A deep learning-based micro-ring resonator reverse design optimization method
By using a deep learning method that cascades forward prediction networks and inverse generation networks, the design of micro-ring resonators is optimized, solving the problems of low design efficiency and high cost of photonic devices. This enables rapid and high-precision design and adaptation to complex target constraints, supporting the automation and industrialization of photonic devices.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- FUDAN UNIVERSITY
- Filing Date
- 2026-01-28
- Publication Date
- 2026-06-05
AI Technical Summary
Existing photonic device design methods are inefficient and computationally expensive in complex and high-dimensional designs, and it is difficult to reuse computation results. Deep learning faces problems such as high data acquisition costs, unbalanced sample distribution, and insufficient physical manufacturability constraints in photonic device design.
By cascading forward prediction networks and backward generation networks, combined with a forward consistency constraint mechanism, and through closed-loop verification of hybrid density backward and forward networks, the design of the micro-ring resonator is optimized. Deep learning is used to achieve bidirectional verification and optimization of performance and structure.
It enables rapid and high-precision design of micro-ring resonators, reduces computational costs, improves design efficiency, and can adapt to complex target constraints, making it suitable for the automated design and industrialization of large-scale photonic devices.
Smart Images
Figure CN122154400A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of photonic device design technology, specifically relating to a deep learning-based reverse design optimization method for microring resonators. Background Technology
[0002] Photonic devices are the core foundation of optical communication, optical computing, and sensing. Common devices include couplers, microring resonators, wavelength division multiplexers, and filters. Most existing designs rely on analytical theory, but these methods have significant limitations in expanding device size and structure, especially for complex structures lacking analytical models. Reverse engineering, on the other hand, starts with the target performance and iteratively updates the device structure using optimization algorithms and Maxwell's equations. This not only yields designs that meet requirements but also expands the design space. However, as the dimensions of device design increase, traditional optimization becomes significantly less efficient. The mapping relationship between target performance and structural parameters is unclear, and different target requirements often require separate optimization processes, making it difficult to reuse existing calculation results and resulting in high computational costs.
[0003] In recent years, deep learning has developed rapidly. Based on existing structural and performance data, models can learn the relationship between the two. Introducing deep learning into device design enables a two-way function of predicting performance from structure and deriving structure from performance, improving the efficiency of device design. In addition, deep learning network models have also shown good generalization ability, providing new ideas for the development of complex photonic devices. Although existing data-driven methods have been verified to be effective on some devices, they still face problems such as high data acquisition costs, uneven sample distribution, insufficient integration of physical manufacturability constraints, and multiple solutions for the target-structure relationship, which affect their widespread application in complex devices and engineering scenarios. Summary of the Invention
[0004] The purpose of this invention is to provide a deep learning-based reverse design optimization method for microring resonators to meet the needs of next-generation high-speed optical communication and on-chip integrated photonic systems.
[0005] This invention combines a forward prediction network and a backward generation network to improve design efficiency while maintaining accuracy. Specifically, a forward consistency constraint mechanism is introduced, inputting the backward-predicted structural parameters into the trained forward network for correction. This achieves bidirectional verification and optimization between performance and structure, thus avoiding bias issues caused by relying solely on mean square error. This method enables rapid and high-precision back-calculation of the free spectral range and quality factor of microring resonators, offering advantages such as high computational efficiency, strong scalability, and adaptability to complex target constraints. It has significant application value for the future automated design and industrialization of large-scale photonic devices.
[0006] The deep learning-based resonant cavity reverse design optimization method provided by this invention includes the following steps:
[0007] Step 1, Data Acquisition and Partitioning; specifically includes:
[0008] Obtain structural parameters for different devices, including the radius R of the microring resonator, the waveguide thickness H, the waveguide width W, and the coupling distance between the waveguide and the microring Gap, as well as data samples consisting of the corresponding performance indicators, free spectral range (FSR) and quality factor Q.
[0009] The obtained data samples are divided into training set, validation set and test set; the training set is used to train the network model, the validation set is used to check whether the model is overfitting during the training process, and the test set is used to test the performance of the final network model.
[0010] Step 2: Pre-train the forward prediction network to learn the mapping from micro-ring structure parameters to performance indicators on the training set. After the target accuracy is achieved, the parameters are frozen for subsequent constraints.
[0011] Step 3: Under the condition of freezing the forward prediction network parameters obtained in Step 2, the inverse network is cascaded with the forward prediction network so that the structural parameters output by the inverse network can obtain the corresponding prediction performance through the frozen forward prediction network, thus forming a closed loop of "performance → structure → performance".
[0012] Step 4: Under the condition of freezing the forward network parameters obtained in Step 2, the inverse network is trained using joint loss. During the training process, only the parameters of the inverse network are updated. The joint loss includes: the mean square error between the target performance and the prediction performance of the forward network and the negative log-likelihood with respect to the structural parameters, which are weighted and summed.
[0013] Step 5: Change the weights in the joint loss and compare them to select the optimal network model; specifically:
[0014] Repeat step 4 within the preset discrete weight set to complete the training and save each model. Then, compare the models based on the minimum combined error of FSR and Q on the validation set to determine the optimal weights and the final model.
[0015] Finally, after determining the weight settings, the corresponding cascaded network model (final model) is saved, and reverse design is performed on the test set with the target performance as input to obtain the corresponding structural parameters. The predicted / simulated response of the structural parameters is compared with the actual response of the test set, and the mean absolute percentage error is used for evaluation to measure the effect of reverse design.
[0016] In this invention, the data samples containing structural parameters R, W, H, and Gap, as well as performance indices FSR and Q, are first cleaned, standardized, and divided into training, validation, and test sets. Then, a forward prediction network is built, taking the device structural parameters as input and outputting the performance indices. Based on this, a hybrid density inverse network is constructed, taking the target performance as input and the device structure as output, and cascaded with the forward network. The joint loss function is used to update the inverse network parameters until convergence. Finally, the network is validated on the test set, and device structures are generated in batches within a preset target performance range. After physical feasibility and forward network screening, the structures are output. Simulations are then conducted to obtain the actual performance, which is compared with the predicted performance to evaluate the effectiveness of the inverse design.
[0017] The main technical features and beneficial effects of this invention are as follows:
[0018] By cascading a hybrid density inverse network and a forward prediction network to construct a closed loop from target performance to structure and back to performance, the multiple solutions to the inverse problem are resolved, improving prediction stability and accuracy. Probabilistic modeling is used to learn the structure distribution, generating several structures under a given target. Physical feasibility screening and forward model evaluation are combined to select the best structure, ensuring the engineering usability of the output structure. The proposed method does not rely on analytical models, can directly interface with simulation and experimental data, and can operate stably under complex targets and constraints, shortening the design iteration cycle and reducing costs. Attached Figure Description
[0019] Figure 1 This is a device structure diagram of the micro-ring resonator in an embodiment of the present invention.
[0020] Figure 2 This is a network model diagram of the reverse design method in an embodiment of the present invention.
[0021] Figure 3 This is a diagram illustrating the steps involved in implementing the algorithm of this invention.
[0022] Figure 4 The diagram shows the predictions of the forward model of this invention. In the diagram, (a) is the prediction of the free spectral range (FSR) by the forward model, and (b) is the prediction of the quality factor Q by the forward model.
[0023] Figure 5 This is a comparison graph of the actual response and the model predicted response in an embodiment of the present invention. Among them, (a) is a comparison of the actual and predicted free spectral range (FSR), and (b) is a comparison of the actual and predicted quality factor (Q).
[0024] Figure 6 To verify the physical realizability in this embodiment of the invention, the response obtained by inputting the structural parameters output by the reverse network into electromagnetic simulation software is compared with the actual response of the target. Figure 5(a) Comparison between the actual FSR response and the response obtained from the simulation of the predicted structure output by the inverse network. Figure 5 (b) Comparison of the actual Q response and the predicted structural response output by the inverse network. Figure 5 (c) and Figure 5 (d) Error histograms for FSR and Q are given respectively, and the results show that the predicted structure can also reproduce the target performance well in actual simulation. Detailed Implementation
[0025] The present invention will be further explained in detail below with reference to the accompanying drawings and specific embodiments, but it should be understood that the scope of protection of the present invention is not limited to the specific embodiments.
[0026] This invention addresses the inverse design problem of microring resonators, proposing a cascaded fully connected network consisting of a forward prediction network and a hybrid density inverse network. The cascaded network model is as follows: Figure 2 As shown in Table 1, the inverse network takes the free spectral range (FSR) and quality factor (Q) of the target performance as input, passes them through three fully connected hidden layers consisting of 128, 256, and 128 neurons, and outputs the corresponding structural parameters R, W, H, and Gap. The forward network takes these structural parameters as input and predicts their corresponding FSR and Q through three fully connected hidden layers consisting of 121, 109, and 96 neurons. This is equivalent to a forward simulation operation, used to compare the predicted performance of the new design structure with the target performance. The hyperparameter settings of the cascaded network model are shown in Table 1.
[0027] Table 1 Model hyperparameter settings
[0028] .
[0029] The algorithm steps in this embodiment are as follows: Figure 3 As shown, to support the modeling and verification of the above network, a structure-performance aligned data sample is first constructed. The device structure set in this embodiment is as follows. Figure 1 As shown, the structural parameters and their sampling ranges are as follows: the radius R of the microrings ranges from 20 μm to 90 μm, the thickness from 0.3 μm to 0.45 μm, the width from 0.4 μm to 1 μm, and the coupling spacing from 250 nm to 850 nm. The total dataset size is 1698 records. The data comes from full-wave electromagnetic simulations using the finite-difference time-domain method and is divided into training, validation, and test sets in an 8:1:1 ratio. The training set is used to train the network model, the validation set is used to check for overfitting during training and determine the model with the lowest validation error, and the test set is used to test the performance of the final network model.
[0030] In network model training, the forward network is first pre-trained, and the parameters of the trained forward network model are fixed. Then, the overall inverse design network is trained. The loss function of the forward prediction network model is:
[0031]
[0032] Where λ is the weight and FSR is the true free spectral range. Here, represents the free spectral range predicted by the forward network, and Q is the quality factor of the true freedom. This represents the quality factor predicted by the positive network.
[0033] The mean absolute percentage error of the trained network model on the test set is shown in Table 2. The final weight was set to 5. Figure 4 This demonstrates the performance of the positive prediction model on the test set. Figure 4 (a) shows the prediction results for FSR. Figure 4 (b) Prediction results for the corresponding quality factor Q
[0034] Table 2 Weights for different loss functions MAPE performance of forward network in predicting FSR and Q
[0035] .
[0036] The loss function for training the inverse network includes the mean squared error between the target performance and the prediction performance of the forward network, as well as the negative log-likelihood with respect to the structural parameters. Its expression is as follows:
[0037]
[0038] In formula (2), K is the number of mixed distributions, which is set to 5 in this example; D is the dimension of the structural geometric parameters R, W, H and Gap, i.e. D is 4; This represents the component of the d-th dimension of the structural parameter in the normalized space; and The first The Gaussian component in the th... The mean and standard deviation of the dimension are given, and α is the weight of the loss function, where α > 0.
[0039] During training, the network calculates the loss of the input data samples and then uses backpropagation to optimize and adjust the weights and biases of each layer in the neural network. Through multiple iterations, a network model with the minimum error is obtained.
[0040] To obtain the optimal weights, different numerical weights α were scanned and substituted into the loss function for training. The mean absolute percentage error of the trained network model on the test set is shown in Table 3. Finally, the weight α was set to 1.
[0041] Table 3 shows the MAPE performance of the cascaded network model for FSR and Q under different loss function weights α.
[0042] .
[0043] After obtaining the network model with the lowest validation error, reverse design validation was performed on the test set samples, and the results are as follows. Figure 5 As shown. The specific process is as follows: First, the performance metrics corresponding to the real structural parameters in the test set are used as input, and the predicted structural parameters are obtained by inverse designing the network; then, the predicted structure is input into the forward network to calculate the corresponding performance metrics. Finally, the predicted performance output by the forward network is compared with the performance corresponding to the real structure in the test set. Figure 5 (a) shows a comparison between the actual response and the predicted response of the FSR. Figure 5 (b) shows a comparison between the actual response and the predicted response of Q.
[0044] Furthermore, to verify the physical realizability of the model's predicted structure, the predicted structure parameters output by the inverse network were re-input into the electromagnetic simulation software, and the resulting performance metrics were compared with the target performance, such as... Figure 6 As shown, where Figure 6 (a) Comparison between the actual FSR response and the response obtained from the simulation of the predicted structure output by the inverse network. Figure 6 (b) Comparison of the actual Q response and the predicted structural response output by the inverse network. Figure 6 (c) and Figure 6 (d) Error histograms for FSR and Q are given respectively, and the results show that the predicted structure can also reproduce the target performance well in actual simulation.
[0045] The above are merely preferred embodiments of the present invention, and the present invention is not limited to the contents of these embodiments. For those skilled in the art, various changes and modifications can be made within the scope of the technical solutions of the present invention, and any changes and modifications made are within the protection scope of the present invention.
Claims
1. A reverse design method for microring resonators based on deep learning, characterized in that, By combining a forward prediction network and a reverse generation network, design efficiency is improved while maintaining accuracy. Specifically, a forward consistency constraint mechanism is introduced, inputting the structural parameters predicted by the reverse generation network into the trained forward network for correction. This achieves bidirectional verification and optimization between performance and structure, avoiding bias caused by relying solely on mean square error, and enabling rapid and high-precision back-calculation of the free spectral range and quality factor of the microring resonator. The specific steps are as follows: Step 1, Data Acquisition and Partitioning; specifically includes: The data sample consists of structural parameters of different devices, including the radius R of the microring resonator, the waveguide thickness H, the waveguide width W, and the coupling distance between the waveguide and the microring, as well as the corresponding performance indicators, the free spectral range (FSR) and the quality factor Q. The obtained data samples are divided into training set, validation set and test set; the training set is used to train the network model, the validation set is used to check whether the model is overfitting during the training process, and the test set is used to test the performance of the final network model. Step 2: Pre-train the forward prediction network to learn the mapping from micro-ring structure parameters to performance indicators on the training set. After the target accuracy is achieved, the parameters are frozen for subsequent constraints. Step 3: Under the condition of freezing the forward prediction network parameters obtained in Step 2, the inverse network is cascaded with the forward prediction network, so that the structural parameters output by the inverse network can obtain the corresponding prediction performance through the frozen forward prediction network, thus forming a closed loop of "performance → structure → performance". Step 4: Under the condition of freezing the forward network parameters obtained in Step 2, the inverse network is trained using joint loss. During the training process, only the parameters of the inverse network are updated. The joint loss includes: the mean square error between the target performance and the prediction performance of the forward network and the negative log-likelihood with respect to the structural parameters, which are weighted and summed. Step 5: Change the weights in the joint loss and compare them to select the optimal network model; specifically: Repeat step 4 within the preset discrete weight set to complete the training and save each model. Then, compare the models based on the minimum combined error of FSR and Q on the validation set to determine the optimal weights and the final model.
2. The reverse design method for a microring resonator according to claim 1, characterized in that, After determining the weight settings, the corresponding cascaded network model is saved, and reverse design is performed on the test set with the target performance as input to obtain the corresponding structural parameters. The predicted / simulated response of the structural parameters is compared with the actual response of the test set, and the mean absolute percentage error is used for evaluation to measure the effect of reverse design.
3. The reverse design method for a microring resonator according to claim 1, characterized in that, The loss function for the pre-training of the forward prediction network described in step 2 is: ； Where λ is the weight and FSR is the true free spectral range. Here, represents the free spectral range predicted by the forward network, and Q is the quality factor of the true freedom. This represents the quality factor predicted by the positive network.
4. The reverse design method for a microring resonator according to claim 3, characterized in that, After pre-training the forward network, the parameters of the forward network model obtained from the training are fixed, and then the inverse network is trained. Its loss function includes the mean squared error between the target performance and the prediction performance of the forward network, as well as the negative log-likelihood with respect to the structural parameters. Its expression is as follows: ； In formula (2), K is the number of mixed distributions; D is the dimension of the structural geometric parameters R, W, H and Gap, i.e., D is 4; This represents the component of the d-th dimension of the structural parameter in the normalized space; and They are respectively the nth The Gaussian component in the th... The mean and standard deviation of the dimension are given, and α is the weight of the loss function, where α >
0.
5. The reverse design method for a microring resonator according to claim 4, characterized in that, During training, the loss of the input data samples is calculated, and then the weights and biases of each layer in the neural network are optimized and adjusted using backpropagation. Through multiple iterations, a network model with the minimum error is obtained.