Coupler reverse design method and system, electronic device and storage medium

Abstract

The application discloses a coupler reverse design method and system, an electronic device and a storage medium. The method adopts a training data set containing different working wavelengths, transmission coefficients and different structure parameter combinations to train a forward performance prediction model. A joint loss function is constructed based on the trained forward performance prediction model. The trained conditional diffusion model is trained based on the joint loss function and the training data set. The target working wavelength and the target transmission coefficient are input into the trained conditional diffusion model to determine a target initial coupler structure parameter candidate set. The structure parameters in the target initial coupler structure parameter candidate set are input into the trained reinforcement learning model to determine a target coupler structure parameter meeting the target working wavelength and the target transmission coefficient. The coupler reverse design is performed through the target coupler structure parameter. The application can improve the accuracy, flexibility and search efficiency of coupler parameter design.

Description

Technical Field

[0001] This application relates to the field of coupler design technology, and in particular to a coupler reverse design method, system, electronic device and storage medium. Background Technology

[0002] Optical couplers are fundamental components in photonic integrated circuits, widely used in optical communication, sensing, and quantum information processing. These devices achieve optical energy transfer through evanescent field coupling between waveguides, and their coupling characteristics directly determine the performance of the entire optical system. The working principle of couplers is based on mode coupling theory, which states that when waveguides approach each other, their transmission modes interact, and optical energy is periodically transferred between the waveguides. By adjusting the waveguide geometry, spacing, and interaction length, specific proportions of optical power distribution or wavelength selection can be achieved.

[0003] Traditional coupler design often employs a combination of parameter scanning and numerical simulation, iteratively searching for the optimal device design that meets the requirements. However, this approach suffers from high simulation costs and limitations imposed by human intuition. In recent years, artificial intelligence (AI) technology has provided a new approach to optical device design. Unlike traditional methods that rely on iterative simulation, AI achieves a network mapping from design parameters to optical responses, directly learning the underlying physical laws, thereby significantly improving computational efficiency. This type of method not only has broad applicability, but also demonstrates significant advantages, particularly in handling complex structure optimization and multiphysics coupling problems. However, it still suffers from limitations such as low flexibility, low accuracy in coupler parameter design, and low search efficiency. Summary of the Invention

[0004] This application aims to propose a coupler reverse design method, system, electronic device, and storage medium, which can improve the accuracy of coupler parameter design, increase the flexibility of coupler parameter design, and improve search efficiency.

[0005] In a first aspect, embodiments of this application provide a coupler reverse design method, the method comprising: A waveguide coupler simulation model, which includes a waveguide and a substrate waveguide, is constructed in simulation software, and the initial structural parameters of the waveguide coupler simulation model are determined. Calculate the refractive index of the waveguide and the refractive index of the substrate waveguide at different operating wavelengths; input the refractive index of the waveguide, the refractive index of the substrate waveguide, and the initial structural parameters into the waveguide coupler simulation model for intrinsic simulation analysis to determine the target structural parameter range and the transmission coefficients corresponding to different combinations of structural parameters within the target structural parameter range; The forward performance prediction model is trained using a training dataset containing different operating wavelengths, transmission coefficients, and different combinations of structural parameters to obtain a trained forward performance prediction model; the different operating wavelengths and different combinations of structural parameters are used as inputs to the forward performance prediction model, and the transmission coefficients are used as training labels for the forward performance prediction model. Based on the trained forward performance prediction model, a joint loss function is constructed; based on the joint loss function, the constructed conditional diffusion model is trained using the training dataset to obtain the trained conditional diffusion model; the different operating wavelengths and the transmission coefficients are used as conditions input to the constructed conditional diffusion model. The target operating wavelength and target transmission coefficient are input into the trained conditional diffusion model to determine the candidate set of target initial coupler structure parameters; The structural parameters in the candidate set of the initial coupler structural parameters are input into the trained reinforcement learning model to determine the target coupler structural parameters that satisfy the target operating wavelength and the target transmission coefficient; the coupler is designed in reverse using the target coupler structural parameters; the reinforcement learning model is constructed through a policy network and a value network.

[0006] In some implementations, constructing a joint loss function based on the trained forward performance prediction model includes: Based on the standard Gaussian noise actually added during the forward diffusion process of the constructed conditional diffusion model and the noise predicted by the constructed conditional diffusion model, a noise prediction loss is constructed. Based on the transmission coefficients predicted by the trained forward performance prediction model and the transmission coefficients input to the constructed conditional diffusion model, a transmission coefficient consistency constraint loss is constructed. Based on the structural parameters predicted by the constructed conditional diffusion model, construct structural parameter process constraint loss that meets the actual processing requirements; Based on the constructed conditional diffusion model, multiple candidate structure parameters are generated, and a candidate structure diversity constraint loss is constructed. The noise prediction loss, the transmission coefficient consistency constraint loss, the structural parameter process constraint loss, and the candidate structure diversity constraint loss are weighted and summed to construct a joint loss function.

[0007] In some implementations, constructing a transmission coefficient consistency constraint loss based on the transmission coefficients predicted by the trained forward performance prediction model and the transmission coefficients input to the constructed conditional diffusion model includes: The current coupler structure parameters and their corresponding operating wavelengths, predicted by the conditional diffusion model during the reverse denoising process, are input into the trained forward performance prediction model to predict the transmission coefficients. The transmission coefficients predicted by the trained forward performance prediction model and the transmission coefficients input to the constructed conditional diffusion model are subjected to squared difference calculation to construct the transmission coefficient consistency constraint loss.

[0008] In some implementations, constructing the structural parameter process constraint loss that meets actual processing requirements based on the structural parameters predicted by the constructed conditional diffusion model includes: ; in, This represents the loss due to structural parameter process constraints. This represents the first structural parameter predicted by the constructed conditional diffusion model during the reverse denoising process. One parameter, This indicates that the structure parameters predicted by the constructed conditional diffusion model during the reverse denoising process are of the th order. The maximum value of each parameter. This indicates that the structure parameters predicted by the constructed conditional diffusion model during the reverse denoising process are of the th order. The minimum value of each parameter. This indicates taking the maximum value.

[0009] In some implementations, the step of generating multiple candidate structure parameters based on the constructed conditional diffusion model and constructing a candidate structure diversity constraint loss includes: ; in, This represents the candidate structure diversity constraint loss. Indicates the number of candidate structure parameters. Represents the normalized i-th One candidate structural parameter, Represents the normalized i-th One candidate structural parameter, This represents the distance adjustment coefficient. Denotes the square of the L2 norm. This represents an exponential function.

[0010] In some implementations, the step of training the constructed conditional diffusion model using the training dataset based on the joint loss function to obtain the trained conditional diffusion model includes: A fully connected neural network is used to construct a conditional diffusion model, resulting in the constructed conditional diffusion model. The operating wavelength, transmission coefficient, diffusion steps, and noisy samples in the training dataset are used as inputs to the constructed conditional diffusion model; the diffusion steps are mapped to a diffusion step embedding vector through sinusoidal position encoding; the operating wavelength and transmission coefficient are linearly mapped to obtain wavelength embedding vector and transmission coefficient embedding vector, respectively. The condition vector is obtained by adding the diffusion step embedding vector, the wavelength embedding vector, and the transmission coefficient embedding vector. Based on the joint loss function, the conditional vector and the noisy sample are input into the constructed conditional diffusion model for training until the joint loss function converges or reaches the preset number of training rounds, thus obtaining the trained conditional diffusion model.

[0011] In some implementations, the step of inputting the structural parameters from the candidate set of initial target coupler structural parameters into a trained reinforcement learning model to determine the target coupler structural parameters that satisfy the target operating wavelength and the target transmission coefficient includes: A policy network and a value network are constructed using a multi-layer fully connected neural network; a reinforcement learning model is then constructed based on the policy network and the value network. The structural parameters and their corresponding transmission coefficients obtained in the initial coupler structural parameter candidate set during the training of the constructed conditional diffusion model are used as the initial state for reinforcement learning. The initial state is input into the policy network to obtain the action output by the policy network; the current structure parameters are updated according to the action output by the policy network to obtain the updated structure parameters. Based on the updated structural parameters, a reward function is constructed; an immediate reward is obtained through the reward function, and an advantage function is calculated based on the immediate reward; Based on the advantage function, a loss function for the policy network is constructed; using mean squared error as the optimization objective, a loss function for the value network is constructed. Based on the loss function of the policy network and the loss function of the value network, the policy network and the value network are iteratively trained until the loss function converges or the preset number of training rounds is reached, thus obtaining a trained reinforcement learning model. The structural parameters in the candidate set of initial target coupler structural parameters are input into the trained reinforcement learning model to determine the target coupler structural parameters that satisfy the target operating wavelength and the target transmission coefficient; the target coupler structural parameters are the structural parameters with the highest corresponding transmission coefficient.

[0012] Secondly, embodiments of this application also provide a coupler reverse design system, the system comprising: The model building module is used to build a waveguide coupler simulation model in the simulation software, which includes a waveguide and a substrate waveguide, and to determine the initial structural parameters of the waveguide coupler simulation model. The simulation analysis module is used to calculate the refractive index of the waveguide and the refractive index of the substrate waveguide at different operating wavelengths; the refractive index of the waveguide, the refractive index of the substrate waveguide, and the initial structural parameters are input into the waveguide coupler simulation model for intrinsic simulation analysis to determine the target structural parameter range and the transmission coefficients corresponding to different combinations of structural parameters within the target structural parameter range; The first training module is used to train the forward performance prediction model using a training dataset containing different operating wavelengths, transmission coefficients, and different combinations of structural parameters to obtain a trained forward performance prediction model; the different operating wavelengths and different combinations of structural parameters are used as inputs to the forward performance prediction model, and the transmission coefficients are used as training labels for the forward performance prediction model. The second training module is used to construct a joint loss function based on the trained forward performance prediction model; to train the constructed conditional diffusion model using the training dataset based on the joint loss function to obtain the trained conditional diffusion model; and to use the different operating wavelengths and the transmission coefficients as conditions for inputting the constructed conditional diffusion model. The parameter determination module is used to input the target operating wavelength and the target transmission coefficient into the trained conditional diffusion model to determine the candidate set of target initial coupler structure parameters; The coupler inverse design module is used to input the structural parameters from the candidate set of the initial coupler structural parameters into the trained reinforcement learning model to determine the target coupler structural parameters that satisfy the target operating wavelength and the target transmission coefficient; and to perform coupler inverse design using the target coupler structural parameters; the reinforcement learning model is constructed through a policy network and a value network.

[0013] Thirdly, embodiments of this application also provide an electronic device, including at least one control processor and a memory for communicatively connecting to the at least one control processor; the memory stores instructions executable by the at least one control processor, the instructions being executed by the at least one control processor to enable the at least one control processor to perform a coupler reverse design method as described above.

[0014] Fourthly, embodiments of this application also provide a computer-readable storage medium storing computer-executable instructions for causing a computer to execute a coupler reverse design method as described above.

[0015] Compared with the prior art, this application has the following beneficial effects: This application constructs a waveguide coupler simulation model containing a waveguide and a substrate waveguide in simulation software and determines the initial structural parameters of the waveguide coupler simulation model; calculates the refractive index of the waveguide and the substrate waveguide at different operating wavelengths; inputs the refractive index of the waveguide, the refractive index of the substrate waveguide, and the initial structural parameters into the waveguide coupler simulation model for intrinsic simulation analysis to determine the target structural parameter range and the transmission coefficients corresponding to different combinations of structural parameters within the target structural parameter range; trains the forward performance prediction model using a training dataset containing different operating wavelengths, transmission coefficients, and different combinations of structural parameters to obtain a trained forward performance prediction model; uses different operating wavelengths and different combinations of structural parameters as inputs to the forward performance prediction model, and uses the transmission coefficients as... Training labels for the forward performance prediction model; constructing a joint loss function based on the trained forward performance prediction model; training the constructed conditional diffusion model using the training dataset based on the joint loss function to obtain the trained conditional diffusion model; using different operating wavelengths and transmission coefficients as inputs to construct the conditions of the conditional diffusion model; inputting the target operating wavelength and target transmission coefficient into the trained conditional diffusion model to determine the candidate set of target initial coupler structure parameters; inputting the structure parameters from the candidate set of target initial coupler structure parameters into the trained reinforcement learning model to determine the target coupler structure parameters that satisfy the target operating wavelength and target transmission coefficient; performing coupler inverse design using the target coupler structure parameters; the reinforcement learning model is constructed through a policy network and a value network.

[0016] Thus, by constructing a joint loss function based on a pre-trained forward performance prediction model, the conditional diffusion model can learn the data distribution of the real coupler structure parameters. Subsequent training of the pre-trained conditional diffusion model using the joint loss function improves its prediction accuracy. Inputting the target operating wavelength and target transmission coefficient into the trained conditional diffusion model determines the candidate set of initial coupler structure parameters, realizing a reverse design paradigm. This significantly reduces the computational resources and time costs associated with repetitive simulations and improves the flexibility and efficiency of device design. Furthermore, a reinforcement learning model determines the target coupler structure parameters that satisfy the target operating wavelength and target transmission coefficient from the candidate set of initial coupler structure parameters. This avoids the problem of the structure deviating from the physically realizable range due to a simple pursuit of performance, and effectively reduces the possibility of missing the global optimum due to reliance on empirical design, resulting in accurate coupler structure parameters. This improves the accuracy, flexibility, and search efficiency of coupler parameter design. Attached Figure Description

[0017] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which: Figure 1 This is a flowchart illustrating an embodiment of the coupler reverse design method provided in this application; Figure 2 This is a schematic diagram of the coupler in the preferred embodiment of the coupler reverse design method provided in this application; Figure 3 yes Figure 2 Schematic diagram of section A-A' in the diagram; Figure 4 This is a schematic diagram of an embodiment of the coupler reverse design system provided in this application; Figure 5 This is a schematic diagram of the structure of an embodiment of the electronic device provided in this application. Detailed Implementation

[0018] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application.

[0019] In the description of this application, the use of terms such as "first," "second," etc., is for the purpose of distinguishing technical features only and should not be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated or the order of the technical features indicated.

[0020] In the description of this application, it should be understood that the orientation descriptions, such as up, down, etc., are based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this application.

[0021] In the description of this application, it should be noted that, unless otherwise explicitly defined, terms such as "setup," "installation," and "connection" should be interpreted broadly, and those skilled in the art can reasonably determine the specific meaning of the above terms in this application in conjunction with the specific content of the technical solution.

[0022] To address the problems of limited flexibility in coupler parameter design, low accuracy in coupler parameter design, and low search efficiency in related technologies, this application proposes a coupler reverse design method, system, electronic device, and storage medium.

[0023] Reference Figure 1This application provides a schematic flowchart of a coupler reverse design method. This coupler reverse design method is applied to electronic devices, such as servers or mobile terminals. Figure 1 As shown, the reverse design method for the coupler may include the following steps S101 to S106.

[0024] Step S101: Construct a waveguide coupler simulation model in the simulation software, which includes the waveguide and the substrate waveguide, and determine the initial structural parameters of the waveguide coupler simulation model.

[0025] In this step, a waveguide coupler simulation model is constructed in the simulation software, comprising a waveguide and a substrate waveguide. Optical energy coupling between the two waveguides is achieved through evanescent field interaction. The waveguide structure uses a rectangular cross-section, with the waveguide core surrounded by a cladding material. Waveguide materials include, but are not limited to, high-refractive-index dielectric materials such as silicon and silicon nitride; substrate waveguide materials include, but are not limited to, low-refractive-index materials such as silicon dioxide. The initial structural parameters of the waveguide coupler simulation model are determined, including: waveguide width, waveguide height, coupling length, bending waveguide radius, and coupling gap.

[0026] Step S102: Calculate the refractive index of the waveguide and the refractive index of the substrate waveguide at different operating wavelengths; input the refractive index of the waveguide, the refractive index of the substrate waveguide, and the initial structural parameters into the waveguide coupler simulation model for intrinsic simulation analysis, and determine the target structural parameter range and the transmission coefficients corresponding to different combinations of structural parameters within the target structural parameter range.

[0027] In this step, the refractive index of the waveguide and the refractive index of the substrate waveguide at different operating wavelengths are calculated as follows: ; ; in, Indicates the operating wavelength. Indicates the operating wavelength The refractive index of the lower waveguide, Indicates the operating wavelength The refractive index of the underlying substrate waveguide.

[0028] Step S103: Use a training dataset containing different operating wavelengths, transmission coefficients and different combinations of structural parameters to train the forward performance prediction model to obtain a trained forward performance prediction model; use different operating wavelengths and different combinations of structural parameters as inputs to the forward performance prediction model, and use the transmission coefficients as training labels for the forward performance prediction model.

[0029] In this step, different operating wavelengths and combinations of structural parameters are used as inputs to the forward performance prediction model, and the transmission coefficients are used as training labels for the model. The forward performance prediction model is trained using a training dataset containing different operating wavelengths, transmission coefficients, and combinations of structural parameters, resulting in a trained model. This trained model is then used to quickly predict the transmission coefficients of the coupler structural parameters at a given operating wavelength.

[0030] The forward performance prediction model in this embodiment is constructed using a neural network model known to those skilled in the art, such as a forward performance prediction model constructed using a fully connected network model. This embodiment does not impose any specific limitations on this.

[0031] Step S104: Based on the trained forward performance prediction model, construct a joint loss function; based on the joint loss function, train the constructed conditional diffusion model using the training dataset to obtain the trained conditional diffusion model; use different operating wavelengths and transmission coefficients as inputs to the constructed conditional diffusion model.

[0032] In this step, a noise prediction loss is constructed based on the standard Gaussian noise actually added during the forward diffusion process of the constructed conditional diffusion model and the noise predicted by the constructed conditional diffusion model; a transmission coefficient consistency constraint loss is constructed based on the transmission coefficients predicted by the trained forward performance prediction model and the transmission coefficients input to the constructed conditional diffusion model; a structural parameter process constraint loss that meets the actual processing requirements is constructed based on the structural parameters predicted by the constructed conditional diffusion model; a candidate structure diversity constraint loss is constructed based on the multiple candidate structural parameters generated by the constructed conditional diffusion model; and a joint loss function is constructed by weighting and summing the noise prediction loss, transmission coefficient consistency constraint loss, structural parameter process constraint loss, and candidate structure diversity constraint loss.

[0033] The above-mentioned noise prediction loss is constructed based on the standard Gaussian noise actually added during the forward diffusion process of the constructed conditional diffusion model and the noise predicted by the constructed conditional diffusion model, including: ; in, Indicates the noise prediction loss. This represents the standard Gaussian noise actually added during the forward diffusion process of the constructed conditional diffusion model. This represents the noise predicted using a pre-constructed conditional diffusion model. Indicates the first Step with noisy samples, This represents the initial noisy sample. Indicates the operating wavelength. Indicates the transmission coefficient. Indicates the number of diffusion steps.

[0034] The above-mentioned transmission coefficient consistency constraint loss is constructed based on the transmission coefficients predicted by the trained forward performance prediction model and the transmission coefficients input to the constructed conditional diffusion model, including: The current coupler structure parameters and their corresponding operating wavelengths, predicted by the conditional diffusion model during the reverse denoising process, are input into the trained forward performance prediction model to predict the transmission coefficients. The squared difference between the transmission coefficients predicted by the trained forward performance prediction model and the transmission coefficients input into the constructed conditional diffusion model is calculated to construct the transmission coefficient consistency constraint loss.

[0035] Based on the structural parameters predicted by the constructed conditional diffusion model, the above-mentioned structural parameter process constraint loss that meets the actual processing requirements is constructed, including: ; in, This represents the loss due to structural parameter process constraints. This represents the first structural parameter predicted by the constructed conditional diffusion model during the reverse denoising process. One parameter, This indicates that the structure parameters predicted by the constructed conditional diffusion model during the reverse denoising process are of the th order. The maximum value of each parameter. This indicates that the structure parameters predicted by the constructed conditional diffusion model during the reverse denoising process are of the th order. The minimum value of each parameter. This indicates taking the maximum value.

[0036] The above-mentioned candidate structure parameters are generated based on the constructed conditional diffusion model, and a candidate structure diversity constraint loss is constructed, including: ; in, This represents the candidate structure diversity constraint loss. Indicates the number of candidate structure parameters. Represents the normalized i-th One candidate structural parameter, Represents the normalized i-th One candidate structural parameter, This represents the distance adjustment coefficient. Denotes the square of the L2 norm. This represents an exponential function.

[0037] A conditional diffusion model is constructed using a fully connected neural network. The working wavelength, transmission coefficient, diffusion steps, and noisy samples from the training dataset are used as inputs to the constructed conditional diffusion model. The diffusion steps are mapped to diffusion step embedding vectors using sinusoidal position encoding. The working wavelength and transmission coefficient are linearly mapped to obtain wavelength embedding vectors and transmission coefficient embedding vectors, respectively. The diffusion step embedding vector, wavelength embedding vector, and transmission coefficient embedding vector are added together to obtain the conditional vector. Based on the joint loss function, the conditional vector and noisy samples are input into the constructed conditional diffusion model for training until the joint loss function converges or the preset number of training rounds is reached, resulting in a trained conditional diffusion model.

[0038] Step S105: Input the target operating wavelength and target transmission coefficient into the trained conditional diffusion model to determine the candidate set of target initial coupler structure parameters.

[0039] In this step, when a target operating wavelength is given and target transmission coefficient Then, initial noise samples are sampled from a standard normal distribution, and then reverse denoising is performed in T iterations: ; in, This represents the sample of noisy coupler structure parameters at step i; Let represent the noise variance scheduling corresponding to the i-th diffusion step; and define the cumulative coefficient. , indicating each step in the diffusion process The cumulative product; This represents the structural parameter sample after denoising in step i-1; This represents the noise predicted by the conditional diffusion model; This represents the standard deviation of the backsampling noise in the i-th step; This represents standard Gaussian noise that follows a normal distribution. Repeating this process K times yields a candidate set of initial coupler structure parameters. This serves as the input for the subsequent reinforcement learning optimizer.

[0040] Step S106: Input the structural parameters from the candidate set of initial coupler structural parameters into the trained reinforcement learning model to determine the target coupler structural parameters that satisfy the target operating wavelength and target transmission coefficient; perform coupler inverse design using the target coupler structural parameters; the reinforcement learning model is constructed through the policy network and the value network.

[0041] In this step, a multi-layer fully connected neural network is used to construct a policy network and a value network. Based on the policy network and the value network, a reinforcement learning model is constructed. The structural parameters and their corresponding transmission coefficients obtained from the initial coupler structural parameter candidate set during the training of the constructed conditional diffusion model are used as the initial state of reinforcement learning. The initial state is input into the policy network to obtain the action output by the policy network. The current structural parameters are updated according to the action output by the policy network to obtain the updated structural parameters. A reward function is constructed based on the updated structural parameters. An immediate reward is obtained through the reward function, and an advantage function is calculated based on the immediate reward. A loss function for the policy network is constructed based on the advantage function. A loss function for the value network is constructed using mean squared error as the optimization objective. Based on the loss functions of the policy network and the value network, the policy network and the value network are iteratively trained until the loss function converges or the preset number of training rounds is reached to obtain a trained reinforcement learning model. The structural parameters from the target initial coupler structural parameter candidate set are input into the trained reinforcement learning model to determine the target coupler structural parameters that satisfy the target operating wavelength and the target transmission coefficient. The target coupler structural parameters are the structural parameters with the highest corresponding transmission coefficient.

[0042] To facilitate understanding by those skilled in the art, a set of preferred embodiments is provided below: The purpose of this embodiment is to provide a reverse design method for a coupler that coordinates the generator and optimizer. First, based on Figure 1 and Figure 2 The schematic diagram of the coupler shown illustrates the construction of a simulation dataset and the creation of a structural parameter space. Then, based on the conditional diffusion model, a candidate set of coupler structural parameters that conforms to the physical laws under the conditions of the target operating wavelength and target transmission coefficient is generated in the structural parameter space. Finally, the trained conditional diffusion model is used as a generator to learn the distribution law of the coupler structural parameters; the constructed reinforcement learning agent is used as an optimizer to achieve goal-oriented parameter tuning. Through the cascade collaboration of the conditional diffusion model and the reinforcement learning model, collaborative optimization training and prediction are performed, achieving a fast inverse solution from the high coupling efficiency target (including the target wavelength and target transmission coefficient) to the coupler structural parameters. The detailed implementation method is described below: Step 1: Modeling and constructing the coupler device. Waveguide coupler simulation model.

[0043] This embodiment models the coupler device in simulation software (such as COMSOL Multiphysics) and constructs it. A simulation model of a waveguide coupler. This coupler achieves optical energy coupling between the two waveguides through evanescent field interaction. For example... Figure 2 As shown, the input ports of the coupler are respectively and The output ports are respectively and Light waves exchange energy within the coupling region. In this embodiment, the coupling characteristics are determined by the transmission coefficient. Description. For example... Figure 3 As shown, the waveguide structure employs a rectangular cross-section, with the waveguide core layer surrounded by a cladding material. The waveguide core layer material includes, but is not limited to, high-refractive-index dielectric materials such as silicon (Si) and silicon nitride (SiN); the substrate includes, but is not limited to, silicon dioxide (SiN). Low refractive index materials such as waveguides are used. This embodiment is suitable for various types of couplers, such as microring combined straight waveguides and dual microrings. For better illustration, this embodiment uses a general coupler shape as an example for analysis, with waveguides as an example. and substrate waveguide This example illustrates the optimized coupler design. Furthermore, the initial structural parameters of the coupler include the waveguide width. Waveguide height Coupling length Bending waveguide radius and coupling gap By adjusting the initial structural parameters of the coupler described above, the transmission coefficient can be adjusted. Regulation.

[0044] Step 2: Data collection and training dataset construction.

[0045] Before collecting and constructing the training dataset, this embodiment first obtains the coupler structure parameters based on actual production needs and process requirements. The initial range of values. In order to perform... Waveguide coupler simulation requires determining and Material at operating wavelength The refractive index of the material. Waveguide and The refractive index can be determined according to the wavelength of light. Calculated using the Sellmeier formula: ; ; in, Indicates the operating wavelength Down The refractive index (i.e., the refractive index of the waveguide). Indicates the operating wavelength Down The refractive index of the substrate waveguide (i.e., the refractive index of the material). The calculated material refractive index will be used as a material parameter in the simulation software for... Waveguide simulation is used to determine the waveguide modes and effective refractive index under different combinations of waveguide size parameters. Within the range of initial structural parameters determined in step one, the simulation employs... A waveguide coupler simulation model is used to perform intrinsic simulation analysis on different combinations of structural parameters under different material refractive indices at different operating wavelengths. The initial structural parameter range is verified by using the TE mode in the transverse electric field to satisfy single-mode transmission as the selection criterion, eliminating structural parameter combinations that do not meet the single-mode condition, thus obtaining a structural parameter range that meets the design requirements. After determining the final structural parameter range (i.e., the target structural parameter range) and the transmission coefficients corresponding to each combination of structural parameters within the final structural parameter range through simulation, a training dataset is constructed based on each structural parameter, transmission coefficient, and different operating wavelengths within the final structural parameter range. The transmission coefficients are then used as training labels or conditions for the training datasets of subsequent neural network models.

[0046] For a segment of length in the coupler The coupling region, where the input and output optical fields are related, can be described by the transfer matrix: ; in, The imaginary unit satisfies ; For the first The average phase delay of the segment coupling region; For the first The transmission coefficient of the segment coupling region, For the first The coupling coefficient of the coupling region. Characteristic mode analysis is performed on the coupler cross-section in simulation software to extract the effective refractive index of the even modes. and odd mode effective refractive index Based on supermode theory, the odd-mode propagation constant and the even-mode propagation constant are respectively: ; ; in, For wavelength, Represents the odd-mode propagation constant. Let represent the even-mode propagation constant. For a length of , ... The coupling region, average phase delay It is determined by the average of the propagation constants of the two supermodes: ; The average phase delay describes the common phase accumulated by the light wave after passing through the coupling region and is reflected in the common terms of the transmission matrix. Transmission coefficient and coupling coefficient This is determined by the difference in propagation constants between the two supermodes. The phase difference between the two supermodes is... Half of the phase difference determines the distribution of light energy between the two waveguides, therefore the transmission coefficient... and coupling coefficient They are respectively: ; ; For lossless couplers, the following conditions must be met: Based on the above principles, this embodiment constructs a coupler optimization design dataset through simulation and calculation. The operating wavelength is... As a conditional control variable, the transmission coefficient As input to the neural network, the structural parameters of the coupler This is the output.

[0047] Furthermore, to further improve the ability to constrain target transmission performance during subsequent generator training, this embodiment utilizes the aforementioned simulation data to further train the forward performance prediction model. This model is used to quickly predict the transmission coefficients of coupler structural parameters at a given operating wavelength. This forward performance prediction model serves only as a performance evaluation module in the joint loss function during the training phase of the conditional diffusion model and does not participate in the final inference process.

[0048] It should be noted that the forward performance prediction model in this embodiment can adopt a neural network model known to those skilled in the art, such as a forward performance prediction model constructed using a fully connected network model. This embodiment does not impose any specific restrictions on this.

[0049] Step 3: Construction of the conditional diffusion model.

[0050] This embodiment is constructed based on the operating wavelength. and transmission coefficient A conditional diffusion model is used as the input condition, and this model is used as a generator to generate a candidate set of initial coupler structure parameters that meet the target operating wavelength requirements. The coupler structure parameter vector is represented as follows: ; in, For waveguide width, For waveguide height, The length of the coupling region. Where is the bending radius, The distance between the two waveguides, indicated by the superscript. This represents the vector transpose. The conditional diffusion model learns the vector transpose at a given operating wavelength. and transmission coefficient The data distribution of coupler structure parameters under certain conditions is used to gradually recover coupler structure parameter samples from random noise. The training dataset consists of the operating wavelength. Coupler structural parameters and its corresponding transmission coefficient Composition. By learning from the simulation data above, the conditional diffusion model can learn the conditional probability distribution, thereby generating a candidate set of initial coupler structure parameters that meet the target operating wavelength and target transmission performance requirements.

[0051] (1) Forward diffusion process.

[0052] To learn the stepwise generation process from random noise to real coupler structure parameter samples, the conditional diffusion model includes two processes: forward diffusion and backward denoising. The forward diffusion process involves stepwise feedforward to the real coupler structure parameter samples. Add Gaussian noise to generate noisy sample sequences The actual coupler structure parameter samples are represented as follows: The i-th forward diffusion process can be represented as: ; in, Indicates that in the known number of... Step Sample Under the condition that the i-th step noisy sample The conditional probability distribution; Indicates a Gaussian distribution; The noise variance scheduling corresponds to the i-th diffusion step. T is the total number of diffusion steps. Indicates the first Step with noisy samples, Indicates the first Step noise, This represents the identity matrix with the same dimension as the structural parameter vector. During forward diffusion, as the number of diffusion steps *i* increases, the noisy samples... The Gaussian noise contained in the sample gradually increases, while the information of the original structural parameters gradually decreases. When the total number of diffusion steps T is large enough, the final noisy sample... It approximately follows a standard Gaussian distribution. To simplify subsequent formulas, let... and define the cumulative coefficient. , indicating each step in the diffusion process The cumulative product. According to the reparameterization technique, it can be obtained from the actual coupler structure parameter samples. The formula for directly calculating any noisy sample in step i is as follows: ; in, This indicates the added noise. It follows standard Gaussian noise with zero mean and an identity matrix. This represents the original structural parameter information retained in the noisy sample of step i; This represents the random noise added in step i; Represents the identity matrix.

[0053] (2) Reverse denoising process.

[0054] Corresponding to the forward diffusion process, the reverse denoising process starts with pure noise samples approximating a standard Gaussian distribution. By progressively predicting and removing noise, it recovers the coupler structure parameters that conform to the sample distribution of the actual coupler structure parameters. This process is modeled as a parameterized Markov chain, and its i-th step reverse conditional probability distribution is expressed as: ; in, These are the trainable parameters for the conditional diffusion model; This indicates that, given the noisy sample at step i... Operating wavelength and transmission coefficient Conditions, No. Step Sample The inverse denoising probability distribution; This represents the predicted mean corresponding to the reverse denoising process in the conditional diffusion model. This represents the prediction covariance matrix corresponding to the backward denoising process in the conditional diffusion model. In this embodiment, the conditional diffusion model is used to predict the noise added to the structural parameter samples during the forward diffusion process; the prediction result is... , indicating a given noisy sample diffusion step number i, operating wavelength and transmission coefficient Under the given conditions, the noise vector is predicted by the conditional diffusion model. Based on the noise predicted by the conditional diffusion model, the mean of the reverse denoising process can be expressed as: ; in, Indicates the predicted first denoising step during the inverse denoising process. Step average; Indicates the first Noisy samples; The first term predicted by the conditional diffusion model is represented by... Step noise; This represents the trainable parameters of the conditional diffusion model.

[0055] (3) Conditional diffusion model network structure.

[0056] To achieve the aforementioned noise prediction, this embodiment employs a fully connected neural network to construct a conditional diffusion model. The neural network input includes the noisy sample from step i. diffusion step number i, operating wavelength and transmission coefficient The diffusion step number i is mapped to a d-dimensional diffusion step embedding vector through sinusoidal position encoding. The 2m-th and 2m+1-th dimensions of the embedding vector are represented as follows: ; in, Indicates the embedded vector dimension index. This represents the dimension of the positional encoding; in this embodiment, we take... Operating wavelength and transmission coefficient After linear mapping, wavelength embedding vectors with the same dimension as the embedding vectors in the diffusion step are obtained. and transmission coefficient embedding vector Their expressions are as follows: ; ; in, and These represent the linear mapping weights corresponding to the operating wavelength and the transmission coefficient, respectively; and These represent the corresponding bias vectors. Adding the diffusion step embedding vector, wavelength embedding vector, and transmission coefficient embedding vector yields the conditional vector: ; in, This is used to simultaneously provide the neural network with information on the number of diffusion steps, operating wavelength, and transmission coefficients during noise prediction. This embodiment uses a four-layer fully connected structure to form the backbone network of the conditional diffusion model. Each fully connected layer is followed by an Adaptive Layer Normalization (AdaLN) activation function and a ReLU activation function. Let the... Layer input features are Then the first The linear change process of the layer is represented as: ; Among them, in the first layer of command ; and They represent the first Weight matrix and bias vector of a fully connected layer; Indicates the first Intermediate features obtained after linear transformation of the layer. To transform the conditional vector... Injected into the backbone network, for intermediate features Perform adaptive layer normalization: ; ; ; in, This represents the features after adaptive normalization. ; and Representing features respectively The value and variance; Indicates the first Dimension of layer features; ⊙ indicates a preset positive number used to improve numerical stability; ⊙ indicates element-wise multiplication. and They represent the condition vectors respectively. The generated scaling and offset vectors are obtained through linear mapping: ; ; in, , , as well as For the first The trainable parameters of the layer AdaLN module, and their corresponding and Let represent the weight matrix and bias vector, respectively. After AdaLN processing, the ReLU activation function is used to obtain the _th Layer output features: ; After four layers of fully connected feature transformation, the final hidden features are then passed through the output layer. Mapped to prediction noise: ; in, This indicates that the conditional diffusion model, given the noisy sample at step i, is in operation. diffusion step number i, operating wavelength and transmission coefficient The noise vector predicted under the given conditions; This represents the set of all trainable parameters in the conditional diffusion model, including the weights and biases of each fully connected layer, conditional mapping layer, AdaLN module, and output layer. This represents the weight matrix of the output layer; This represents the bias vector of the output layer.

[0057] (4) The training and prediction process of the conditional diffusion model.

[0058] In this embodiment, during the training of the conditional diffusion model, not only is the error between the neural network's predicted noise and the actual Gaussian noise added during the forward diffusion process used as the training objective, but it also further introduces constraints on transmission coefficient consistency, structural parameter manufacturing processes, and candidate structure diversity to construct a joint loss function for coupler reverse design. Its expression is: ; in, Indicates the noise prediction loss; This represents the loss of consistency in transmission coefficients; This represents the loss due to structural parameter process constraints. This represents the loss of candidate structure diversity; These represent the weighting coefficients for each loss term. The noise prediction loss for each part is calculated as follows: ; in, Indicates the noise prediction loss; This represents the standard Gaussian noise actually added during the forward diffusion process; This represents the noise predicted by the conditional diffusion model; Indicates the first The loss is applied to noisy samples, ensuring that the diffusion model can learn the data distribution of the true coupler structural parameters. The loss for transmission coefficient consistency constraints is calculated as follows: ; The established coupler forward performance prediction model calculates the transmission coefficient corresponding to this structure: ; in, This represents the coupler transmission performance prediction model obtained by training the simulation data from step two; This represents the transmission coefficients corresponding to the predicted structural parameters; This represents the current coupler structure parameters predicted by the conditional diffusion model during the reverse denoising process, calculated as follows: ; By minimizing this loss, the generated structure not only satisfies the real data distribution but also meets the target transmission performance requirements. To ensure that the generated structure meets actual processing requirements, this embodiment introduces a structural parameter process constraint loss: ; in, This indicates the number of structural parameters predicted during the inverse denoising process. One parameter; and They represent the first, second, and third structural parameters predicted during the inverse denoising process, respectively. The minimum and maximum allowed values ​​for each parameter. This applies when the predicted structure is within the process-allowed range. Otherwise, it will increase as the degree of boundary violation increases.

[0059] To avoid the conditional diffusion model generating a large number of similar structures under the same objective conditions, this embodiment further introduces a candidate structure diversity constraint loss: ; in, Indicates the number of candidate structure parameters generated; and They represent the normalized i-th and the One candidate structural parameter; This represents the distance adjustment coefficient. This loss is used to improve the differentiation between candidate structures, providing a richer initial search space for subsequent reinforcement learning.

[0060] During training, training samples are randomly sampled from the dataset constructed in the simulation. And uniformly sample the number of diffusion steps from the diffusion step set. , This represents a discrete uniform distribution. Standard Gaussian noise is then sampled. And construct the i-th noisy sample according to the forward diffusion formula. Then, process the noisy sample... diffusion step number i, operating wavelength and transmission coefficient Input the conditional diffusion model to obtain the predicted noise. Calculate the error between the predicted noise and the actual noise based on the joint loss function, and update the network parameters using the AdamW optimizer until the joint loss function converges or the preset number of training epochs is reached.

[0061] After training, this embodiment uses the trained conditional diffusion model as a generator. Given a target operating wavelength... and target transmission coefficient Then, initial noise samples are sampled from a standard normal distribution, and then reverse denoising is performed in T iterations: ; in, This represents the sample of noisy coupler structure parameters at step i; This represents the structural parameter sample after denoising in step i-1; This represents the noise predicted by the conditional diffusion model; This represents the standard deviation of the backsampling noise in the i-th step; This represents standard Gaussian noise that follows a normal distribution. Repeating this process K times yields a candidate set of initial coupler structure parameters. This serves as input to the subsequent reinforcement learning optimizer, further optimizing the coupler structure parameters.

[0062] Step 4: Building the reinforcement learning model.

[0063] To further improve the accuracy of coupler structure parameters generated by the conditional diffusion model, this embodiment constructs a reinforcement learning model as an optimizer to further optimize the candidate set of target initial coupler structure parameters generated by the conditional diffusion model. The reinforcement learning agent continuously interacts with the simulation environment, adjusting the design scheme based on the current device structure parameters to further improve the target transmission coefficient while meeting process constraints, thereby obtaining coupler structure parameters with superior performance.

[0064] (1) State space and action space.

[0065] In reinforcement learning, the state space describes all environmental information that the agent can acquire at the current moment, while the action space describes the structural parameter adjustment operations that the agent can perform. In this embodiment, the reinforcement learning state space is defined as: ; in, Let represent the initial state of the i-th reinforcement learning iteration. This indicates the current waveguide width of the coupler. This indicates the current waveguide height of the coupler. This indicates the current waveguide spacing of the coupler. This indicates the length of the coupling region of the current coupler. This indicates the bending radius of the current coupler. This represents the target transmission coefficient corresponding to the current i-th combination of structural parameters. As reinforcement learning training progresses, the state vector is updated in real time according to the actions performed by the agent, but its dimension remains unchanged throughout the training process. At each step, the agent adjusts the coupler structural parameters by performing actions; its action space is defined as: ; in, These represent the adjustment amounts for each structural parameter in step i. Since the device structural parameters are limited by the manufacturability of the process, the structural parameters after the action must satisfy: ; in, , and The process allowable ranges for each structural parameter, To be the minimum value, This is the maximum value. When the updated structural parameters exceed the allowable range of the process, the reinforcement learning environment will impose a penalty in the reward function to ensure that the optimization result meets the actual manufacturing requirements.

[0066] (2) Reward function design.

[0067] The reward function evaluates the contribution of each action of the agent to the performance improvement of the coupler device and guides the agent to continuously optimize the target transfer coefficient. This embodiment uses the transfer coefficient as the reward, while introducing a parameter out-of-bounds penalty term to balance coupler device performance and manufacturability. The reward function is specifically defined as follows: ; in, This represents the immediate reward obtained in step i; The transmission coefficients are the values ​​corresponding to the updated coupler structure. This is the penalty coefficient; This is a parameter out-of-bounds penalty term. The parameter out-of-bounds penalty term is defined as follows: ; in, This represents the k-th structural parameter; and These represent the minimum and maximum allowable values ​​for the corresponding structural parameters, respectively. When the updated structural parameters are within the allowable range of the process, At this point, the agent receives the full reward; when the structural parameters exceed the allowable range of the process, the penalty increases with the degree of deviation, thereby reducing the reward obtained for the corresponding action and guiding the agent to gradually learn structural parameters that meet the process constraints and have high transmission performance.

[0068] (3) Construction of strategy network and value network.

[0069] In this embodiment, the reinforcement learning agent consists of two parts: a policy network and a value network. Its structure remains a simple fully connected neural network, with both the policy network and the value network designed as three-layer fully connected neural network structures. The policy network generates actions based on the current state, while the value network evaluates the value of the current state. The two networks have similar structures but their parameters are independent. The policy network uses a Gaussian policy, and its output action probability distribution is expressed as: ; in, Represents a policy network; Indicates the policy network parameters; and These represent the mean and standard deviation of the action, respectively. Indicates an action, The state is represented as follows. The output of the l-th fully connected layer of the policy network is represented as: ; in, and They represent the first Layer weight matrix and bias vector; and They represent the first Layer and first The hidden layer outputs are represented by ReLU, where ReLU denotes the activation function. The input of the first layer satisfies... Strategy Network Mean Output Layer Represented as: ; in, This represents the weight matrix of the mean output layer. This represents the bias vector of the mean output layer; This indicates the maximum allowable range of adjustment for the action. The output of the third fully connected layer of the policy network is represented by tanh, where tanh represents the hyperbolic tangent function. The action standard deviation output layer is represented as follows: ; in, and These represent the standard deviation output layers respectively. The weight matrix and bias vector are given. Furthermore, the value network is primarily used to evaluate the state value corresponding to the current state; its network structure is consistent with the policy network, but it is trained with independent parameters. The output of the l-th hidden layer is represented as: ; in, and Represent the value network of the first Layer weight matrix and bias vector. Value network output representation: ; in, Represents the parameters of the value network. and These represent the weight matrix and bias vector of the output layer of the value network, respectively. This represents the output of the third hidden layer of the value network. The network structure described above is determined before training begins, and the network weight parameters are determined during training. and It will be continuously updated.

[0070] (4) Near-end strategy optimization algorithm.

[0071] This embodiment employs a proximal policy optimization algorithm to train a reinforcement learning agent. This algorithm improves policy optimization efficiency while maintaining training stability by introducing a shearing mechanism during policy updates. During training, the agent first interacts with the simulation environment to collect experience trajectories. Each experience trajectory consists of a state, action, reward, and next state. For each step i, the temporal difference error is calculated: ; in, As a reward discount factor, Representing the value network in relation to state The value estimate, Representing the value network in relation to state Value estimation. To further reduce the variance of policy gradient estimation, this embodiment uses the generalized dominance estimation method to calculate the dominance function, which is defined as: ; in, This represents the smoothing coefficient of the generalized advantage estimation (GAE). Indicates the first The temporal difference error of the step. The advantage function is used to measure the superiority or inferiority of an action relative to the average level of the current policy. When When, it indicates that the current action is better than the average strategy; when If the current action is worse than the average policy, it indicates that the current action is inferior to the average policy. To avoid excessive policy update fluctuations leading to training instability, near-end policy optimization uses an importance sampling ratio: ; in, This indicates the importance sampling ratio for step i; Indicates the current state of the policy network. Select action The probability of; This indicates that the policy network selects actions under the same conditions before the update. The probability of; Indicates the current policy network parameters; This represents the policy network parameters before the update. It measures the probability ratio of choosing the same action under the same conditions given the old and new policies. The policy network is optimized using a surrogate-cutting objective function, which is expressed as: ; in, Represents the policy network loss function; This represents the mathematical expectation of an empirical sample. This represents the estimated value of the advantage function at step i; Represents the shearing function; The shearing coefficient is set to 0.2 in this embodiment to limit the policy update magnitude and improve training stability. The value network uses mean squared error as the optimization objective, and its loss function is expressed as: ; ; in, Represents the value network loss function; Representing the value network in relation to state Value estimation; Indicates the parameters of the value network; Indicates the value of the target state; This represents the estimated value of the advantage function.

[0072] During training, after collecting a certain number of experience trajectories, small batches of training data are randomly sampled from the experience pool to jointly optimize the policy network and value network. As training continues, the agent gradually learns action policies that can obtain higher cumulative rewards, thereby further optimizing the initial coupler structure parameters generated by the conditional diffusion model, and finally obtaining the target coupler structure parameters that meet the target operating wavelength and target transmission performance requirements.

[0073] Step 5: Co-optimization between generator and optimizer.

[0074] In this embodiment, the generator and optimizer employ a phased training and cascaded inference approach to achieve coupler inverse design. During the training phase, the generator first independently completes pre-training based on the training dataset constructed in step two, aiming to learn the operating wavelength through a denoising and diffusion process. and transmission coefficient The generator analyzes the distribution patterns of coupler structure parameters under certain conditions, thereby enabling it to generate initial coupler structure parameter candidates that conform to physical laws. After the generator is trained, its network parameters are saved in a fixed manner and used to generate initial coupler structure parameter candidates for the reinforcement learning optimizer.

[0075] The optimizer is trained independently after the generator has completed its pre-training. At the beginning of each training round, the optimizer first samples the candidate set of initial coupler structure parameters from the already trained generator: ; in, This represents the candidate set of initial coupler structure parameters output by the generator; This represents the parameters of the k-th initial coupler structure generated by the generator; This indicates that the conditional diffusion model is in operation at the wavelength. and transmission coefficient The distribution of structural parameters learned under the given conditions; Indicates the generator network parameters; This represents the number of candidate samples for the structural parameters. Subsequently, the structural parameters output by the generator are used as the initial state of the reinforcement learning optimizer. The policy network outputs an action based on the current state: ; in, This represents the action output by the policy network at step i; Indicates the policy network in state The probability distribution of the next output action; This represents the policy network parameters. Then, based on... Update the coupler structural parameters. The updated structural parameters are used to calculate the corresponding transmission coefficients through electromagnetic simulation, and then applied to the reward function: ; in, This represents the immediate reward at step i; This represents the transmission coefficients corresponding to the updated structure. Indicates the target transmission coefficient; Indicates the penalty coefficient; This represents the penalty term for parameters exceeding the bounds. An immediate reward is obtained, and the advantage function is further calculated: ; in, This represents the estimated value of the advantage function at step i; Indicates time difference error; Indicates the reward discount factor; Let represent the smoothing coefficient of the Generalized Advantage Estimation (GAE); T represents the length of a single training trajectory. Combine this with the agent's objective function, which is then pruned by the policy network. ; And the loss function of the value network: ; in, The loss function of the policy network is represented; This represents the mathematical expectation of an empirical sample. Indicates the importance sampling ratio; Represents the shearing function; This represents the shear coefficient. Value network loss function; Represents the network parameters; Representation value network pairs of states Value estimation; Represents the target state value. (Regarding policy network parameters) Sum value network parameters Iterative updates are performed to continuously improve the transmission performance of the coupler structure parameters.

[0076] In the prediction phase, the generator and optimizer work in series, with data flowing sequentially to form a complete coupler reverse design process. The input target operating wavelength... and target transmission coefficient Then, the generator first uses standard Gaussian noise as initial input and generates a set of candidate parameters for the target initial coupler structure through a reverse denoising process: ; Subsequently, the candidate structural parameters from the candidate set of initial coupler structural parameters are sequentially input into the reinforcement learning optimizer as its initial state: ; in, This represents the initial reinforcement learning state corresponding to the k-th candidate structural parameter; This represents the target transmission coefficient corresponding to the k-th candidate structural parameter. The optimizer then performs multi-step fine-tuning on each group of candidate structures, adjusting the coupler structural parameters by outputting actions through the policy network, calculating the updated transmission coefficient using electromagnetic simulation, and continuously updating the agent state based on the reward function. Finally, the coupler structural parameter with the highest transmission coefficient is selected from the endpoints of all optimized trajectories as the final design result. ; in, This represents the final optimized coupler structure parameters (i.e., the target coupler structure parameters). This represents the final structure parameters corresponding to the k-th optimized trajectory; This represents the transmission coefficient corresponding to the final structural parameters.

[0077] In this embodiment, the generator is responsible for generating initial coupler structure parameter candidates that conform to physical laws, and the optimizer is responsible for target-oriented local fine optimization of the candidate structure. The two are effectively connected through one-way data transmission, realizing end-to-end mapping from the target operating wavelength and target transmission coefficient to the optimal coupler structure parameters (i.e., the target coupler structure parameters), thereby achieving efficient reverse design of the coupler.

[0078] This embodiment, based on the simulation of a small number of coupler devices, implements a general reverse design method for couplers using a conditional diffusion model and a reinforcement learning model in synergy. The technical solution of this embodiment has the following advantages: 1. This embodiment proposes a reverse design paradigm that uses the target operating wavelength and target transmission coefficient to jointly guide the generation of coupler structural parameters. By embedding the target operating wavelength and target transmission coefficient as core design conditions into a conditional diffusion model, the generation process is directly guided by the target conditions. This realizes a paradigm shift from the traditional "coupler structural parameters determine device performance" to "target operating wavelength and target transmission coefficient guide the generation of coupler structural parameters," significantly reducing the computational resources and time costs caused by repeated simulations, and improving the flexibility and efficiency of device design.

[0079] 2. This embodiment designs a cascaded optimization architecture in which the generator and optimizer work together. The generator is responsible for generating candidate initial coupler structure parameters that conform to physical laws based on the target operating wavelength and the target transmission coefficient, effectively alleviating the problems of low efficiency and high simulation resource consumption in the initial random exploration of traditional reinforcement learning in high-dimensional design space; the optimizer is responsible for target-oriented local fine optimization of the candidate initial coupler structure parameters, making up for the shortcomings of the conditional diffusion model in generating results that lack performance orientation and are difficult to directly obtain high transmission coefficient structures.

[0080] 3. This embodiment achieves a balance between physical feasibility and optimal performance through the cascaded collaboration of a generator and an optimizer. The generator, based on parameter distribution patterns learned from training data, ensures that the generated initial coupler structure parameter candidates satisfy physical constraints and manufacturability. Under process constraints, the optimizer, with the goal of improving the target transmission coefficient, performs local fine-tuning of the candidate structures, ultimately obtaining coupler structure parameters that meet the target operating wavelength requirements and have a high transmission coefficient. This embodiment avoids the problem of structural deviations from physical realizability due to a simple pursuit of performance, and effectively reduces the possibility of missing the global optimal solution by relying on empirical design, thereby improving the accuracy of coupler parameter design.

[0081] Reference Figure 4 This application also provides a coupler reverse design system, which includes: The model building module 100 is used to build a waveguide coupler simulation model in the simulation software, which includes a waveguide and a substrate waveguide, and to determine the initial structural parameters of the waveguide coupler simulation model. The simulation analysis module 200 is used to calculate the refractive index of the waveguide and the refractive index of the substrate waveguide at different operating wavelengths; the refractive index of the waveguide, the refractive index of the substrate waveguide, and the initial structural parameters are input into the waveguide coupler simulation model for intrinsic simulation analysis to determine the target structural parameter range and the transmission coefficients corresponding to different combinations of structural parameters within the target structural parameter range; The first training module 300 is used to train the forward performance prediction model using a training dataset containing different operating wavelengths, transmission coefficients and different combinations of structural parameters, to obtain a trained forward performance prediction model; different operating wavelengths and different combinations of structural parameters are used as inputs to the forward performance prediction model, and the transmission coefficients are used as training labels for the forward performance prediction model. The second training module 400 is used to construct a joint loss function based on the trained forward performance prediction model; based on the joint loss function, the constructed conditional diffusion model is trained using the training dataset to obtain the trained conditional diffusion model; different operating wavelengths and transmission coefficients are used as input conditions for the constructed conditional diffusion model. The parameter determination module 500 is used to input the target operating wavelength and the target transmission coefficient into the trained conditional diffusion model to determine the candidate set of target initial coupler structure parameters; The coupler inverse design module 600 is used to input the structural parameters from the candidate set of initial coupler structural parameters into the trained reinforcement learning model to determine the target coupler structural parameters that satisfy the target operating wavelength and target transmission coefficient; to perform coupler inverse design using the target coupler structural parameters; and to construct the reinforcement learning model through a policy network and a value network.

[0082] It should be noted that since the coupler reverse design system in this embodiment is based on the same inventive concept as the coupler reverse design method described above, the corresponding content in the method embodiment is also applicable to this system embodiment, and will not be described in detail here.

[0083] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with relevant regulations. The acquisition, storage, use and processing of data in the technical solution of this application all comply with the relevant provisions of national laws and regulations.

[0084] Reference Figure 5 This application also provides an electronic device, which includes: At least one memory; At least one processor; At least one program; The program is stored in memory, and the processor executes at least one program to implement the coupler reverse design method described above in this disclosure.

[0085] This electronic device can be any smart terminal, including mobile phones, tablets, personal digital assistants (PDAs), and in-vehicle computers.

[0086] The electronic devices according to embodiments of this application will now be described in detail.

[0087] The processor 1600 can be implemented using a general-purpose central processing unit (CPU), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this disclosure. The memory 1700 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 1700 can store the operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 1700 and is called and executed by the processor 1600 using the coupler reverse design method of the embodiments of this disclosure.

[0088] The input / output interface 1800 is used to implement information input and output. The communication interface 1900 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.). Bus 2000 transmits information between various components of the device (e.g., processor 1600, memory 1700, input / output interface 1800, and communication interface 1900); The processor 1600, memory 1700, input / output interface 1800 and communication interface 1900 are connected to each other within the device via bus 2000.

[0089] This disclosure also provides a storage medium, which is a computer-readable storage medium storing computer-executable instructions for causing a computer to perform the above-described coupler reverse design method.

[0090] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0091] The embodiments described in this disclosure are for the purpose of more clearly illustrating the technical solutions of this disclosure and do not constitute a limitation on the technical solutions provided by this disclosure. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by this disclosure are also applicable to similar technical problems.

[0092] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this disclosure, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0093] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0094] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0095] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0096] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0097] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus 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 coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0098] 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.

[0099] Furthermore, the functional units in the various embodiments of this application 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.

[0100] 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 this application, 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 multiple instructions to cause an electronic device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks. The embodiments of this application have been described in detail above with reference to the accompanying drawings, but this application is not limited to the above embodiments. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of this application.

[0101] The embodiments of this application have been described in detail above with reference to the accompanying drawings. However, this application is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of this application.

Claims

1. A method for reverse design of a coupler, characterized in that, The method includes: A waveguide coupler simulation model, which includes a waveguide and a substrate waveguide, is constructed in simulation software, and the initial structural parameters of the waveguide coupler simulation model are determined. Calculate the refractive index of the waveguide and the refractive index of the substrate waveguide at different operating wavelengths; input the refractive index of the waveguide, the refractive index of the substrate waveguide, and the initial structural parameters into the waveguide coupler simulation model for intrinsic simulation analysis to determine the target structural parameter range and the transmission coefficients corresponding to different combinations of structural parameters within the target structural parameter range; The forward performance prediction model is trained using a training dataset containing different operating wavelengths, transmission coefficients, and different combinations of structural parameters to obtain a trained forward performance prediction model; the different operating wavelengths and different combinations of structural parameters are used as inputs to the forward performance prediction model, and the transmission coefficients are used as training labels for the forward performance prediction model. Based on the trained forward performance prediction model, a joint loss function is constructed; based on the joint loss function, the constructed conditional diffusion model is trained using the training dataset to obtain the trained conditional diffusion model; the different operating wavelengths and the transmission coefficients are used as conditions input to the constructed conditional diffusion model. The target operating wavelength and target transmission coefficient are input into the trained conditional diffusion model to determine the candidate set of target initial coupler structure parameters; The structural parameters in the candidate set of the initial coupler structural parameters are input into the trained reinforcement learning model to determine the target coupler structural parameters that satisfy the target operating wavelength and the target transmission coefficient; the coupler is designed in reverse using the target coupler structural parameters; the reinforcement learning model is constructed through a policy network and a value network.

2. The coupler reverse design method according to claim 1, characterized in that, The joint loss function is constructed based on the trained forward performance prediction model, including: Based on the standard Gaussian noise actually added during the forward diffusion process of the constructed conditional diffusion model and the noise predicted by the constructed conditional diffusion model, a noise prediction loss is constructed. Based on the transmission coefficients predicted by the trained forward performance prediction model and the transmission coefficients input to the constructed conditional diffusion model, a transmission coefficient consistency constraint loss is constructed. Based on the structural parameters predicted by the constructed conditional diffusion model, construct structural parameter process constraint loss that meets the actual processing requirements; Based on the constructed conditional diffusion model, multiple candidate structure parameters are generated, and a candidate structure diversity constraint loss is constructed. The noise prediction loss, the transmission coefficient consistency constraint loss, the structural parameter process constraint loss, and the candidate structure diversity constraint loss are weighted and summed to construct a joint loss function.

3. The coupler reverse design method according to claim 2, characterized in that, The step of constructing a transmission coefficient consistency constraint loss based on the transmission coefficients predicted by the trained forward performance prediction model and the transmission coefficients input to the constructed conditional diffusion model includes: The current coupler structure parameters and their corresponding operating wavelengths, predicted by the conditional diffusion model during the reverse denoising process, are input into the trained forward performance prediction model to predict the transmission coefficients. The transmission coefficients predicted by the trained forward performance prediction model and the transmission coefficients input to the constructed conditional diffusion model are subjected to squared difference calculation to construct the transmission coefficient consistency constraint loss.

4. The coupler reverse design method according to claim 2, characterized in that, The step of constructing the structural parameter process constraint loss that meets actual processing requirements based on the structural parameters predicted by the constructed conditional diffusion model includes: ； in, This represents the loss due to structural parameter process constraints. This represents the first structural parameter predicted by the constructed conditional diffusion model during the reverse denoising process. One parameter, This indicates that the structure parameters predicted by the constructed conditional diffusion model during the reverse denoising process are of the th order. The maximum value of each parameter. This indicates that the structure parameters predicted by the constructed conditional diffusion model during the reverse denoising process are of the th order. The minimum value of each parameter. This indicates taking the maximum value.

5. The coupler reverse design method according to claim 2, characterized in that, The step of generating multiple candidate structure parameters based on the constructed conditional diffusion model and constructing a candidate structure diversity constraint loss includes: ； in, This represents the candidate structure diversity constraint loss. Indicates the number of candidate structure parameters. Represents the normalized i-th One candidate structural parameter, Represents the normalized i-th One candidate structural parameter, This represents the distance adjustment coefficient. Denotes the square of the L2 norm. This represents an exponential function.

6. The coupler reverse design method according to claim 1, characterized in that, The step of training the constructed conditional diffusion model using the training dataset based on the joint loss function to obtain the trained conditional diffusion model includes: A fully connected neural network is used to construct a conditional diffusion model, resulting in the constructed conditional diffusion model. The operating wavelength, transmission coefficient, diffusion steps, and noisy samples in the training dataset are used as inputs to the constructed conditional diffusion model; the diffusion steps are mapped to a diffusion step embedding vector through sinusoidal position encoding; the operating wavelength and transmission coefficient are linearly mapped to obtain wavelength embedding vector and transmission coefficient embedding vector, respectively. The condition vector is obtained by adding the diffusion step embedding vector, the wavelength embedding vector, and the transmission coefficient embedding vector. Based on the joint loss function, the conditional vector and the noisy sample are input into the constructed conditional diffusion model for training until the joint loss function converges or reaches the preset number of training rounds, thus obtaining the trained conditional diffusion model.

7. The coupler reverse design method according to claim 1, characterized in that, The step of inputting the structural parameters from the candidate set of initial coupler structural parameters into a trained reinforcement learning model to determine the target coupler structural parameters that satisfy the target operating wavelength and the target transmission coefficient includes: A policy network and a value network are constructed using a multi-layer fully connected neural network; a reinforcement learning model is then constructed based on the policy network and the value network. The structural parameters and their corresponding transmission coefficients obtained in the initial coupler structural parameter candidate set during the training of the constructed conditional diffusion model are used as the initial state for reinforcement learning. The initial state is input into the policy network to obtain the action output by the policy network; the current structure parameters are updated according to the action output by the policy network to obtain the updated structure parameters. Based on the updated structural parameters, a reward function is constructed; an immediate reward is obtained through the reward function, and an advantage function is calculated based on the immediate reward; Based on the advantage function, a loss function for the policy network is constructed; using mean squared error as the optimization objective, a loss function for the value network is constructed. Based on the loss function of the policy network and the loss function of the value network, the policy network and the value network are iteratively trained until the loss function converges or the preset number of training rounds is reached, thus obtaining a trained reinforcement learning model. The structural parameters in the candidate set of initial target coupler structural parameters are input into the trained reinforcement learning model to determine the target coupler structural parameters that satisfy the target operating wavelength and the target transmission coefficient; the target coupler structural parameters are the structural parameters with the highest corresponding transmission coefficient.

8. A coupler reverse design system, characterized in that, The system includes: The model building module is used to build a waveguide coupler simulation model in the simulation software, which includes a waveguide and a substrate waveguide, and to determine the initial structural parameters of the waveguide coupler simulation model. The simulation analysis module is used to calculate the refractive index of the waveguide and the refractive index of the substrate waveguide at different operating wavelengths; the refractive index of the waveguide, the refractive index of the substrate waveguide, and the initial structural parameters are input into the waveguide coupler simulation model for intrinsic simulation analysis to determine the target structural parameter range and the transmission coefficients corresponding to different combinations of structural parameters within the target structural parameter range; The first training module is used to train the forward performance prediction model using a training dataset containing different operating wavelengths, transmission coefficients, and different combinations of structural parameters to obtain a trained forward performance prediction model; the different operating wavelengths and different combinations of structural parameters are used as inputs to the forward performance prediction model, and the transmission coefficients are used as training labels for the forward performance prediction model. The second training module is used to construct a joint loss function based on the trained forward performance prediction model; to train the constructed conditional diffusion model using the training dataset based on the joint loss function to obtain the trained conditional diffusion model; and to use the different operating wavelengths and the transmission coefficients as conditions for inputting the constructed conditional diffusion model. The parameter determination module is used to input the target operating wavelength and the target transmission coefficient into the trained conditional diffusion model to determine the candidate set of target initial coupler structure parameters; The coupler inverse design module is used to input the structural parameters from the candidate set of the initial coupler structural parameters into the trained reinforcement learning model to determine the target coupler structural parameters that satisfy the target operating wavelength and the target transmission coefficient; and to perform coupler inverse design using the target coupler structural parameters; the reinforcement learning model is constructed through a policy network and a value network.

9. An electronic device, characterized in that, It includes at least one control processor and a memory for communicatively connecting to the at least one control processor; the memory stores instructions executable by the at least one control processor to enable the at least one control processor to perform the coupler reverse design method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions for causing a computer to perform the coupler reverse design method as described in any one of claims 1 to 7.

Images