Optical signal synchronization calibration device and method of an optical computing system

Abstract

The application provides an optical signal synchronous calibration device and method of an optical computing system, the device comprising: an optical probe array arranged on a chip, an AI coprocessor, an optical path tuning engine and a low-speed electrical feedback line arranged on a board card; the optical probe array detects optical signals of wavelength channels corresponding to a plurality of optical computing units and converts the optical signals into electrical signals; the AI coprocessor drives the corresponding optical computing unit to emit a detection pulse according to a pre-allocated time slot index variable, extracts a phase error from the electrical signal and calculates a final calibration quantity based on the phase error; the optical path tuning engine performs an optical path tuning operation based on the final calibration quantity to calibrate the error of the corresponding optical computing unit; and the low-speed electrical feedback line transmits the electrical signal to the AI coprocessor and transmits the detection pulse to the chip. The application realizes real-time calibration in a dynamic environment through spatial distributed detection and board-level centralized processing cooperation.

Description

Technical Field

[0001] This invention belongs to the field of optical interconnect and optical computing technology, and specifically relates to an optical signal synchronization calibration device and method for an optical computing system. Background Technology

[0002] With the rapid development of silicon photonics integration, optoelectronic co-packaging, and large-scale neuromorphic optical computing chips, the integration and complexity of optical computing systems have significantly increased. The number of optical computing units is growing exponentially, leading to random drift in phase, timing, and power of optical signals between different boards, chips, and even different optical waveguides on the same chip. This drift not only affects the computational accuracy of the optical computing system but also reduces its stability and reliability. Traditional optical signal calibration methods mainly employ electrical domain clock allocation or offline spectrometer calibration, which have significant limitations: electrical domain clock allocation suffers from large delays due to photoelectric conversion, making it difficult to meet real-time requirements; offline spectrometer calibration suffers from high power consumption, complex operation, and inability to adaptively track temperature / stress changes. Especially in large-scale optical computing systems, these traditional methods struggle to achieve fast and accurate synchronous calibration, severely restricting the large-scale deployment and practical application of optical computing systems.

[0003] Therefore, how to provide an optical signal synchronization calibration device and method for an optical computing system, realize on-chip in-situ detection and board-level rapid tuning, reduce power consumption while ensuring real-time performance and scalability, is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] This invention provides an optical signal synchronization calibration device and method for an optical computing system to solve at least one of the above-mentioned technical problems.

[0005] To address the aforementioned technical problems, in a first aspect, the present invention provides an optical signal synchronization calibration device for an optical computing system. The optical computing system includes a chip and a board, the chip being coupled to the board and having multiple optical computing units. The device includes:

[0006] An optical probe array is deployed at the four corners and the center of the chip, configured to detect optical signals of the corresponding wavelength channels of multiple optical computing units and convert the optical signals into electrical signals;

[0007] An AI coprocessor, deployed on the board, is configured to drive the corresponding optical computing unit to emit detection pulses according to a pre-allocated time slot index variable, and to extract phase error from the electrical signal and calculate the final calibration amount based on the phase error.

[0008] An optical path tuning engine, deployed on the board, is configured to perform optical path tuning operations based on the final calibration value to calibrate the error of the corresponding optical computing unit;

[0009] A low-speed electrical feedback line is connected between the chip and the board, configured to transmit the electrical signal to the AI ​​coprocessor and the probe pulse to the chip.

[0010] Preferably, each optical probe in the optical probe array is composed of a microring resonator and a photodiode. The microring resonator is used to filter and resonate optical signals of a specific wavelength, and the photodiode is used to convert the optical signal into an electrical signal. The Q value of the optical probe is 8000 and the free spectral range is 8 nm.

[0011] Preferably, the phase error is obtained by taking the phase angle of the output voltage of the photodiode.

[0012] Preferably, the final calibration value includes a wide-range calibration value and a precision calibration value. The optical path tuning engine includes a silicon-based thermal tuning module and a piezoelectric tuning module. The silicon-based thermal tuning module drives a micro-thermal wire array based on the wide-range calibration value, and changes the effective refractive index of the micro-ring resonator through the thermo-optic effect to achieve a large-stroke coarse adjustment within a range of ±100 ps. The piezoelectric tuning module controls the deformation of the piezoelectric ceramic based on the precision calibration value, and adjusts the optical path of the micro-ring resonator through deformation to achieve a high-resolution fine adjustment within a range of ±10 ps.

[0013] Preferably, driving the corresponding optical computing unit to emit a probe pulse according to the pre-allocated time slot index variable includes:

[0014] The AI ​​coprocessor transmits a calibration trigger frame to the chip level via the low-speed electrical feedback line. The calibration trigger frame includes a calibration start command and a time slot allocation table of optical computing units and wavelengths.

[0015] After receiving the calibration trigger frame, each optical computing unit at the chip level transmits probe pulses sequentially according to the corresponding time slot index variable allocated in the time slot allocation table.

[0016] Preferably, the step of calculating the final calibration amount based on the phase error includes:

[0017] Perform PID calculations on the phase error to obtain the PI control quantity;

[0018] Obtain the current environmental parameters and input them into the trained prediction model to obtain the feedforward compensation amount;

[0019] The final calibration value is obtained by summing the PI control value and the feedforward compensation value.

[0020] Preferably, the formula for calculating the final calibration amount Δτ is:

[0021] Δτ=Δτ PID + AIpred;

[0022] Δτ PID =Kp·Eφ+Ki·∫Eφdt;

[0023] Where, Δτ PID Here, AIpred is the PI control input, Eφ is the feedforward compensation input, Kp is the proportional control coefficient, and Ki is the integral control coefficient.

[0024] In a second aspect, the present invention also provides an optical signal synchronization calibration method, applied to an optical signal synchronization calibration device as described in any one of the first aspects, the method comprising:

[0025] During the T0 to T1 phase, the AI ​​coprocessor transmits a calibration trigger frame to the chip level. The calibration trigger frame contains a calibration start command and a time slot allocation table for each optical computing unit and wavelength.

[0026] During the T1 to T2 phase, after receiving the calibration trigger frame, each optical computing unit at the chip level sequentially transmits probe pulses according to the time slot index variable allocated in the time slot allocation table.

[0027] In stages T2 to T3, the optical probe array detects the optical signals of the wavelength channels corresponding to the multiple optical computing units and converts the optical signals into electrical signals. The AI ​​coprocessor extracts the phase error from the electrical signals and calculates the final calibration amount based on the phase error.

[0028] During stages T3 to T4, the optical path tuning engine receives the final calibration value and performs optical path tuning operations based on the final calibration value to calibrate the error of the corresponding optical computing unit.

[0029] Preferably, the step of calculating the final calibration amount based on the phase error includes:

[0030] Perform PID calculations on the phase error to obtain the PI control quantity;

[0031] Obtain the current environmental parameters and input them into the trained prediction model to obtain the feedforward compensation amount;

[0032] The final calibration value is obtained by summing the PI control value and the feedforward compensation value.

[0033] Preferably, the method further includes:

[0034] In stages T4 to T5, the AI ​​coprocessor uses the phase error and external environment parameters collected during this calibration as samples to input into the prediction model, and updates the weights of the prediction model using an online incremental learning method.

[0035] Beneficial Effects: This invention proposes an optical signal synchronization calibration device for an optical computing system. The optical computing system includes a chip and a board, with the chip coupled to the board and having multiple optical computing units. The device includes: an optical probe array, an AI coprocessor, an optical path tuning engine, and a low-speed electrical feedback line. The optical probe array is located at the four corners and the center of the chip, configured to detect the optical signals of the corresponding wavelength channels of the multiple optical computing units and convert the optical signals into electrical signals. The AI ​​coprocessor is located on the board and configured to drive the corresponding optical computing units to emit detection pulses according to pre-allocated time slot index variables, and to extract phase errors from the electrical signals and calculate the final calibration amount based on the phase errors. The optical path tuning engine is located on the board and configured to perform optical path tuning operations based on the final calibration amount to calibrate the errors of the corresponding optical computing units. The low-speed electrical feedback line is connected between the chip and the board and configured to transmit electrical signals to the AI ​​coprocessor and to transmit detection pulses to the chip. This application utilizes optical probe arrays deployed at the four corners and center of the chip to capture signals from optical computing units at different locations. These signals are then converted into electrical signals via photoelectric conversion, enabling on-chip in-situ detection. An AI coprocessor employs a time-slotted strategy to sequentially excite each unit, avoiding multi-channel signal overlap and interference. Based on phase error, dynamic calibration parameters are generated to drive an optical path tuning engine that adjusts the effective refractive index or physical deformation of the optical waveguide, achieving rapid board-level tuning. A low-speed electrical feedback line transmits probe pulses and feedback data between the chip and the board, forming a closed-loop control circuit to reduce power consumption while meeting signal transmission requirements. The optical signal synchronous calibration device provided in this application combines on-chip in-situ detection and rapid board-level tuning into an optical computing system. Through the collaborative efforts of spatially distributed detection and centralized board-level processing, real-time calibration in dynamic environments is achieved. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of this specification or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 A schematic diagram of the optical signal synchronization calibration device provided in this application;

[0038] Figure 2 The timing wavelength diagram corresponding to the time slot allocation table of the optical signal synchronization calibration method provided in this application.

[0039] Attached image captions:

[0040] 1. Chip;

[0041] 11. Optical probe array;

[0042] 12. Optical computing unit;

[0043] 2. Circuit board;

[0044] 21. AI coprocessor;

[0045] 22. Optical path tuning engine;

[0046] 23. Low-speed electrical feedback line. Detailed Implementation

[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0048] Example 1

[0049] like Figure 1-2 As shown, this embodiment provides an optical signal synchronization calibration device for an optical computing system. The optical computing system includes a chip and a board. The chip is coupled to the board and has multiple optical computing units. Specifically, the device includes: an optical probe array, an AI coprocessor, an optical path tuning engine, and a low-speed electrical feedback line. The optical probe array is located at the four corners and the center of the chip and is configured to detect the optical signals of the corresponding wavelength channels of the multiple optical computing units and convert the optical signals into electrical signals. The AI ​​coprocessor is located on the board and is configured to operate according to a pre-allocated time slot index variable λ. k The corresponding optical computing unit is driven to emit a probe pulse, and the phase error is extracted from the electrical signal and the final calibration value is calculated based on the phase error; the optical path tuning engine is deployed on the board and configured to perform optical path tuning operation based on the final calibration value to calibrate the error of the corresponding optical computing unit; the low-speed electrical feedback line is connected between the chip and the board and configured to transmit the electrical signal to the AI ​​coprocessor and the probe pulse to the chip.

[0050] Specifically, this invention proposes an optical signal synchronization calibration device for an optical computing system. The optical computing system includes a chip and a board, with the chip coupled to the board and having multiple optical computing units. The device includes: an optical probe array, an AI coprocessor, an optical path tuning engine, and a low-speed electrical feedback line. The optical probe array is located at the four corners and the center of the chip, configured to detect the optical signals of the corresponding wavelength channels of the multiple optical computing units and convert the optical signals into electrical signals. The AI ​​coprocessor is located on the board and configured to operate according to a pre-allocated time slot index variable λ. k The system drives the corresponding optical computing unit to emit probe pulses and extracts phase error from the electrical signal, calculating the final calibration value based on the phase error. An optical path tuning engine is deployed on the board and configured to perform optical path tuning operations based on the final calibration value to calibrate the error of the corresponding optical computing unit. A low-speed electrical feedback line connects the chip and the board, configured to transmit electrical signals to the AI ​​coprocessor and probe pulses to the chip. This application captures signals from optical computing units at different locations by deploying optical probe arrays at the four corners and center of the chip, generating electrical signals through photoelectric conversion to achieve on-chip in-situ detection. The AI ​​coprocessor uses a time-slotted strategy to orderly excite each unit, avoiding multi-channel signal overlap interference, generating dynamic calibration parameters based on the phase error, and driving the optical path tuning engine to adjust the effective refractive index or physical deformation of the optical waveguide to achieve rapid board-level tuning. The low-speed electrical feedback line transmits probe pulses and feedback data between the chip and the board, forming a closed-loop control circuit to reduce power consumption while meeting signal transmission requirements. The optical signal synchronous calibration device provided in this application combines on-chip in-situ detection and board-level rapid tuning and introduces them into an optical computing system. Through the collaboration of spatially distributed detection and board-level centralized processing, it achieves real-time calibration in dynamic environments.

[0051] Among them, the low-speed electrical feedback bus transmits the error information measured by the chip-level optical probe to the board-level AI coprocessor at a rate of less than 1 Gb / s, so as to reduce power consumption while meeting the signal transmission requirements, thereby achieving low-power, nanosecond-level closed-loop feedback.

[0052] At the board level, three main functional modules are set up: an optical path tuning engine, a low-speed electrical feedback bus, and an AI coprocessor. The optical path tuning engine adopts a dual mechanism of "silicon-based thermal tuning + piezoelectric tuning": the thermal tuning module provides a coarse tuning range of ±100 ps, ​​and the piezoelectric module provides a fine tuning range of ±10 ps. When cascaded, they can be continuously adjusted within 0–110 ps. The AI ​​coprocessor is based on the TinyML core, with an inference latency of less than 5 μs and an operating power consumption of less than 2 mW. Its function is to use a predictive model trained with historical drift data such as temperature and stress to output the feedforward compensation quantity AIpred in advance. It works in parallel with the classic PID controller, significantly reducing steady-state error and tuning power consumption.

[0053] Appendix Figure 2As a time-wavelength diagram, its core function is to visualize the time slot and wavelength matching relationship in optical signal synchronization calibration scenarios. It accurately presents the scheduling rhythm of different calibration signals in the "time-wavelength" dimension, ensuring that there are no signal conflicts and no timing deviations during the synchronization calibration process. It can directly assist technicians in verifying the feasibility of the synchronization calibration scheme (such as whether the calibration signal time slot is misaligned with the service signal), quickly locate the root cause of synchronization deviation (such as calibration failure caused by a wavelength time slot offset), and provide intuitive graphical evidence for the optimization of synchronization algorithms.

[0054] As one possible approach, each group of optical probes in the optical probe array includes a microring resonator and a photodiode. The microring resonator is used to filter and resonate optical signals of a specific wavelength, while the photodiode is used to convert the optical signal into an electrical signal. The microring resonator has a Q value (quality factor) of 8000 and a free spectral range (FSR) of 8 nm. One optical probe covers the signal detection of multiple optical computing units.

[0055] Specifically, a microring resonator is a resonant device with a ring waveguide structure, which can be implemented using a silicon-based microring structure. By adjusting the microring radius and coupling coefficient, wavelength-selective filtering is achieved, and its periodic resonance characteristics can enhance the intensity of the target wavelength light field. A photodiode is a semiconductor device that converts light energy into current, which can be implemented using a Ge-Si photodiode structure. It linearly converts the optical signal output from the microring resonator into a voltage signal. This achieves enhanced capture and linear conversion of a specified wavelength light signal, avoiding spectral overlap interference between multiple channels.

[0056] The chip has NxM optical computing units. This application uses a combination of sparse deployment and global coverage to deploy the optical probe array. Specifically, the optical probe array is arranged at key positions such as the four corners and the center of the chip. Each probe collects the output optical signals of several neighboring optical computing units through a coupling waveguide or power beam splitter, thereby realizing the monitoring of the entire NxM optical computing unit matrix.

[0057] One possible approach is to obtain the phase error by taking the phase angle of the output voltage of the photodiode.

[0058] Specifically, the output voltage of a photodiode is the voltage waveform formed after the optical probe converts the received optical signal into an electrical signal. The phase characteristics of this voltage waveform maintain a linear correspondence with the phase of the incident optical signal.

[0059] As one possible approach, the AI ​​coprocessor is also used to extract power error from electrical signals. The power error is obtained by squaring the effective value of the output voltage of the photodiode. The power error is mainly used for link health monitoring and threshold discrimination. By continuously collecting power error data, the signal integrity of the link (such as attenuation, noise interference, etc.) is dynamically reflected, and the current working status of the link is presented intuitively to ensure the stability of data transmission / energy transfer. When the power error exceeds the safety threshold, a graded response (such as yellow alarm, red shutdown) is immediately triggered to prevent the fault from spreading.

[0060] As one possible approach, the final calibration values ​​include wide-range calibration values ​​and precision calibration values. The optical path tuning engine includes a silicon-based thermal tuning module and a piezoelectric tuning module. The silicon-based thermal tuning module drives a micro-thermal wire array based on the wide-range calibration value, changing the effective refractive index of the micro-ring resonator through the thermo-optic effect to achieve a large-path coarse tuning within a range of ±100 ps. The piezoelectric tuning module controls the deformation of the piezoelectric ceramic based on the precision calibration value, adjusting the optical path of the micro-ring resonator through deformation to achieve a high-resolution fine tuning within a range of ±10 ps.

[0061] Specifically, the wide-range calibration parameter is used for adjusting the optical path over a wide range. It can be achieved by using a silicon-based thermally tuned module to drive a micro-thermal wire array to generate a temperature gradient. This temperature gradient alters the effective refractive index of the micro-ring resonator through the thermo-optical effect, thus enabling coarse adjustment of the optical path to cover a large range of errors. Its function is to quickly eliminate the main phase error components caused by temperature drift or stress deformation. The precision calibration parameter is used for fine-tuning the optical path with high precision. It can be achieved by using a piezoelectric tuning module to control the deformation of the piezoelectric ceramic. This mechanical deformation directly changes the physical length of the micro-ring resonator, thus achieving nanometer-level compensation of the optical path. Its function is to eliminate residual errors and adapt to dynamic environmental changes. The silicon-based thermally tuned module is a tuning unit based on the thermo-optical effect. It can be achieved by using a micro-thermal wire array integrated around the micro-ring resonator. Current-driven local temperature changes alter the refractive index of the silicon material, enabling wide-range, low-power coarse adjustment of the optical path. The piezoelectric tuning module is a tuning unit based on mechanical deformation. Specifically, it can use piezoelectric ceramic material attached to the surface of a microring resonator, which generates nanoscale deformation through voltage control. Its function is to achieve high-precision, fast-response fine-tuning of the optical path. The two tuning mechanisms utilize indirect control of refractive index and direct adjustment of physical dimensions, respectively, avoiding the range limitations or insufficient precision of a single tuning method.

[0062] In other words, after the optical path tuning engine receives the final calibration value, the silicon-based thermal tuning module first completes a large-range coarse adjustment, then the piezoelectric tuning module completes a small-range fine adjustment, and finally the optical path is latched to a stable state to achieve closed-loop correction.

[0063] As one possible approach, based on pre-allocated time slot index variables Drive the corresponding optical computing unit to emit probe pulses, including:

[0064] The AI ​​coprocessor transmits calibration trigger frames to the chip level via a low-speed electrical feedback line. The calibration trigger frame contains a calibration start command and a time slot allocation table of optical computing units and wavelengths.

[0065] After receiving the calibration trigger frame, each optical computing unit at the chip level calculates the corresponding time slot index variable according to the time slot allocation table. Probe pulses are emitted sequentially.

[0066] Specifically, the AI ​​coprocessor generates a calibration trigger frame containing a calibration start command and a time slot allocation table, and transmits it to the chip level via a low-speed electrical feedback line. Upon receiving the calibration trigger frame, each optical computing unit at the chip level parses the time slot allocation table and extracts its corresponding time slot index variable. Each optical computing unit, according to... The probe pulses are emitted in the numerical order of their respective values ​​within their assigned independent time windows. For example, when... When the value is an integer from 1 to N, the optical computing unit numbered k will initiate pulse transmission in the k-th time slot window. This time-division triggering mechanism ensures that the detection pulses of different units do not overlap on the time axis, avoiding mutual interference between multiple optical signals on the transmission path. The time slot index variable λ... k This is an independent time window identifier assigned to each optical computing unit, used to isolate the probe pulse emission actions of different units in the time dimension. The calibration trigger frame is a communication data packet containing calibration control commands and resource allocation information, used to transmit global synchronization signals between the board and the chip. The time slot allocation table is configuration information that records the correspondence between optical computing units and wavelength channels, used to establish the mapping rules between optical computing units and probe time slots.

[0067] One possible approach is to calculate the final calibration value based on the phase error, including:

[0068] Perform PID calculations on the phase error to obtain the PI control input;

[0069] Obtain the current environmental parameters, input the current environmental parameters into the trained prediction model, and obtain the feedforward compensation amount;

[0070] The final calibration value is obtained by summing the PI control value and the feedforward compensation value.

[0071] Specifically, the phase error is input to the PID controller for proportional-integral calculation, generating a real-time corrected PI control input. Simultaneously, environmental parameters are synchronously input into a pre-trained predictive model, outputting a feedforward compensation input correlated with environmental disturbances. The two control inputs are linearly superimposed at the summation node to form the final calibration output. This process corrects the current error in real time through a closed-loop feedback mechanism, while simultaneously using a feedforward mechanism to pre-compensate for sudden environmental changes, achieving composite control under dynamic conditions.

[0072] PID control is a method that processes error signals using a proportional-integral-derivative algorithm. In this application, a combination of proportional and integral terms is used. The proportional and integral coefficients can be tuned according to the dynamic characteristics of the system to quickly respond to phase deviations and eliminate steady-state errors.

[0073] The current environmental parameters are real-time monitored temperature and stress data, which can be acquired using onboard temperature sensors and strain gauges to characterize the dynamic impact of the external environment on the optical waveguide path length. The temperature sensor and strain gauge are integrated on the board.

[0074] The predictive model is a machine learning model trained on historical data. Specifically, it can be implemented using neural networks or support vector machines. It is used to predict the trend of optical path change based on environmental parameters and generate compensation amounts.

[0075] The feedforward compensation is a pre-adjustment value based on the prediction of environmental parameters. Specifically, it can be superimposed on the value output by the model inference and the PI control quantity to offset the phase shift caused by environmental disturbances in advance.

[0076] As one feasible approach, the formula for calculating the final calibration value Δτ is:

[0077] Δτ=Δτ PID + AIpred;

[0078] Δτ PID =K p ·E φ +K i ·∫E φ dt;

[0079] Where, Δτ PID E is the PI control input, AIpred is the feedforward compensation input. φ For phase error, K p K is the proportional control coefficient. i This is the integral control coefficient.

[0080] The specific explanation is as follows: Δτ PIDIt is a closed-loop control quantity generated through proportional-integral operations, used to quickly respond to phase deviations and eliminate steady-state errors; AIpred is a feedforward compensation quantity predicted by a machine learning model, specifically a trained neural network model, which outputs a predicted compensation value after inputting environmental parameters such as temperature and stress, used to proactively offset the impact of environmental disturbances on the optical path; E φ It is the phase error, used to characterize the actual phase deviation of the optical signal transmission path; K p This refers to the proportional control coefficient, which can be set according to the system response speed requirements to adjust the instantaneous compensation intensity for phase errors; K i It refers to the integral control coefficient, which can be set according to the steady-state error tolerance and is used to adjust the compensation intensity for the accumulated amount of historical phase error.

[0081] To verify the engineering applicability of this invention, a 64-channel optical convolution accelerator prototype was implemented on a TSMC 45 nm CMOS process and silicon photonics hybrid integration platform. The chip internally features four micro-ring probe arrays, each covering 16 computational units. The board-level optical path tuning engine employs a dual-mechanism tuner scheme, with thermally tuned and piezoelectric tuners cascaded through a 3 dB coupler. The experimental performance targets were phase error < ±0.05 rad and power imbalance < 0.5 dB.

[0082] At a room temperature of 25 °C, the system's setup time from cold start to synchronization completion was 680 ns, with a total steady-state power consumption of 38 mW (including the AI ​​coprocessor). Subsequent temperature cycling tests were conducted: continuous operation for 72 hours at an ambient temperature of 70 °C, during which the RMS value of phase drift was only 0.018 rad, and the power imbalance remained within 0.4 dB, requiring no manual intervention throughout. This prototype has now been deployed in a data center test rack and has operated stably for 30 days under real-world business load conditions, further demonstrating the reliability, real-time performance, and low power consumption advantages of this invention in large-scale optical computing scenarios.

[0083] Example 2

[0084] This invention also provides an optical signal synchronization calibration method, applied to the optical signal synchronization calibration device as described in any one of Embodiment 1. The method provided in this embodiment specifically includes the following steps (1) to (4):

[0085] Step (1): In the T0 to T1 stage, the AI ​​coprocessor transmits the calibration trigger frame to the chip level. The calibration trigger frame contains the calibration start instruction and the time slot allocation table of each optical computing unit and wavelength.

[0086] The calibration trigger frame is a communication data packet containing calibration control instructions and resource allocation information. Specifically, it can be encapsulated using the Ethernet protocol and transmitted to the chip level via a low-speed electrical feedback line to achieve centralized scheduling of the calibration process.

[0087] Step (2): In stages T1 to T2, after each optical computing unit at the chip level receives the calibration trigger frame, it allocates the corresponding time slot index variable λ according to the time slot allocation table. k Probe pulses are emitted sequentially;

[0088] Among them, the time slot index variable λ k It is used to identify the timing number of the probe pulse emitted by the optical computing unit. By binding different optical computing units with specific time periods through a pre-allocated time slot table, signal crosstalk caused by concurrent operation of multiple units is avoided.

[0089] Step (3): In the T2 to T3 stage, the optical probe array detects the optical signals of the corresponding wavelength channels of multiple optical computing units and converts the optical signals into electrical signals. The AI ​​coprocessor extracts the phase error from the electrical signal and calculates the final calibration amount based on the phase error.

[0090] The phase error is obtained by taking the phase angle of the photodiode's output voltage and used as an input parameter for closed-loop control. The final calibration value refers to the compensation value used to adjust the optical path. It can be generated by superimposing the PID control algorithm and the feedforward prediction model, and includes both coarse and fine adjustment components to adapt to calibration requirements under different environmental disturbances.

[0091] Step (4): In the T3 to T4 stage, the optical path tuning engine receives the final calibration amount and performs optical path tuning operation based on the final calibration amount to calibrate the error of the corresponding optical computing unit.

[0092] As one possible approach, the calculation of the final calibration quantity based on phase error involved in step (3) above includes:

[0093] Perform PID calculations on the phase error to obtain the PI control input;

[0094] Obtain the current environmental parameters, input the current environmental parameters into the trained prediction model, and obtain the feedforward compensation amount;

[0095] The final calibration value is obtained by summing the PI control value and the feedforward compensation value.

[0096] Specifically, the phase error is input to the PID controller for proportional-integral calculation, generating a real-time corrected PI control input. Simultaneously, environmental parameters are synchronously input into a pre-trained predictive model, outputting a feedforward compensation input correlated with environmental disturbances. The two control inputs are linearly superimposed at the summation node to form the final calibration output. This process corrects the current error in real time through a closed-loop feedback mechanism, while simultaneously using a feedforward mechanism to pre-compensate for sudden environmental changes, achieving composite control under dynamic conditions.

[0097] PID control is a method that processes error signals using a proportional-integral-derivative algorithm. In this application, a combination of proportional and integral terms is used. The proportional and integral coefficients can be tuned according to the dynamic characteristics of the system to quickly respond to phase deviations and eliminate steady-state errors.

[0098] The current environmental parameters are real-time monitored temperature and stress data, which can be collected using onboard temperature sensors, strain gauges or accelerometers to characterize the dynamic influence of the external environment on the optical waveguide path length.

[0099] The predictive model is a machine learning model trained on historical data. Specifically, it can be implemented using neural networks or support vector machines. It is used to predict the trend of optical path change based on environmental parameters and generate compensation amounts.

[0100] The feedforward compensation is a pre-adjustment value based on the prediction of environmental parameters. Specifically, it can be superimposed on the value output by the model inference and the PI control quantity to offset the phase shift caused by environmental disturbances in advance.

[0101] As one possible approach, after step (4), the method of this embodiment further includes the following step (5):

[0102] Step (5): In the T4 to T5 stage, the AI ​​coprocessor uses the phase error and external environment parameters collected in this calibration as samples to input into the prediction model, and updates the weights of the prediction model using an online incremental learning method.

[0103] Specifically, online incremental learning is a learning method that dynamically adjusts model parameters using new samples without retraining the entire model. Weight updates refer to the parameter optimization process of adjusting the connection strength of neurons in the prediction model. This can be achieved through iterative calculations using backpropagation or adaptive moment estimation algorithms, enabling the prediction model to continuously track changes in the environment. After a single calibration operation, the AI ​​coprocessor uses the phase error data collected during the current calibration cycle and the synchronously recorded environmental parameters to form a training sample set. This sample set is input into the prediction model, where the model weights are fine-tuned using an online incremental learning algorithm. During this process, the model retains the statistical characteristics of historical training data while adjusting the internal parameter mapping relationship based on the distribution characteristics of the new samples. This dynamic update mechanism allows the prediction model to continuously adapt to long-term drift of environmental variables, avoiding the accumulation of compensation calculation biases caused by the solidification of model parameters.

[0104] In summary, the transmission of the calibration trigger frame enables global triggering and time slot allocation for the calibration task, through a predefined time slot index variable λ. k By distributing the probe pulse emission of the large-scale optical computing unit across different time periods, the resource contention problem caused by multi-channel concurrent operation is resolved. During the detection phase, the optical probe array performs in-situ conversion of the optical signals in the wavelength channels, transforming the optical domain signals into electrical signals. Combined with real-time phase error extraction by the AI ​​coprocessor, an electrical domain closed-loop control loop is formed, significantly reducing signal processing latency. The final calibration calculation integrates the steady-state accuracy of PID control and the environmental adaptability of the prediction model. This is achieved through the coordinated execution of thermal tuning by the optical path tuning engine and the piezoelectric tuning module, compensating for wide-range deviations caused by temperature drift and correcting subtle phase shifts caused by stress deformation. The phased, time-sequential operation decomposes the calibration process into four stages: triggering, detection, calculation, and execution. A balance between low latency and high accuracy is achieved through an electro-optic hybrid control architecture.

[0105] As one possible approach, the AI ​​coprocessor is also used to extract power error from electrical signals. The power error is obtained by squaring the effective value of the output voltage of the photodiode. The power error is mainly used for link health monitoring and threshold discrimination. By continuously collecting power error data, the signal integrity of the link (such as attenuation, noise interference, etc.) is dynamically reflected, and the current working status of the link is presented intuitively to ensure the stability of data transmission / energy transfer. When the power error exceeds the safety threshold, a graded response (such as yellow alarm, red shutdown) is immediately triggered to prevent the fault from spreading.

[0106] As one feasible approach, the formula for calculating the final calibration value Δτ is:

[0107] Δτ=Δτ PID + AIpred;

[0108] Δτ PID =K p ·E φ +K i ·∫E φ dt;

[0109] Where, Δτ PID E is the PI control input, AIpred is the feedforward compensation input. φ For phase error, K p K is the proportional control coefficient. i This is the integral control coefficient.

[0110] This application supports parallel operation of up to 256 nodes and 64 wavelengths; as the scale continues to expand, it can be linearly scaled up through a domain-based recursive synchronization method.

[0111] The distributed calibration method provided in this application has a complete calibration cycle of less than 1 μs and is completely transparent to upper-layer computing tasks. It supports parallel operation of up to 256 nodes and 64 wavelengths; as the scale continues to increase, it can be linearly scaled through a domain-based recursive synchronization method.

[0112] In the embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between systems or units may be electrical, mechanical, or other forms.

[0113] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0114] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0115] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0116] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An optical signal synchronization calibration device for an optical computing system, characterized in that, The optical computing system includes a chip and a board, the chip being coupled to the board, and the chip having multiple optical computing units. The device includes: An optical probe array is deployed at the four corners and the center of the chip, configured to detect optical signals of the corresponding wavelength channels of multiple optical computing units and convert the optical signals into electrical signals; An AI coprocessor, deployed on the board, is configured to drive the corresponding optical computing unit to emit detection pulses according to a pre-allocated time slot index variable, and to extract phase error from the electrical signal and calculate the final calibration amount based on the phase error. An optical path tuning engine, deployed on the board, is configured to perform optical path tuning operations based on the final calibration value to calibrate the error of the corresponding optical computing unit; A low-speed electrical feedback line is connected between the chip and the board, configured to transmit the electrical signal to the AI ​​coprocessor and the detection pulse to the chip; The calculation of the final calibration amount based on the phase error includes: Perform PID calculations on the phase error to obtain the PI control quantity; Obtain the current environmental parameters and input them into the trained prediction model to obtain the feedforward compensation amount; The final calibration value is obtained by summing the PI control value and the feedforward compensation value. The formula for calculating the final calibration value Δτ is: Δt=Δt PID + AIpred; Δτ PID =K p ·E φ +K i ·∫E φ dt; Wherein, the Δτ PID The PI control variable is AIpred, which is the feedforward compensation variable, and E is... φ For phase error, the K p K is the proportional control coefficient; i This is the integral control coefficient.

2. The optical signal synchronization calibration device for an optical computing system according to claim 1, characterized in that, Each optical probe in the optical probe array is composed of a micro-ring resonator and a photodiode. The micro-ring resonator is used to filter and resonate optical signals of a specific wavelength, and the photodiode is used to convert the optical signal into an electrical signal.

3. The optical signal synchronization calibration device for the optical computing system according to claim 2, characterized in that, The phase error is obtained by taking the phase angle of the output voltage of the photodiode.

4. The optical signal synchronization calibration device for the optical computing system according to claim 3, characterized in that: The final calibration value includes a wide-range calibration value and a precision calibration value. The optical path tuning engine includes a silicon-based thermal tuning module and a piezoelectric tuning module. The silicon-based thermal tuning module drives a micro-thermal wire array based on the wide-range calibration value, changing the effective refractive index of the micro-ring resonator through the thermo-optic effect. The piezoelectric tuning module controls the deformation of the piezoelectric ceramic based on the precision calibration value, adjusting the optical path of the micro-ring resonator through deformation.

5. The optical signal synchronization calibration device for an optical computing system according to claim 1, characterized in that, The step of driving the corresponding optical computing unit to emit a probe pulse according to the pre-allocated time slot index variable includes: The AI ​​coprocessor transmits a calibration trigger frame to the chip level via the low-speed electrical feedback line. The calibration trigger frame includes a calibration start command and a time slot allocation table of optical computing units and wavelengths. After receiving the calibration trigger frame, each optical computing unit at the chip level transmits probe pulses sequentially according to the corresponding time slot index variable allocated in the time slot allocation table.

6. A method for optical signal synchronization calibration, characterized in that, The method, applied to the optical signal synchronization calibration apparatus as described in any one of claims 1-5, comprises: During the T0 to T1 phase, the AI ​​coprocessor transmits a calibration trigger frame to the chip level. The calibration trigger frame contains a calibration start command and a time slot allocation table for each optical computing unit and wavelength. During the T1 to T2 phase, after receiving the calibration trigger frame, each optical computing unit at the chip level sequentially transmits probe pulses according to the time slot index variable allocated in the time slot allocation table. In stages T2 to T3, the optical probe array detects the optical signals of the wavelength channels corresponding to the multiple optical computing units and converts the optical signals into electrical signals. The AI ​​coprocessor extracts the phase error from the electrical signals and calculates the final calibration amount based on the phase error. During the T3 to T4 phases, the optical path tuning engine receives the final calibration amount and performs optical path tuning operations based on the final calibration amount to calibrate the error of the corresponding optical computing unit. The calculation of the final calibration amount based on the phase error includes: Perform PID calculations on the phase error to obtain the PI control quantity; Obtain the current environmental parameters and input them into the trained prediction model to obtain the feedforward compensation amount; The final calibration value is obtained by summing the PI control value and the feedforward compensation value. The formula for calculating the final calibration value Δτ is: Δt=Δt PID + AIpred; Δτ PID =K p ·E φ +K i ·∫E φ dt; Where, Δτ PID E is the PI control input, AIpred is the feedforward compensation input. φ For phase error, K p K is the proportional control coefficient. i This is the integral control coefficient.

7. The optical signal synchronization calibration method according to claim 6, characterized in that, The method further includes: In stages T4 to T5, the AI ​​coprocessor uses the phase error and external environment parameters collected during this calibration as samples to input into the prediction model, and updates the weights of the prediction model using an online incremental learning method.

Images