the MZI and the PS corresponding voltage.

[0103] Through the cooperation of the MZI and the phase shifter, the weight configuration in the range of [-1, 1] can be realized. Optionally, a configuration sequence is to first configure the weight value in the range of [0, 1] using the MZI, and then control the positive and negative of the weight value using the phase shifter.

[0104] The following two specific embodiments are used to introduce the application of the optical signal compensator provided in the embodiments of the present application in practice.

[0105] In the first specific embodiment, it is applied to numerical simulation of high-speed optical communication signals, and the data format and communication parameters are shown in Table 1.

[0106] Table 1

[0107] For the direct detection application in data center, we numerically simulate the communication rate and distance as 100 Gbaud and 10 km, respectively, and use higher order PAM4 modulation format, which means two bits per symbol (i.e. baud). In order to avoid the impact of nonlinear effects, we set the input power as 0 dBm, so that the impact of linear dispersion is mainly considered. In the training process, 10000 symbols are used to optimize the parameters in the system, and once the training is completed, another 10000 symbols are used to test to obtain the BER value.

[0108] Fig. 11 shows a comparison diagram of an initial signal, a fiber distortion signal and a signal compensated and recovered by the optical signal compensator provided by the embodiment of the application in optical fiber communication. As shown in Fig. 11, we use the fiber distortion signal (b in Fig. 11) as the input and the initial signal (a in Fig. 11) as the target for training. In actual application, we migrate the trained parameters to a silicon optical chip, and use a micro-heater to configure the programmable units in the cascaded reservoir architecture and the reservoir readout architecture. Once the configuration is completed, the fiber distortion signal can directly enter the silicon optical chip for processing, and the compensated and recovered signal as shown in (c) in Fig. 11 is obtained.

[0109] In this embodiment, the embodiment of the application first compensates and recovers the signal with the maximum dispersion (1270 nm) in the O-band scenario with a communication rate of 100 Gbaud and a communication distance of 10 km.

[0110] Fig. 12 shows diagrams of the relationship between BER and readout layer size, and the relationship between BER and the number of cascaded reservoirs. As shown in Fig. 12, the compensation effect of the optical signal with a communication rate of 100 Gbaud, a communication distance of 10 km and a modulation format of PAM4 at a carrier wavelength of 1270 nm, (a) in Fig. 12 shows the relationship between BER and readout layer size; (b) in Fig. 12 shows the relationship between BER and the number of cascaded reservoirs, with the readout layer size fixed at 15. Numerical simulation experiments are sequentially performed on the reservoir readout layer and the cascaded reservoir architecture, and the following two technical effects are obtained:

[0111] (1) As the readout layer size increases, the bit error rate (BER) of the communication signal decreases accordingly: until the readout size reaches 43, the BER can meet the requirements of HD-FEC. Considering that the output size of 43 is difficult to realize on an actual silicon optical chip and is difficult to operate, we increase the number of cascaded reservoirs while fixing the readout size at 15 to explore the relationship between BER and the number of cascaded reservoirs.

[0112] (2) With the increase of the number of cascaded reservoirs, the BER can also be optimized accordingly. Until the number of reservoirs increases to 3, the BER can also meet the requirements of HD-FEC. At this time, the energy consumption is nearly 300 times lower than that of the DSP.

[0113] Next, for the general-purpose WDM system, we consider the following two communication scenarios (communication waveband is O-band, communication rate is 100Gbaud / λ, and communication distance is 10km) to explore the compensation recovery effect of the optical signal compensation scheme provided by the embodiments of the present application:

[0114] FIG. 13 shows the signal compensation effect of the optical signal compensation scheme provided by the embodiments of the present application for the WDM communication scenario with a frequency interval of 400GHz when the fixed reservoir readout size is 5, and the signal compensation effect of the LAN WDM communication scenario when the fixed reservoir readout size is 10, respectively.

[0115] (1) WDM communication scenario with a frequency interval of 400GHz: the wavelength of the WDM system is 1304.58, 1306.85, 1309.14 and 1311.43nm, and the corresponding dispersion range for 10km is -19~+11ps / nm. As shown in (a) of FIG. 13, the fixed readout layer scale is 5, and the optical signal compensation scheme architecture provided by the embodiments of the present application can compensate and recover all dispersion distortions covered by the WDM system.

[0116] As shown in (b) of FIG. 13, the fixed readout layer scale is 5, and the optical signal compensation scheme architecture provided by the embodiments of the present application can compensate and recover all dispersion distortion signals covered by the WDM system.

[0117] (2) LAN WDM communication scenario: the wavelength of the WDM system is 1295.56, 1300.05, 1304.58 and 1309.58nm, and the corresponding dispersion range for 10km is -28.4~+9.4ps / nm. As shown in (b) of FIG. 10, the fixed readout layer scale is 10, and the optical signal compensation scheme architecture provided by the embodiments of the present application can compensate and recover all dispersion distortion signals covered by the WDM system.

[0118] The optical signal compensator provided by the embodiments of the present application can be applied to direct modulation and direct detection communication signals, and compensates and recovers the optical signals directly in the optical domain, eliminates the use of electrical dispersion compensation algorithm in the DSP after photoelectric detection, and significantly reduces the power consumption of the electrical chip in the optical module; by using the combination of the programmable unit of MZI+phase shifter, the full-optical positive-negative matrix deployment of [-1, 1] is realized, so as to match the dispersion wavelengths in multiple wavelength ranges; and by optimizing the number of cascaded reservoirs and the reservoir readout size, the optical signal compensator can match multiple communication scenarios, and the application range of the optical signal compensator is improved.

[0119] In a second specific embodiment, numerical simulation is performed on optical communication signals with a communication band of C-band, and the data format and communication parameters are as shown in Table 2:

[0120] Table 2

[0121] For general transmission bands, the numerical simulation sets the communication rate and distance to be 40 Gbaud and 30-50 km respectively, and uses the OOK modulation format. In order to avoid the influence of nonlinear effects, the input power is set to 0 dBm, so that the influence of linear dispersion is mainly considered. In the training process, 10000 symbols are used to optimize the parameters in the system, and once the training is completed, another 10000 symbols are used to test and obtain the BER value.

[0122] In this embodiment, the optical signal compensation scheme provided by the embodiments of the present application compensates and recovers the communication signals with a communication rate of 40 Gbaud and an OOK modulation format in a 30-50 km transmission distance range in a communication scenario with a communication band of C-band.

[0123] FIG. 14 shows the relationship between BER and transmission distance, and the relationship between Q-factor and transmission distance of the optical signal compensator provided by the embodiments of the present application for a C-band communication system in a communication scenario with a communication rate of 40 Gbaud, an OOK modulation format, and a transmission distance of 30-50 km.

[0124] With the increase of the number of cascaded reservoirs, the BER is also optimized accordingly. As shown in FIG. 14(a), under different transmission distances, when the number of reservoirs is less than 3, the BER gradually decreases.

[0125] The greater the dispersion distortion, the better the compensation effect of the optical signal compensator provided by the embodiments of the present application. As shown in FIG. 14(b), compared with the scenario without reservoirs and with three reservoirs, the transmission distance is from 30 km to 50 km, and the dispersion is enhanced in turn, and the Q-factor is improved from 1.17 dB to 1.57 dB in turn.

[0126] The embodiment of the present application also provides an optical receiver comprising the optical signal compensator as described above, and optionally, the optical signal compensator can be integrated on the same silicon optical chip as a photodetector, so that the distorted optical signal is compensated and recovered in the optical domain, the use of the DSP compensation algorithm in the existing optical module greatly reduces the power consumption in the optical module, and meanwhile, the photoelectric conversion and digital-to-analog conversion are not needed before the optical signal compensation, the calculation complexity is greatly reduced, the processing speed of the optical signal compensation is greatly improved, and the communication delay is greatly reduced.

[0127] The present application provides an optical module comprising the optical receiver as mentioned above (which can also be referred to as a receiving end of the optical module), and the receiving end of the optical module is provided with the optical signal compensator as mentioned above, the optical signal compensator is connected with an optical fiber, receives the distorted optical signal transmitted by the optical fiber, and compensates and recovers the distorted optical signal in the optical domain, so that the high power consumption and high delay caused by the compensation in the electrical domain after the photoelectric conversion and digital-to-analog conversion are avoided.

[0128] The present application provides an optical communication system comprising the optical module as described above, since the optical module provided by the embodiment of the present application adopts the optical signal compensator provided by the embodiment of the present application, the distorted optical signal can be compensated and recovered in the optical domain, the compensation and recovery do not need to use the DSP algorithm of the electrical chip in the optical module, the energy consumption of the optical module is greatly reduced, and the overall energy consumption of the optical communication system is further reduced; and the photoelectric conversion and digital-to-analog conversion are not needed before the optical signal compensation, so that the accurate compensation and recovery of the distorted signal can be completed in the optical domain, the BER of the communication system is reduced, and the communication performance of the communication system is improved.

[0129] In the description of the present specification, specific features, structures, materials or characteristics can be combined in any one or more embodiments or examples in a suitable manner.

[0130] Finally, it should be pointed out that the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit the same; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that the technical solutions recorded in the foregoing embodiments can still be modified, or some technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. An optical signal compensator, characterized by, The application relates to a photonic signal compensation neural network deployed on a silicon-based photonic signal processing chip and used for compensating and recovering a received distorted optical signal. The photonic signal compensation neural network comprises: a reservoir computing neural network layer connected with an output end of an optical fiber and used for receiving the distorted optical signal transmitted by the optical fiber and processing the distorted optical signal to obtain an output optical signal sequence; a reservoir computing readout layer connected with the reservoir computing neural network layer and used for processing the output optical signal sequence to obtain a compensated and recovered optical signal. The reservoir computing neural network layer comprises N reservoir layers and an output layer, the output layer is connected with the N reservoir layers respectively, and is used for performing weighted processing on output optical signals of the N reservoir layers to obtain the output optical signal sequence, wherein N is a positive integer greater than 1.

2. The optical signal compensator of claim 1, wherein, The reservoir computing readout layer comprises: N optical processing loops used for separating the output optical signal sequence in a time dimension into the N optical processing loops; a linear weighting layer used for performing weighted processing on optical signals output by the N optical processing loops to obtain the compensated and recovered optical signal. The weighted weights of the output layer and the weighted weights of the readout weight layer are obtained based on training data, the training data comprises initial optical signal data and distorted optical signal data, the initial optical signal data is optical signal data before transmission by the optical fiber, and the distorted optical signal data is optical signal data after transmission by the optical fiber.

3. The optical signal compensator of claim 2, wherein, The output layer and the readout weight layer are provided with first programmable units, the first programmable units arranged on the output layer are used for adjusting the weighted weights of the output layer, and the first programmable units arranged on the readout weight layer are used for adjusting the weighted weights of the readout weight layer.

4. The optical signal compensator according to claim 2 or 3, characterized in that The first programmable units comprise a feedback intensity modulation unit and a phase modulation unit, the feedback intensity modulation unit is used for modulating feedback intensity of an optical signal, and the phase modulation unit is used for modulating a phase of the optical signal.

5. The optical signal compensator of claim 4, wherein, The feedback intensity modulation unit comprises a Mach-Zehnder interferometer and a first micro-heater, the first micro-heater is used for heating the Mach-Zehnder interferometer, adjusting a temperature of the Mach-Zehnder interferometer to modulate the feedback intensity of the optical signal; 6. The optical signal compensator of claim 5, wherein, The phase modulation unit comprises a phase shifter and a second micro-heater, the second micro-heater is used for heating the phase shifter, adjusting a temperature of the phase shifter to modulate the phase of the optical signal. Each reservoir layer of the N reservoir layers comprises a reservoir feedback loop, the reservoir feedback loop of each reservoir layer comprises a feedback loop length parameter, and the feedback loop length parameter is determined based on a frequency interval of a wavelength division multiplexing system.

7. The optical signal compensator of any of claims 2-6, wherein, The reservoir feedback loop of each reservoir layer further comprises a feedback intensity parameter, and the feedback intensity parameter is obtained based on training data, the training data comprises initial optical signal data and distorted optical signal data, the initial optical signal data is optical signal data before transmission by the optical fiber, and the distorted optical signal data is optical signal data after transmission by the optical fiber.

8. The optical signal compensator of claim 7, wherein, ​ 9. The optical signal compensator of claim 8, wherein, The individual reservoir layers are provided with second programmable units for adjusting a feedback strength parameter of the individual reservoir layers.

10. The optical signal compensator of any of claims 1-9, wherein, The reservoir computing readout layer is configured to perform a 1-dimensional convolution operation on the output optical signal to output the compensated recovered optical signal.

11. The optical signal compensator of any of claims 1-10, wherein, The silicon-based photonic signal processing chip is further provided with a photodetector, and the optical signal compensator is connected with the photodetector.

12. An optical receiver, characterized in that, An optical signal compensator as claimed in any one of claims 1-11.

13. An optical module characterized by comprising: An optical receiver as claimed in claim 12.

14. An optical communication system, characterized by An optical module as claimed in claim 13.

Images