A Predictive Tuning and Virtual Calibration Method for Optical Modules Based on Digital Twins

Abstract

This invention relates to the field of optical communication technology and discloses a predictive debugging and virtual calibration method for optical modules based on digital twins. The steps are as follows: A digital twin model is constructed by fusing the physical characteristics of the optical module with a neural network; the parameters of the digital twin model are fine-tuned using sample measured data to construct a personalized digital twin model; the full working scenario of the optical module is simulated in a digital environment, and the personalized digital twin model generates virtual test samples, which are then virtually debugged; the parameter combination with the highest comprehensive scoring function is selected from the virtual test samples that pass the virtual debugging as the optimal driving parameters for virtual calibration; the physical testing system, in conjunction with the optimal driving parameters, determines whether the measured performance is qualified. If qualified, a test report is generated; if unqualified, the virtual debugging and virtual calibration iterations are repeated. This invention significantly reduces the time spent on testing equipment and the corresponding material consumption, improving the utilization rate of production line resources.

Description

Technical Field

[0001] This invention relates to the field of optical communication technology, and in particular to a predictive debugging and virtual calibration method for optical modules based on digital twins. Background Technology

[0002] In the production and R&D of optical modules, testing and debugging are crucial steps to ensure their performance meets standards. Currently, the industry mainly uses traditional methods such as fixed-process testing, experience-based debugging, and traversal scanning. While these methods can complete module performance verification and parameter adjustment to a certain extent, they have gradually revealed many shortcomings in practical applications, hindering further improvements in production efficiency and quality.

[0003] The fixed-process testing method uses a pre-set fixed test sequence to perform the same test steps on all optical modules. The test parameter combination is fixed during the test and cannot be adjusted according to the individual differences of the modules. This results in a long test time, low efficiency, and a large number of unnecessary test items.

[0004] The experience-based debugging method relies mainly on the engineer's experience to debug parameters. During the process, a trial-and-error method or a simple binary search strategy is used. The debugging results are greatly affected by human factors and have poor consistency.

[0005] The traversal scanning method scans the entire range of parameters such as voltage, temperature, and current during the testing process. Although it covers all test items, it takes a very long time and cannot meet the high efficiency requirements of large-scale production.

[0006] The aforementioned problems collectively led to serious waste of resources, with testing equipment being occupied by inefficient processes for extended periods, resulting in rising production costs. At the same time, frequent physical debugging and testing not only increased energy consumption but also affected the overall throughput capacity of the production line. Summary of the Invention

[0007] The purpose of this invention is to address the technical problems of low efficiency, slow convergence speed, and resource waste in existing optical module testing and debugging technologies by providing a predictive debugging and virtual calibration method for optical modules based on digital twins.

[0008] To achieve the above-mentioned objectives, the embodiments of the present invention provide the following technical solutions:

[0009] A digital twin model is constructed by fusing the physical characteristics of optical modules with neural networks;

[0010] By fine-tuning the parameters of the digital twin model using sample measurement data, a personalized digital twin model can be constructed.

[0011] Simulate the full working scenario of the optical module in a digital environment, generate virtual test samples using a personalized digital twin model, and perform virtual debugging on the virtual test samples;

[0012] The parameter combination with the highest comprehensive scoring function is selected from the virtual test samples that have passed virtual debugging and used as the optimal driving parameters for virtual calibration.

[0013] The physical testing system, in conjunction with the optimal driving parameters, determines whether the measured performance is qualified. If qualified, a test report is generated; if not qualified, the virtual debugging and virtual calibration iterations are repeated.

[0014] To address the technical problem of low testing and debugging efficiency of optical modules, this application solves the technical problems of traditional fixed-process testing methods being unable to skip unnecessary test items and the time-consuming traversal scanning method by constructing a personalized digital twin model and simulating and virtually debugging the entire working scenario in a digital environment. The generation and debugging of virtual test samples in this application replaces most of the physical trial and error process, enabling the testing process to focus on key verification links, thereby significantly shortening the overall testing time.

[0015] To address the technical problems of slow convergence speed and resource waste in optical module testing and debugging, this application utilizes a digital twin model to perform rapid parameter optimization and virtual calibration in a virtual space, and outputs the optimal driving parameters to guide physical testing. This solves the technical problems of experience-based debugging methods relying on manual trial and error, inconsistent convergence, and long equipment downtime. This application places the exploration of massive parameter combinations within an efficient digital simulation, achieving intelligent guidance and rapid convergence in the debugging process, while significantly reducing the downtime and energy consumption of physical testing equipment, thus lowering production costs.

[0016] Compared with existing technologies, the beneficial effects of this invention are as follows: By completing a large number of simulation tests and parameter optimizations in a virtual environment using a digital twin model, physical testing is compressed to the final verification stage, breaking the constraints of fixed processes and effectively avoiding the time waste caused by full-range scanning, thus adapting to the high-paced requirements of large-scale production; Model-based virtual debugging can simulate the entire working scenario, intelligently and quickly explore the parameter space and locate the optimal solution, replacing experience-based manual trial and error, solving the problems of blind and slow debugging processes, and achieving optimization of the debugging path and stable convergence of results; Fine-tuning the basic model using measured data to construct a personalized digital twin reflecting individual differences allows virtual debugging and calibration to be performed on the characteristics of specific modules, overcoming the shortcomings of traditional methods that ignore individual differences, thereby uncovering the best performance of each module; Time-consuming and energy-intensive testing and debugging work is transferred to the digital domain to the greatest extent, greatly reducing the time spent on expensive testing equipment and the corresponding material consumption, lowering overall production costs, and improving the utilization rate of production line resources.

[0017] Furthermore, the predictive debugging and virtual calibration method for optical modules based on digital twins includes the following sub-steps for constructing the digital twin model:

[0018] Based on the physical characteristics of the core components of the optical module, a physical equation module is constructed to provide a digital twin model of basic performance constraints;

[0019] A neural network model is introduced to compensate for the prediction error of the physical equation module;

[0020] Train a neural network model, and then combine the trained neural network model with the compensated physical equation module to construct a digital twin model.

[0021] In the aforementioned solutions, existing technologies typically employ either pure physical equation modeling or pure data-driven modeling when constructing performance models for optical modules used for debugging and calibration. Pure physical models rely on idealized assumptions and struggle to accurately characterize nonlinear errors introduced by manufacturing tolerances, batch material variations, and environmental disturbances. Pure data models, on the other hand, require massive amounts of labeled data for training, which can easily lead to overfitting or insufficient generalization in the early stages of production when samples are scarce or when working with individual modules. This results in models with limited fidelity, failing to accurately reflect the true dynamic characteristics of a specific module. Inaccurate models undermine the reliable foundation for subsequent virtual debugging and calibration, rendering optimization results in the digital space ineffective in guiding physical testing. This necessitates extensive physical trial and error for correction, severely limiting overall efficiency and wasting resources. This application integrates physical characteristics with neural networks to construct a physical equation module that provides fundamental constraints. The neural network module dynamically compensates for errors in the digital twin model. Based on the physical principles of the device, an equation module with clear physical meaning is established, laying the foundation for the model's basic framework and constraints. Subsequently, a neural network is introduced to specifically learn and compensate for residual errors caused by simplification assumptions and unmodeled factors in the physical model. This solves the technical problem of unreliable virtual debugging results due to insufficient accuracy of the basic model. Through neural network compensation, this application enables the digital twin model to absorb limited measured data, accurately approximating the individual characteristics of specific optical modules, significantly improving model fidelity. This ensures that full-scene simulations and parameter optimizations performed in the digital environment highly map the real behavior of the physical world, making the conclusions of virtual debugging highly valuable for guidance.

[0022] Furthermore, in the predictive debugging and virtual calibration method for optical modules based on digital twins, the processing formula of the neural network model is as follows:

[0023] ;

[0024] in, This is the weight matrix of the first hidden layer, with a shape of 16×3. This is the weight matrix for the second hidden layer, with a shape of 8×16. This is the weight matrix of the output layer, with a shape of 3×8. This is the paranoia vector of the first hidden layer, with a shape of 16×1. This is the paranoia vector of the second hidden layer, with a shape of 8×1. This is the bias vector for the output layer, with a shape of 3×1. , , The initial value is set to 0. The ReLU activation function for the first hidden layer. The ReLU activation function for the second hidden layer. It is the Sigmoid function of the output layer.

[0025] In the aforementioned solutions, existing technologies often face key challenges in model structure design when using neural networks to compensate for physical models. If the network structure is too simple (such as a single-layer linear model), its nonlinear expressive power is insufficient, making it difficult to effectively compensate for complex errors. If the structure is too complex or the activation function is inappropriately chosen, overfitting is likely to occur with limited samples, resulting in poor model generalization ability and unstable predictions. This inherent defect of the compensation model directly affects the fused digital twin, limiting its ability to approximate the characteristics of individual modules and creating a bottleneck in accuracy improvement. This application employs a hidden layer structure with the ReLU activation function to learn and map complex error patterns not covered by the physical model in a hierarchical and nonlinear manner. The output layer uses the Sigmoid function to constrain the final compensation amount within a reasonable range, ensuring output stability. The dimensions of the weight matrix and bias vector are specifically designed to ensure sufficient expressive power while adapting to small-sample training scenarios by controlling the parameter scale, preventing overfitting. This solves the technical problem of the compensation neural network's inability to balance nonlinear fitting ability and generalization stability with limited data. This application ensures that the neural network can robustly learn reliable error patterns from a small amount of measured data, thereby achieving stable, efficient and accurate integration with the physical equation module. This provides a reliable data-driven component for building high-fidelity personalized digital twin models and solidifies the accuracy foundation of the virtual debugging process.

[0026] Furthermore, the predictive debugging and virtual calibration method for optical modules based on digital twins includes the following steps in constructing a personalized digital twin model:

[0027] Three environmental conditions were selected: extremely cold temperature, normal temperature, and extremely hot temperature. Different sets of typical driving conditions were selected under each working environment. Output performance data were collected under each driving condition, and several performance data constituted a sample set.

[0028] The parameters of the second hidden layer and output layer of the neural network model are fine-tuned and trained using the fine-tuning target of the sample set. The fine-tuned neural network model is then combined with the compensated physical equation module to obtain a personalized digital twin model.

[0029] Furthermore, the predictive debugging and virtual calibration method for optical modules based on digital twins, wherein generating virtual test samples and performing virtual debugging on the virtual test samples includes the following sub-steps:

[0030] Based on the industry standard of optical modules, the parameter range and step size of virtual testing are determined, virtual test samples are generated, and a parallel computing framework is used to perform batch prediction on the virtual samples to obtain prediction performance indicators.

[0031] The predictive performance metrics of all virtual test samples are evaluated, and qualified parameter sets that meet the requirements of different application scenarios are selected.

[0032] In the aforementioned solutions, existing technologies typically collect data under limited and singular operating conditions for the personalized fine-tuning and virtual testing of optical modules. This results in a lack of environmental robustness in the fine-tuned model, and unreliable prediction results under boundary conditions. Furthermore, in the virtual testing phase, simulations involving massive parameter combinations often employ serial computation, which is slow, leading to a long cycle from model construction to obtaining effective debugging conclusions and high computational resource consumption. This application systematically collects data under three typical environmental conditions—extremely cold, normal, and extremely hot—to construct a sample set capable of representing the entire operating boundary. It then performs targeted fine-tuning only on key layers of the neural network, thereby efficiently constructing a high-fidelity and environmentally adaptable personalized model. Subsequently, test samples are automatically generated according to industry standards, and a parallel computing framework is used for batch performance prediction and rapid automatic judgment. This achieves efficient exploration and qualified set screening of the massive virtual parameter space, solving the two core technical problems directly affecting overall efficiency: low model personalization efficiency, poor environmental adaptability, and excessively long virtual testing computation and screening time. The layered fine-tuning strategy of this application achieves reliable personalization with a small amount of data, and the parallel computing framework transforms virtual testing from a bottleneck into an efficient process. This ensures high efficiency and reliability across the entire process from personalized modeling to virtual debugging, laying a solid foundation for quickly and accurately locating the optimal driving parameters. It is the key to significantly shortening the overall debugging cycle and reducing dependence on physical testing resources.

[0033] Furthermore, in the predictive debugging and virtual calibration method for optical modules based on digital twins, the generation of virtual test samples includes the following sub-steps:

[0034] Virtual test samples are generated by combining the parameter ranges and step sizes of temperature range, bias current, TEC current, and modulation current.

[0035] The temperature range is -40℃ to 85℃, with a step size of 5℃, and a total of 26 temperature points;

[0036] The bias current range is 5mA~30mA, with a step size of 1mA, for a total of 26 current points;

[0037] The TEC current range is -1A to 1A, with a step size of 0.1A, and a total of 21 current points;

[0038] The modulation current range is 10mA~40mA, with a step size of 2mA, and a total of 16 current points.

[0039] Furthermore, the application scenario requirements for the predictive debugging and virtual calibration method for optical modules based on digital twins are as follows:

[0040] Data center scenario: Transmitted optical power Pout∈[1dBm, 3dBm], extinction ratio ER≥12dB, emission dispersion eye diagram closed quadrature TDECQ≤8ps;

[0041] 5G base station scenario: transmit optical power Pout∈[0dBm, 5dBm], extinction ratio ER≥10dB, transmit dispersion eye diagram closed four-phase TDECQ≤10ps;

[0042] Industrial control scenario: Emitted optical power Pout∈[0.5dBm, 4dBm], extinction ratio ER≥11dB, emission dispersion eye diagram closed four-phase TDECQ≤9ps.

[0043] In the aforementioned solutions, existing technologies often use coarse-grained parameter exploration space definitions during virtual debugging, and performance qualification standards are typically singular and disconnected from application scenarios, resulting in virtual test results that cannot accurately guide actual production. This application solves the problems of imprecise virtual parameter exploration and disconnect between qualification standards and practical applications by combining a refined multi-dimensional parameter space grid definition with a multi-scenario differentiated performance index threshold system. First, scanning rules covering the entire working range and with fine step sizes are set for key driving variables such as temperature and current, constructing a high-resolution parameter simulation grid to ensure the comprehensiveness and continuity of virtual testing. Second, specific thresholds for core performance indicators are defined for three typical scenarios: data centers, 5G base stations, and industrial control, forming quantitative judgment standards strictly linked to applications. The refined grid of this application avoids parameter blind spots, and the differentiated standards ensure the practicality of the debugging target, upgrading virtual debugging from coarse simulation to high-fidelity, scenario-based, and precise pre-verification, thereby significantly improving the reliability and first-pass yield of virtual calibration results and reducing physical testing rework.

[0044] Furthermore, the predictive debugging and virtual calibration method for optical modules based on digital twins includes the following sub-steps for selecting the parameter combination with the highest comprehensive scoring function:

[0045] The variance of emitted optical power Pout, extinction ratio ER, and emission dispersion eye diagram closed four-phase TDECQ over the entire temperature range is calculated using the following formula:

[0046] ;

[0047] in, For the variance of the fluctuation, Let Pout be the emitted optical power, ER be the extinction ratio, and TDECQ be the values ​​of the closed four-phase emission dispersion eye diagram at temperature t. The values ​​are the average of emitted light power Pout, extinction ratio ER, and emission dispersion eye diagram closed four-phase TDECQ over the entire temperature range;

[0048] The formulas for calculating the changes in emitted optical power Pout, extinction ratio ER, and emission dispersion eye diagram of a closed four-phase TDECQ under bias current fluctuations of ±1mA and modulation current fluctuations of ±2mA are as follows:

[0049] ;

[0050] in, The values ​​represent the emitted optical power Pout, extinction ratio ER, and emission dispersion eye diagram closed four-phase TDECQ under rated current. This represents the maximum change in performance when the bias current fluctuates by ±1mA or the modulation current by ±2mA. This is the maximum value after the current fluctuation. This is the minimum value after current fluctuation;

[0051] The formula for calculating the operating power consumption of an optical module is:

[0052] ;

[0053] in, The operating power consumption of the optical module, This is the bias current output by the laser diode equation submodule. The TEC current output by the TEC temperature control unit. The modulation current is the modulated current of the modulator equation submodule. , , ;

[0054] A comprehensive scoring function is obtained by considering temperature stability, current stability, and operating power consumption. The formula is as follows:

[0055] ;

[0056] in, The mean temperature fluctuation variance of the emitted optical power Pout, extinction ratio ER, and emission dispersion eye diagram closed four-phase TDECQ is given. This represents the mean of the maximum temperature variance of the optical module. The values ​​represent the average current fluctuations in the emitted optical power Pout, extinction ratio ER, and emission dispersion eye diagram of the closed four-phase TDECQ. This represents the average value of the maximum current fluctuation. The operating power consumption of the optical module, This represents the maximum rated power consumption of the optical module. , , These are the weighting coefficients. For comprehensive scoring functions, ;

[0057] The comprehensive scoring function is solved using particle swarm optimization to select the optimal driving parameters, and the prediction performance corresponding to the parameter combinations is recorded. The formula is as follows:

[0058] ;

[0059] in, For optimal driving parameters, This is the optimal bias current output by the laser diode equation submodule. The optimal TEC current output by the TEC temperature control unit. This represents the optimal modulation current for the modulator equation submodule. To predict performance, This is a performance prediction function for personalized digital twin models.

[0060] In the aforementioned schemes, existing methods typically perform static performance evaluation and ranking of driving parameters only at a single temperature point, or employ computationally intensive methods such as traversal scanning for multi-condition evaluation. This results in a one-sided evaluation dimension, focusing only on instantaneous performance indicators (such as output optical power at a certain temperature) and completely ignoring the performance stability of parameter combinations across the entire operating temperature range and current disturbances. Consequently, the calibration results lack reliability under actual complex operating conditions. Furthermore, the optimization methods are inefficient. Even when considering multi-dimensional objectives, traditional optimization algorithms with extremely high computational costs and slow convergence are often used, making it difficult to quickly and globally locate the comprehensive optimal solution in a massive parameter space. This makes the virtual calibration process itself an efficiency bottleneck. This application addresses the technical problem of traditional virtual calibration methods failing to guarantee parameter stability across all operating conditions due to the single evaluation dimension by introducing a technical solution that combines a "multi-dimensional comprehensive scoring function integrating temperature stability, current stability, and power consumption" with "efficient global optimization based on particle swarm optimization algorithm." This application constructs a quantitative evaluation system that simultaneously incorporates the variance of key performance indicators across the entire temperature range, the changes under current perturbations, and the module's power consumption into the evaluation, forming a unified comprehensive scoring function. This transforms the parameter optimization objective from single-point performance optimization to comprehensive robustness optimization under all operating conditions. This application ensures the practicality of the optimization results through a multi-dimensional scoring function and ensures the efficiency of the search process through intelligent optimization algorithms. As a result, while improving the quality of virtual calibration results, it significantly shortens the computation time required to obtain the results, enhancing the practicality and efficiency of the entire predictive debugging process.

[0061] Furthermore, the predictive debugging and virtual calibration method for optical modules based on digital twins, wherein the physical testing system, in conjunction with optimal driving parameters, determines whether the measured performance is qualified, includes the following sub-steps:

[0062] The industrial control computer sends the optimal driving parameters to the optical module and controls the optical module to operate according to the parameter combination. After the optical module's working state stabilizes, the control test equipment collects the measured performance.

[0063] The error norm between predicted and measured performance is calculated using the following formula:

[0064] ;

[0065] in, For actual performance testing, Let the error norm be... This represents the measured emitted optical power. For the predicted emitted optical power, This is the measured extinction ratio. For the predicted extinction ratio, This is a measured emission dispersion eye diagram of a closed four-phase system. The predicted emission dispersion eye diagram is a closed four-phase diagram;

[0066] Whether it passes or fails is determined based on the error norm:

[0067] like If the measured performance is satisfactory, then the system will generate a list including the optical module number and optimal driving parameters. Predictive performance Actual performance Error norm The inspection report;

[0068] like If the actual performance fails, the virtual debugging and virtual calibration iterations will be repeated.

[0069] In the above scheme, the physical testing system of this application receives the optimal driving parameters from virtual calibration to drive the optical module. After collecting the measured performance, it calculates the error norm between the measured performance and the predicted performance of the digital twin model, which serves as a quantitative criterion for the model's prediction accuracy. Furthermore, it makes a clear judgment based on a preset strict error threshold (Error≤0.5): if qualified, it directly generates an inspection report containing complete comparison data; if not qualified, it automatically triggers a new round of virtual debugging and calibration iterations, using new measured data to optimize the model. By introducing quantitative comparison and automatic decision-making, it not only achieves the final verification of the fidelity of the digital twin model, but more importantly, it establishes a highly efficient iterative optimization closed loop driven by "virtual-physical" data. This ensures that when there is a deviation between the model prediction and the measured performance, the entire system can automatically and purposefully improve the model and parameters, thereby significantly improving the accuracy of subsequent virtual debugging and the first-time success rate of the overall calibration process, ultimately achieving the goal of reducing the total number of debugging rounds and saving testing resources.

[0070] Furthermore, in the predictive debugging and virtual calibration method for optical modules based on digital twins, the iterative process of re-performing virtual debugging and virtual calibration includes the following sub-steps:

[0071] Point out the unqualified parameters Supplement to the sample set To form a new sample set ;

[0072] Based on the new sample set Iteratively perform virtual debugging and virtual calibration until the measured performance is satisfactory or the number of iterations is reached. Attached Figure Description

[0073] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0074] Figure 1 This is a flowchart of a predictive commissioning and virtual calibration method for optical modules based on digital twins.

[0075] Figure 2 A flowchart for constructing a digital twin model.

[0076] Figure 3 This is a schematic diagram of the structure of a neural network model.

[0077] Figure 4 A flowchart for selecting the parameter combination with the highest comprehensive scoring function.

[0078] Figure 5 This is a schematic diagram of the interaction of a physical testing system. Detailed Implementation

[0079] 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 some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0080] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, the terms "first," "second," etc., are used only for distinguishing descriptions and should not be construed as indicating or implying relative importance, or suggesting any such actual relationship or order between these entities or operations. Additionally, the terms "connected," "linked," etc., can refer to a direct connection between elements or an indirect connection via other elements.

[0081] Example 1: Predictive Commissioning and Virtual Calibration Method for Optical Modules Based on Digital Twin

[0082] This invention is achieved through the following technical solutions, such as... Figure 1 As shown, the predictive commissioning and virtual calibration method for optical modules based on digital twins includes the following steps:

[0083] S1: Construct a digital twin model by fusing the physical characteristics of the optical module with a neural network.

[0084] Specifically, such as Figure 2 As shown, S1 includes the following sub-steps:

[0085] S11: Based on the physical characteristics of the core components of the optical module, a physical equation module is constructed to provide a digital twin model of basic performance constraints;

[0086] The core components of the optical module include a laser diode, a TEC temperature control unit, and a modulator.

[0087] The physical equation module includes a laser diode equation submodule, a TEC temperature control unit equation submodule, and a modulator equation submodule;

[0088] The laser diode equation submodule calculates the relationship between output optical power and bias current using a rate equation, the formula of which is:

[0089] ;

[0090] in, The output optical power of the laser diode submodule. The external quantum efficiency of a laser diode. Let be Planck's constant. , The laser frequency, The amount of electron charge. , For bias current, This is the threshold current of the laser diode. Let be the laser loss coefficient within the cavity. The length of the laser cavity;

[0091] In the embodiments, Typical values ​​range from 0.6 to 0.8, provided in the device datasheet, with batch variations of ±5%. Determined by the operating wavelength of the optical module (e.g., 1310nm wavelength corresponds to) ), Typical values ​​are in the range of 5-10 mA, increasing with increasing temperature. The typical value is , The typical value is 300-500 μm.

[0092] The TEC temperature control unit equation submodule calculates the relationship between the TEC current and the core temperature of the optical module using the heat conduction equation. The formula is as follows:

[0093] ;

[0094] in, This refers to the core temperature of the optical module. For ambient temperature, For TEC thermal resistance, For TEC current, For the power consumption of the optical module, This is the heat dissipation coefficient.

[0095] The modulator equation submodule calculates the relationship between the modulation current and the signal modulation depth using the Markus equation, as follows:

[0096] ;

[0097] in, For signal modulation depth, This reflects the degree to which the modulation current controls the intensity of the optical signal. This is the modulation current (the drive current input to the modulator). The reference value for the modulation current. The recombination energy (in eV) of carrier transfer in the modulator is determined by the properties of the modulator material (such as lithium niobate or silicon). The standard free energy change (in eV) of carrier transfer reflects the energy barrier of the modulation process. is the Boltzmann constant.

[0098] In the embodiments, Take the median value of the modulation current range. , Typical values ​​are 0.2–0.5 eV. Typical values ​​are -0.1 eV to 0 eV. .

[0099] S12: Introduce a neural network model to compensate for the prediction error of the physical equation module;

[0100] like Figure 3 As shown, the neural network model adopts a lightweight fully connected neural network, which includes an input layer, a first hidden layer, a second hidden layer, and an output layer connected in series.

[0101] The input layer includes 3 neurons, the first hidden layer includes 16 neurons, the second hidden layer includes 8 neurons, and the output layer includes 3 neurons.

[0102] S12 includes the following sub-steps:

[0103] S121: Input the driving vector to the input layer. The input layer normalizes the driving vector to eliminate the influence of dimensions. The formula is:

[0104] ;

[0105] in, This is the bias current output by the laser diode equation submodule. The TEC current output by the TEC temperature control unit. The modulation current is the modulated current of the modulator equation submodule. As the driving vector, For transpose, For normalization processing Mapped to the interval [0,1] for The minimum value, for The maximum value.

[0106] S122: The normalized driving vector is input to the first hidden layer. The first hidden layer solves the gradient vanishing problem through the ReLU activation function and uses random deactivation Dropout to prevent overfitting.

[0107] Where Dropout rate = 0.2, and Dropout rate is the random inactivation rate;

[0108] S123: The output data of the first hidden layer is input into the second hidden layer, and the second hidden layer further extracts nonlinear features through the ReLU activation function;

[0109] S124: The output data of the second hidden layer is input to the output layer. The output layer uses the Sigmoid function to map the error and outputs a performance error vector. The formula is as follows:

[0110] ;

[0111] in, This is the performance error vector. It is the error compensation amount for the emitted light power. It is the error compensation amount for the extinction ratio (unit: dB). It is the error compensation amount (in ps) for the closed four-phase emission dispersion eye diagram. The Sigmoid function maps the error to [-0.5, 0.5].

[0112] Therefore, the formulas for S121-S124 are:

[0113] ;

[0114] in, This is the weight matrix of the first hidden layer, with a shape of 16×3. This is the weight matrix for the second hidden layer, with a shape of 8×16. This is the weight matrix of the output layer, with a shape of 3×8. This is the paranoia vector of the first hidden layer, with a shape of 16×1. This is the paranoia vector of the second hidden layer, with a shape of 8×1. This is the bias vector for the output layer, with a shape of 3×1. , , The initial value is set to 0. The ReLU activation function for the first hidden layer. The ReLU activation function for the second hidden layer. It is the Sigmoid function of the output layer.

[0115] S13: Train the neural network model. After training, the neural network model and the compensation physical equation module are combined to construct a digital twin model.

[0116] Specifically, S13 includes the following sub-steps:

[0117] S131: Collect historical measured data of similar optical modules, preprocess the data, and divide the preprocessed data into training set, validation set, and test set;

[0118] The historical measured data covers different production batches, different production equipment, and different working scenarios;

[0119] In this embodiment, the historical measured data is generated in batches of ≥10, using ≥3 different production devices and operating under different conditions such as -40℃ temperature and 30mA bias current, with a total sample size of ≥2000 to ensure data diversity and representativeness. Outliers in the historical measured data are removed using the 3σ criterion (e.g., abnormal power fluctuations caused by testing equipment malfunctions, the judgment criterion being...). ,in The mean, Standard deviation, (The data is historical measured data). After removing the data, it is normalized using Min-Max. After normalization, it is divided into training set, validation set and test set in a ratio of 7:2:1.

[0120] S132: The mean squared error (MSE) loss function is used to minimize the difference between the neural network prediction error and the actual error. The formula is as follows:

[0121] ;

[0122] in, This represents the mean squared error loss value during the training phase. The number of samples in the training set. This represents the actual performance error compensation amount for the j-th optical module. Let $\frac{j}{j}$ be the neural network prediction error compensation amount for the $j$-th sample. Let j be the measured performance vector of the optical module for the j-th sample. Let represent the prediction performance of the physical equation module for the j-th sample.

[0123] It is important to note that It is the core indicator for measuring the prediction performance of neural networks; the smaller the value, the more accurate the prediction.

[0124] S133: The Adam optimizer is used to iteratively train the neural network model. After training, the neural network model is evaluated using a test set to determine whether the model accuracy meets the target.

[0125] In this embodiment, the learning rate of the Adam optimizer is initially 0.001, decaying to 0.8 every 100 iterations, with 500 iterations. Training stops when the loss function on the validation set shows no decrease for 20 consecutive iterations. The test set requires... Otherwise, the neural network structure should be readjusted or additional training data should be provided.

[0126] S134: After the neural network model meets the requirements, the neural network model and the physical equation module are integrated to construct a digital twin model. The formula is as follows:

[0127] ;

[0128] in, For the predictive performance of the optical module, To assess the prediction performance of the physics equations module, Here, Pout represents the error compensation value output by the neural network, ER represents the extinction ratio, and TDECQ represents the emission dispersion eye diagram closed four-phase. This is the theoretically predicted emitted optical power. The extinction ratio is the theoretically predicted value. The theoretically predicted emission dispersion eye diagram is a closed four-phase diagram.

[0129] In the embodiment, the prediction accuracy of the digital twin model can reach: emitted optical power Pout≤0.2dBm, extinction ratio ER≤0.1dB, and emission dispersion eye diagram closed four-phase TDECQ≤0.5ps, which meets the accuracy requirements of optical module testing.

[0130] S2: Fine-tune the parameters of the digital twin model using sample measured data to construct a personalized digital twin model;

[0131] S21: Select three environmental conditions: extremely cold temperature, normal temperature, and extremely hot temperature. Under each working environment, select different sets of typical driving conditions. Collect output performance data under each driving condition. Several performance data constitute a sample set.

[0132] Specifically, S21 includes the following sub-steps:

[0133] S211: Select three environmental conditions: 40℃, 25℃, and 85℃, covering both extreme and normal working environments of the optical module;

[0134] S212: Under each environmental condition, select 3 sets of typical driving conditions (low, medium, and high bias current, corresponding to 5mA, 15mA, and 25mA; TEC current is set according to temperature control requirements, such as 1A TEC current for heating at -40℃, -1A TEC current for cooling at 85℃, and modulation current of 10mA, 25mA, and 40mA).

[0135] S213: Collect output performance data three times under each driving condition, and take the average value as the measured value of that sample point;

[0136] S214: All sample points constitute the sample set, as shown in the formula:

[0137] ;

[0138] in, The driving condition for the k-th sample (dimension 3×1). The measured performance (3×1 dimension) is the value of the k-th sample. For the sample set, .

[0139] S22: The second hidden layer and output layer of the neural network model are fine-tuned and trained with the fine-tuning target of the sample set. The fine-tuned neural network model is combined with the compensated physical equation module to obtain a personalized digital twin model.

[0140] S22 includes the following sub-steps:

[0141] S221: The fine-tuning objective is to minimize the error between the model's predicted and measured values ​​for the sample set. The loss function is defined as follows:

[0142] ;

[0143] in, For loss function, Predict the performance vector for the physical equation module of the k-th sample. This represents the neural network prediction error compensation amount for the k-th sample. The parameters of the neural network to be fine-tuned (weights and biases of the second hidden layer and the output layer);

[0144] S222: Mini-batch gradient descent is used to train and iterate the parameters of the neural network to be fine-tuned. Simultaneously, L2 regularization is applied to the loss function for each iteration to prevent overfitting. The formula is as follows:

[0145] ;

[0146] in, The corrected loss function, This is the weight decay coefficient. , As weight;

[0147] S223: When the corrected loss function is less than the weight decay coefficient or the number of iterations reaches the target, the fine-tuning training of the neural network model is completed. The completed neural network model is combined with the compensated physical equation module to obtain a personalized digital twin model.

[0148] In this example, BatchSize=3, the initial learning rate is 0.0005, lower than in the pre-training stage, to avoid drastic parameter fluctuations, and it decays to 0.9 every 50 iterations. When the loss function... (When the predicted emitted light power Pout ≤ 0.1dBm, extinction ratio ER ≤ 0.05dB, and emission dispersion eye diagram closed four-phase TDECQ ≤ 0.2ps) or when the number of iterations reaches 300 rounds, stop fine-tuning.

[0149] S3: Personalized digital twin models generate virtual test samples by simulating the full working scenario of optical modules in a digital environment, and perform virtual debugging on the virtual test samples;

[0150] Specifically, step S3 includes the following sub-steps:

[0151] S31: Based on the industry standard of optical modules, determine the parameter range and step size of virtual test, generate virtual test samples, and use a parallel computing framework to perform batch prediction on virtual samples to obtain prediction performance indicators.

[0152] Specifically, virtual test samples are generated by combining the parameter ranges and step sizes of temperature range, bias current, TEC current, and modulation current.

[0153] The temperature range is -40℃ to 85℃, with a step size of 5℃, and a total of 26 temperature points;

[0154] The bias current range is 5mA~30mA, with a step size of 1mA, for a total of 26 current points;

[0155] The TEC current range is -1A to 1A, with a step size of 0.1A, and a total of 21 current points;

[0156] The modulation current range is 10mA~40mA, with a step size of 2mA, and a total of 16 current points.

[0157] The total number of virtual test samples is 226,464.

[0158] In this embodiment, a parallel computing framework is used to process 1000 samples in a single batch, with a total prediction time of ≤5 minutes.

[0159] S32: Determine the predictive performance indicators of all virtual test samples and select qualified parameter sets that meet the indicator requirements of different application scenarios;

[0160] The application scenario metrics requirements are as follows:

[0161] Data center scenario: Transmit optical power Pout∈[1dBm, 3dBm], extinction ratio ER≥12dB, emission dispersion eye diagram closed quad-phase TDECQ≤8ps (high signal quality requirements, suitable for short-distance high bandwidth transmission);

[0162] 5G base station scenario: transmit optical power Pout∈[0dBm, 5dBm], extinction ratio ER≥10dB, transmit dispersion eye diagram closed four-phase TDECQ≤10ps (high coverage requirements, more lenient indicators, suitable for medium and long distance transmission);

[0163] Industrial control scenario: Transmitted optical power Pout∈[0.5dBm, 4dBm], extinction ratio ER≥11dB, emission dispersion eye diagram closed four-phase TDECQ≤9ps (balancing signal quality and coverage).

[0164] In the embodiment, taking the data center scenario as an example, if all the indicators of the data center scenario meet the requirements (emitted optical power Pout, extinction ratio ER, emission dispersion eye diagram closed four-phase TDECQ), it is included in the qualified parameter set; otherwise, it is removed. The size of the qualified parameter set is about 15% of the total virtual sample size (about 45,000 parameter points).

[0165] S4: Select the parameter combination with the highest comprehensive scoring function from the virtual test samples that have passed virtual debugging, and use it as the optimal working parameter for virtual calibration;

[0166] Specifically, such as Figure 4 As shown, step S4 includes the following sub-steps:

[0167] S41: Calculate the fluctuation variance of emitted light power Pout, extinction ratio ER, and emission dispersion eye diagram closed four-phase TDECQ over the entire temperature range. The formula is:

[0168] ;

[0169] in, For the variance of the fluctuation, Let Pout be the emitted optical power, ER be the extinction ratio, and TDECQ be the values ​​of the closed four-phase emission dispersion eye diagram at temperature t. The values ​​are the average of emitted light power Pout, extinction ratio ER, and emission dispersion eye diagram closed four-phase TDECQ over the entire temperature range;

[0170] S42: Calculate the changes in emitted optical power Pout, extinction ratio ER, and emission dispersion eye diagram of the closed four-phase TDECQ under bias current ±1mA and modulation current ±2mA fluctuations. The formula is:

[0171] ;

[0172] in, The values ​​represent the emitted optical power Pout, extinction ratio ER, and emission dispersion eye diagram closed four-phase TDECQ under rated current. This represents the maximum change in performance when the bias current fluctuates by ±1mA or the modulation current by ±2mA. This is the maximum value after the current fluctuation. This is the minimum value after current fluctuation;

[0173] S43: Calculate the operating power consumption of the optical module, using the following formula:

[0174] ;

[0175] in, The operating power consumption of the optical module, This is the bias current output by the laser diode equation submodule. The TEC current output by the TEC temperature control unit. The modulation current is the modulated current of the modulator equation submodule. , , ;

[0176] It is important to note that the smaller the fluctuation variance, the better the temperature stability, the smaller the change, and the stronger the current anti-interference capability, that is, the better the current stability. The operating power consumption of the optical module only considers the bias, TEC, and modulation power consumption. The smaller the values ​​of these three, the lower the operating power consumption of the optical module, the better.

[0177] S44: A comprehensive scoring function is obtained by considering temperature stability, current stability, and operating power consumption. The formula is as follows:

[0178] ;

[0179] in, The mean temperature fluctuation variance of the emitted optical power Pout, extinction ratio ER, and emission dispersion eye diagram closed four-phase TDECQ is given. This represents the mean of the maximum temperature variance of the optical module. The values ​​represent the average current fluctuations in the emitted optical power Pout, extinction ratio ER, and emission dispersion eye diagram of the closed four-phase TDECQ. This represents the average value of the maximum current fluctuation. The operating power consumption of the optical module, This represents the maximum rated power consumption of the optical module. , , These are the weighting coefficients. For comprehensive scoring functions, .

[0180] In the embodiments, Typical values ​​are derived from historical data. Typical value Typical value , , , It can be adjusted according to the needs of the scenario (e.g., increasing the size for energy-sensitive scenarios). To 0.4, while reducing up to 0.3);

[0181] It is important to note that The higher the score, the better the stability of the virtual calibration working parameters.

[0182] S45: Particle swarm optimization is used to solve the comprehensive scoring function, the optimal driving parameters are selected, and the prediction performance corresponding to the parameter combination is recorded. The formula is as follows:

[0183] ;

[0184] in, For optimal driving parameters, This is the optimal bias current output by the laser diode equation submodule. The optimal TEC current output by the TEC temperature control unit. This represents the optimal modulation current for the modulator equation submodule. To predict performance, This is a performance prediction function for personalized digital twin models.

[0185] Specifically, S54 includes the following sub-steps:

[0186] S541: Set the number of particles to 50, with each particle corresponding to each candidate parameter combination. The initial inertia weight is 0.9, which decreases linearly to 0.4 with each iteration. This balances the global and local search, using the following formula:

[0187] ;

[0188] in, For inertial weights, This represents the current iteration number. This represents the total number of iterations. ;

[0189] S542: Setting Learning Factors This accelerates the particles toward their individual and global optimum by setting boundary constraints and iterating the particles continuously to select the optimal driving parameters.

[0190] The boundary constraint is that the particle's parameter values ​​are within the range of the virtual test parameters, and if they exceed the range, the boundary is forcibly pulled back.

[0191] S5: The physical testing system, in conjunction with the optimal driving parameters, determines whether the measured performance is qualified. If qualified, a test report is generated; if not qualified, the virtual debugging and virtual calibration iterations are repeated.

[0192] Specifically, such as Figure 5 As shown, the physical testing system includes an optical power meter, an eye diagram meter, a temperature control chamber, and an industrial control computer;

[0193] In the embodiment, the optical power meter is required to have an accuracy of ±0.1dBm, the eye diagram meter is required to have a sampling rate of ≥20GSa / s, and the temperature control box is required to have a temperature control accuracy of ±0.5℃, in order to reduce the time spent on physical measurements;

[0194] The industrial control computer is the core of the collaborative control system and is responsible for communication with the optical module and testing equipment.

[0195] In this embodiment, the communication between the industrial control computer and the optical module uses an RS485 interface with a baud rate of 115200bps. The data frame format is "start bit 1 + data bits 8 + parity bit 1 (even parity) + stop bit 1", and the protocol used is Modbus-RTU for distributing optimal drive parameters. (The instruction format is “010600000005CRC”, where “01” is the module address, “06” is the write instruction, “0000” is the parameter address, “0005” is the parameter value, and “CRC” is the checksum) and read the module’s real-time working status (the instruction format is “010300000003CRC”, which reads 3 performance parameters).

[0196] In this embodiment, communication between the industrial control computer and the testing equipment uses an Ethernet interface based on the TCP / IP protocol. Standardized control commands adapted to the corresponding testing equipment are employed. Three core performance parameters are specifically collected: emitted optical power Pout, extinction ratio ER, and emitted dispersion eye diagram closed four-phase TDECQ. The measurement range of these core performance parameters is set according to the application scenario requirements. The sampling frequency is 1Hz, and the core performance parameters are sampled five times. The average of the five sampling results is taken as the measured performance. .

[0197] S51: The industrial control computer sends the optimal driving parameters to the optical module, controls the optical module to operate according to the parameter combination, and controls the test equipment to collect the measured performance after the optical module's working state is stable.

[0198] The logic for determining the stable working state of the optical module is that the optical module remains stable for 30 seconds, and in the performance value fluctuations of 5 consecutive samplings, the emitted optical power Pout ≤ 0.05dBm, the extinction ratio ER ≤ 0.02dB, and the emission dispersion eye diagram closed four-phase TDECQ ≤ 0.1ps.

[0199] S52: Calculate the error norm between predicted and measured performance, using the following formula:

[0200] ;

[0201] in, For actual performance testing, Let the error norm be... This represents the measured emitted optical power. For the predicted emitted optical power, This is the measured extinction ratio. For the predicted extinction ratio, This is a measured emission dispersion eye diagram of a closed four-phase system. The predicted emission dispersion eye diagram is a closed four-phase diagram;

[0202] S53: Determine whether it is qualified based on the error norm:

[0203] like (Corresponding to emitted optical power Pout≤0.3dBm, extinction ratio ER≤0.2dB, and emission dispersion eye diagram closed quad-phase TDECQ0.8ps), then the measured performance is qualified, and the following information is generated: optical module number, optimal driving parameters. Predictive performance Actual performance Error norm The inspection report;

[0204] like If the actual performance fails, the virtual debugging and virtual calibration iterations will be repeated.

[0205] The iterative process of re-performing virtual debugging and virtual calibration includes the following sub-steps:

[0206] S531: Point out unqualified parameters Supplement to the sample set To form a new sample set ;

[0207] S532: Based on a new sample set Iterate through S22-S5 until the measured performance is satisfactory or the number of iterations is reached.

[0208] During the execution of S22-S5, the initial learning rate is reduced to 0.0003 during fine-tuning to reduce parameter fluctuations, and the weight decay coefficient is increased to... This enhances the ability to resist overfitting.

[0209] The maximum number of iterations is 3. If the optical module is still not qualified after 3 iterations, it is determined to be a defective product (for example, it may have chip failure, packaging defects, etc.) and is transferred to the rework process to avoid infinite loops that lead to a decrease in production efficiency.

[0210] Example 2: Application Steps of Predictive Commissioning and Virtual Calibration Method for Optical Modules Based on Digital Twin

[0211] 1. Offline training of a general digital twin model:

[0212] Input: Historical measured data of similar modules (≥2000 samples);

[0213] Output: A general model consisting of "physical equation module + pre-trained neural network";

[0214] Execution Entity: Cloud server (with GPU acceleration capabilities, reducing training time to ≤2 hours);

[0215] 2. Small sample test of a single module:

[0216] Input: The physics module to be tested;

[0217] Output: 9 small sample data (3 temperature points × 3 sets of driving conditions);

[0218] Execution subject: Testing station on the production line (equipped with temperature control chamber and basic testing equipment, time ≤10 minutes);

[0219] 3. Personalized model fine-tuning:

[0220] Input: General model and sample data;

[0221] Output: A personalized digital twin model adapted to a single module;

[0222] Execution entity: Edge computing unit (deployed at the test station, fine-tuning time ≤ 5 minutes);

[0223] 4. Virtual debugging and calibration:

[0224] Input: Personalized models and application scenario metrics;

[0225] Output: Optimal driving parameters With predictive performance ;

[0226] Execution entity: Edge computing unit (virtual prediction and optimization time ≤ 5 minutes);

[0227] 5. Physical verification test:

[0228] Input: Optimal driving parameters With optical modules;

[0229] Output: Measured performance With the verification results (pass / fail);

[0230] Execution subject: Physical testing system (IPC + testing equipment, time ≤ 5 minutes);

[0231] 6. Iterative optimization:

[0232] Input: Unqualified parameter points ;

[0233] Output: New optimal driving parameters ;

[0234] Or fail to meet the standard; execution subject: edge computing unit and physical testing system work together (single iteration time ≤ 15 minutes).

[0235] Throughout the entire process, physical testing only involves small-sample testing and physical verification testing of a single module, with a total time of ≤15 minutes, which is more than 50% shorter than the traditional solution (30-60 minutes). The equipment resource utilization rate is reduced from 100% in the traditional solution (the entire process occupies testing instruments) to less than 20%, which significantly reduces production costs. At the same time, personalized digital twin models ensure testing accuracy, achieving dual optimization of efficiency and accuracy.

[0236] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A predictive debugging and virtual calibration method for optical modules based on digital twins, characterized in that, Includes the following sub-steps: A digital twin model is constructed by fusing the physical characteristics of optical modules with neural networks; By fine-tuning the parameters of the digital twin model using sample measurement data, a personalized digital twin model can be constructed. Simulate the full working scenario of the optical module in a digital environment, generate virtual test samples using a personalized digital twin model, and perform virtual debugging on the virtual test samples; The parameter combination with the highest comprehensive scoring function is selected from the virtual test samples that have passed virtual debugging and used as the optimal driving parameters for virtual calibration. The physical testing system, in conjunction with the optimal driving parameters, determines whether the measured performance is qualified. If qualified, a test report is generated; if not qualified, the virtual debugging and virtual calibration iterations are repeated. The construction of the digital twin model includes the following sub-steps: Based on the physical characteristics of the core components of the optical module, a physical equation module is constructed to provide a digital twin model of basic performance constraints; A neural network model is introduced to compensate for the prediction error of the physical equation module; Train a neural network model, and then combine the trained neural network model with the compensated physical equation module to construct a digital twin model; The processing formula of the neural network model is as follows: ； in, This is the weight matrix of the first hidden layer, with a shape of 16×3. This is the weight matrix for the second hidden layer, with a shape of 8×16. This is the weight matrix of the output layer, with a shape of 3×8. This is the paranoia vector of the first hidden layer, with a shape of 16×1. This is the paranoia vector of the second hidden layer, with a shape of 8×1. This is the bias vector for the output layer, with a shape of 3×1. , , The initial value is set to 0. The ReLU activation function for the first hidden layer. The ReLU activation function for the second hidden layer. It is the Sigmoid function of the output layer; The construction of a personalized digital twin model includes the following steps: Three environmental conditions were selected: extremely cold temperature, normal temperature, and extremely hot temperature. Different sets of typical driving conditions were selected under each working environment. Output performance data were collected under each driving condition, and several performance data constituted a sample set. The parameters of the second hidden layer and output layer of the neural network model are fine-tuned and trained using the fine-tuning target of the sample set. The fine-tuned neural network model is then combined with the compensated physical equation module to obtain a personalized digital twin model.

2. The predictive debugging and virtual calibration method for optical modules based on digital twins according to claim 1, characterized in that, The process of generating virtual test samples and performing virtual debugging on the virtual test samples includes the following sub-steps: Based on the industry standard of optical modules, the parameter range and step size of virtual testing are determined, virtual test samples are generated, and a parallel computing framework is used to perform batch prediction on the virtual samples to obtain prediction performance indicators. The predictive performance metrics of all virtual test samples are evaluated, and qualified parameter sets that meet the requirements of different application scenarios are selected.

3. The predictive debugging and virtual calibration method for optical modules based on digital twins according to claim 2, characterized in that, The generation of virtual test samples includes the following sub-steps: Virtual test samples are generated by combining the parameter ranges and step sizes of temperature range, bias current, TEC current, and modulation current. The temperature range is -40℃ to 85℃, with a step size of 5℃, and a total of 26 temperature points; The bias current range is 5mA~30mA, with a step size of 1mA, for a total of 26 current points; The TEC current range is -1A to 1A, with a step size of 0.1A, and a total of 21 current points; The modulation current range is 10mA~40mA, with a step size of 2mA, and a total of 16 current points.

4. The predictive debugging and virtual calibration method for optical modules based on digital twins according to claim 2, characterized in that, The application scenario metrics requirements are as follows: Data center scenario: Transmitted optical power Pout∈[1dBm, 3dBm], extinction ratio ER≥12dB, emission dispersion eye diagram closed quadrature TDECQ≤8ps; 5G base station scenario: transmit optical power Pout∈[0dBm, 5dBm], extinction ratio ER≥10dB, transmit dispersion eye diagram closed four-phase TDECQ≤10ps; Industrial control scenario: Emitted optical power Pout∈[0.5dBm, 4dBm], extinction ratio ER≥11dB, emission dispersion eye diagram closed four-phase TDECQ≤9ps.

5. The predictive debugging and virtual calibration method for optical modules based on digital twins according to claim 1, characterized in that, The process of selecting the parameter combination with the highest comprehensive scoring function includes the following sub-steps: The variance of emitted optical power Pout, extinction ratio ER, and emission dispersion eye diagram closed four-phase TDECQ over the entire temperature range is calculated using the following formula: ； in, For the variance of the fluctuation, Let Pout be the emitted optical power, ER be the extinction ratio, and TDECQ be the values ​​of the closed four-phase emission dispersion eye diagram at temperature t. The values ​​are the average of emitted light power Pout, extinction ratio ER, and emission dispersion eye diagram closed four-phase TDECQ over the entire temperature range; The formulas for calculating the changes in emitted optical power Pout, extinction ratio ER, and emission dispersion eye diagram of a closed four-phase TDECQ under bias current fluctuations of ±1mA and modulation current fluctuations of ±2mA are as follows: ； in, The values ​​represent the emitted optical power Pout, extinction ratio ER, and emission dispersion eye diagram closed four-phase TDECQ under rated current. This represents the maximum change in performance when the bias current fluctuates by ±1mA or the modulation current by ±2mA. This is the maximum value after the current fluctuation. This is the minimum value after current fluctuation; The formula for calculating the operating power consumption of an optical module is: ； in, The operating power consumption of the optical module, This is the bias current output by the laser diode equation submodule. The TEC current output by the TEC temperature control unit. The modulation current is the modulated current of the modulator equation submodule. , , ; A comprehensive scoring function is obtained by considering temperature stability, current stability, and operating power consumption. The formula is as follows: ； in, The mean temperature fluctuation variance of the emitted optical power Pout, extinction ratio ER, and emission dispersion eye diagram closed four-phase TDECQ is given. This represents the mean of the maximum temperature variance of the optical module. The values ​​represent the average current fluctuations in the emitted optical power Pout, extinction ratio ER, and emission dispersion eye diagram of the closed four-phase TDECQ. This represents the average value of the maximum current fluctuation. The operating power consumption of the optical module, This represents the maximum rated power consumption of the optical module. , , These are the weighting coefficients. For comprehensive scoring functions, ; The comprehensive scoring function is solved using particle swarm optimization to select the optimal driving parameters, and the prediction performance corresponding to the parameter combinations is recorded. The formula is as follows: ； in, For optimal driving parameters, This is the optimal bias current output by the laser diode equation submodule. The optimal TEC current output by the TEC temperature control unit. This represents the optimal modulation current for the modulator equation submodule. To predict performance, This is a performance prediction function for personalized digital twin models.

6. The predictive debugging and virtual calibration method for optical modules based on digital twins according to claim 1, characterized in that, The physical testing system, in conjunction with the optimal driving parameters, determines whether the measured performance is qualified, including the following sub-steps: The industrial control computer sends the optimal driving parameters to the optical module and controls the optical module to operate according to the parameter combination. After the optical module's working state stabilizes, the control test equipment collects the measured performance. The error norm between predicted and measured performance is calculated using the following formula: ； in, For actual performance testing, Let the error norm be... This represents the measured emitted optical power. For the predicted emitted optical power, This is the measured extinction ratio. For the predicted extinction ratio, This is a measured emission dispersion eye diagram of a closed four-phase system. The predicted emission dispersion eye diagram is a closed four-phase diagram; Whether it passes or fails is determined based on the error norm: like If the measured performance is satisfactory, then the system will generate a list including the optical module number and optimal driving parameters. Predictive performance Actual performance Error norm The inspection report; like If the actual performance fails, the virtual debugging and virtual calibration iterations will be repeated.

7. The predictive debugging and virtual calibration method for optical modules based on digital twins according to claim 6, characterized in that, The iterative process of re-performing virtual debugging and virtual calibration includes the following sub-steps: Point out the unqualified parameters Supplement to the sample set To form a new sample set ; Based on the new sample set Iteratively perform virtual debugging and virtual calibration until the measured performance is satisfactory or the number of iterations is reached.

Images