Hybrid modeling method, system and device for improving avalanche photodiode current gain prediction and medium

Abstract

The invention discloses a hybrid modeling method, system and device for improving avalanche photodiode current gain prediction and a medium, and belongs to the technical field of semiconductor device simulation, the core of the hybrid modeling method is to construct a physically guided improved one-dimensional U-Net neural network, a third-order polynomial regression module is embedded at the bottleneck of the U-Net network, and the current gain prediction of the avalanche photodiode is improved. Variable parameters related to the gain are explicitly learned and input, the dependency relationship of the gain on the parameters such as the electric field and the temperature is directly modeled, and the physical rationality is ensured; meanwhile, the inherent jump connection structure of the U-Net is utilized to transmit fine features captured by the encoder to the decoder, so that high-fidelity reconstruction of key details of gain distribution is realized; the method aims at solving the core contradiction between high-precision numerical simulation and calculation efficiency in the design process. Through the design of mechanism driving, the hybrid prediction model is enabled to deeply accord with the physical nature of the avalanche photodiode, and second-level, high-fidelity and physically credible gain prediction is realized.

Description

Technical Field

[0001] This invention relates to the field of semiconductor device simulation technology, specifically to a hybrid modeling method, system, device, and medium for improving the prediction of current gain in avalanche photodiodes. Background Technology

[0002] Avalanche photodiodes (APDs) play an irreplaceable role in military reconnaissance, space communication, lidar, and weak light signal detection due to their high sensitivity resulting from the internal avalanche multiplication effect of charge carriers. APDs based on mercury cadmium telluride (HgCdTe) materials possess the ability to achieve high signal-to-noise ratio detection at the single-photon level, making them particularly suitable for complex detection scenarios such as low illumination and high dynamic range. They are also core devices for mid- and far-infrared detection. Currently, the commonly used device structure is a planar junction structure based on ion implantation technology. This structure is achieved by sequentially constructing N-type ... + Contact area, N - doubling zone and P + The absorption region forms a typical N + -N - -P + Structure; wherein, P is located at the bottom + The region, as a light absorption region, is responsible for efficiently absorbing mid-wave infrared photons and generating photogenerated electron-hole pairs; the N in the middle - The region forms a strong electric field under reverse bias and is a key functional region for avalanche multiplication; the N at the top + The absorption layer is used to achieve efficient collection and output of electrons; when infrared photons enter the absorption layer, photogenerated electrons are driven by a strong electric field to move towards the N-type electrons. - During the migration to the multiplication region, under the influence of the high electric field in this region, electrons gain sufficient kinetic energy and trigger secondary avalanche multiplication of charge carriers through collisional ionization, thereby achieving an internal gain of more than 100 times and significantly improving the detector's responsivity and signal output capability.

[0003] Currently, the design and performance evaluation of APDs heavily rely on Technology Computer Aided Design (TCAD) numerical simulation. TCAD can accurately simulate device characteristics by solving complex physical equations, but it consumes huge computational resources, and a single simulation often takes several hours, which seriously restricts the efficiency of device design iteration and parameter optimization.

[0004] In recent years, purely data-driven artificial intelligence models (such as deep learning) have been attempted as surrogate models for TCAD to improve simulation speed. However, these "black box" models have significant drawbacks: (1) Physical irrationality: The model’s predictions may seriously violate basic physical laws, such as outputting negative gain under specific parameters or abnormal results that violate the basic relationship between electric field and gain, making the prediction results unreliable.

[0005] (2) Poor detail fidelity: During the feature extraction process, the microscopic details of the sequence or image are easily lost, resulting in a distorted smooth gain curve in the prediction, making it impossible to capture key feature points and resulting in low prediction accuracy.

[0006] (3) High data dependence: Training high-performance models requires massive amounts of high-quality labeled data, and the cost of acquiring this data (whether through experiments or simulations) is extremely high; when data is insufficient, the model is prone to overfitting and has weak generalization ability.

[0007] Therefore, there is an urgent need in this field for a new technical solution that can achieve rapid prediction of avalanche photodiode current gain while ensuring physical rationality and detail fidelity. Summary of the Invention

[0008] Technical problem solved: In view of the technical problems of low efficiency of TCAD simulation and serious loss of physical reliability and detail in the prediction results of pure data-driven AI models in the existing technology, the present invention provides a hybrid modeling method, system, device and medium to improve the prediction of current gain of avalanche photodiode. The method can complete the high-fidelity prediction of current gain of avalanche photodiode in seconds, and the prediction results strictly conform to physical laws.

[0009] Technical Solution: The present invention discloses a hybrid modeling method for improving the prediction of current gain in avalanche photodiodes. The hybrid prediction model is based on a physically guided improved one-dimensional U-Net neural network architecture. The hybrid modeling method includes the following steps: Step 1, Data Preparation: Obtain multiple sets of variable parameters of the avalanche photodiode and their corresponding current gain simulation data to construct a training dataset; wherein, the variable parameters include at least one of geometric parameters, process parameters, material parameters and external operating conditions; Step 2, Model Construction: Construct a hybrid prediction model based on a one-dimensional U-Net architecture. The hybrid prediction model includes a feature extraction encoder, a feature reconstruction decoder, and a parameterized physical module integrated at the bottleneck between the encoder and decoder. Step 3, Model Training: The hybrid prediction model is trained using the training dataset; wherein, during the training process, the current gain data is input to the feature extraction encoder, and the corresponding variable parameters are input to the parameterized physics module. The parameterized physics module learns the mapping relationship between the variable parameters and the gain features processed by the encoder, and merges them with the encoder output at the bottleneck, and then reconstructs the predicted current gain through the decoder. Step 4, Gain Prediction: Input the target variable parameters of the avalanche photodiode to be predicted into the trained hybrid prediction model. The parameterized physics module outputs the physical prior information, which is then processed by the decoder to output a high-fidelity and physically reasonable current gain prediction result.

[0010] Preferably, the parameterized physics module is a multinomial regression model, and more preferably a third-order multinomial regression model.

[0011] Preferably, the feature extraction encoder and the feature reconstruction decoder are connected by a skip connection structure, which is used to stitch together the high-resolution feature maps extracted by each layer of the encoder with the feature maps of the corresponding layers of the decoder, so as to preserve the detailed features of the current gain curve during the reconstruction process.

[0012] Preferably, the variable parameter includes at least one of the following: Geometric parameters include the thickness of each zone, the width of each zone, the inclination angle of the table sidewall, the table diameter, the width of the protective ring, and the spacing between the protective rings; Process parameters include doping concentration in each region, spatial distribution of impurities, surface roughness of the etched area, and thickness of the filling layer; Material parameters include the cadmium composition, defect density, and dielectric constant of the filler material in the HgCdTe material; External operating conditions include operating temperature, incident light angle, and incident light wavelength.

[0013] Preferably, the avalanche photodiode has a structure that is planar, mesa, guard ring, or etch-fill structure; wherein the planar avalanche photodiode structure includes planar PIN type and planar N-on-P type; and the mesa avalanche photodiode structure includes mesa PIN type and mesa N-on-P type.

[0014] Preferably, the hybrid modeling method uses an incremental learning strategy to optimize the trained model. When new training data is obtained, the model is trained and fine-tuned using all existing data based on the original model parameters.

[0015] This invention also discloses a fast prediction system for avalanche photodiode gain, used to implement the above-mentioned hybrid modeling method for improving avalanche photodiode current gain prediction, comprising: The data acquisition module is used to acquire the variable parameters and current gain data; The preprocessing module is used to standardize the variable parameter and current gain data; The gain prediction module stores the hybrid prediction model trained by the above hybrid modeling method. It is used to receive the processed target variable parameters and output the gain prediction value. The result output module is used to display or transmit the predicted gain value.

[0016] The present invention also discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements a hybrid modeling method when executing the program.

[0017] The present invention also discloses a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described hybrid modeling method.

[0018] The working principle of the hybrid modeling method of this invention is as follows: the native U-Net model is trained by inputting the corresponding sequence, then propagates forward through the encoder-decoder, calculates the error through the loss function, and then propagates backward to update the network weights. Its predictive ability depends entirely on the quality and coverage of the training data, and it does not involve the influence factors of the device in the whole process. Therefore, it cannot reveal the causal laws behind the performance, and due to its poor extrapolation ability, it is insufficient in guiding the design of novel and groundbreaking devices. Figure 1 As shown, the multinomial regression module introduced in the hybrid model of this invention serves as the input port for variable parameters that are directly related to the semiconductor device, such as process parameters, material parameters, and external operating conditions. These parameters autonomously establish a relationship with the input sequence through this module, which is equivalent to giving the model the ability to physically interpret the input sequence. This allows the model's predictions to no longer be limited to sequences already existing in the training set, but to generate reasonable sequences based on the variable parameters in the device under the premise of conforming to physical laws. This module performs two functions in the entire model: first, during the training process, it receives variable parameters from the outside and sequences at the bottleneck, but does not prevent information in the encoder from being transmitted to the decoder through the bottleneck; second, during the prediction process, it receives variable parameters from the outside, transforms them into a sequence through internal conversion, transmits it to the bottleneck, and then outputs it through the decoder. The encoder does not work during this process.

[0019] Compared with the prior art, the present invention achieves the following technical effects: 1. This invention uses a U-Net architecture as its core and integrates a third-order PR hybrid prediction model to achieve fast prediction of avalanche photodiode gain. The feature extraction encoder and feature reconstruction decoder in the hybrid prediction model are connected through a residual network, avoiding the loss of original input information and thus improving prediction accuracy. The multinomial regression model introduced in this invention transforms the "black box" functional relationship between input and output in the hybrid prediction model into an understandable form, ensuring the physical rationality of the prediction results and reducing overfitting caused by insufficient training dataset. Applying the U-Net convolutional neural network used for processing medical images to predict the performance of APD devices, the prediction model for predicting the net gain of APD devices is significantly shortened compared to TCAD simulation, effectively reducing R&D costs. 2. High physical rationality: By introducing a parameterized physics module, the key physical parameters that affect the gain are explicitly embedded in the model, so that the prediction process of the model is subject to soft constraints of physical laws, effectively avoiding the generation of non-physical interpretations, and maintaining a reasonable prediction trend even in areas not covered by training data. 3. High fidelity of details: By utilizing the jump connection structure of U-Net, the shallow detail information of the encoder is directly transmitted to the decoder, which enables the predicted gain curve to highly reproduce the undulations, inflection points and other key details of the real curve, and the prediction accuracy is significantly improved. 4. Extreme prediction efficiency: After the model is trained, a single prediction only takes seconds, which is more than a thousand times faster than TCAD simulation, which takes several hours. This greatly accelerates the design iteration and parameter scanning process of APD. 5. Strong generalization and evolution capabilities: The method is applicable to various APD structures such as planar and tabletop types, requiring only the addition of corresponding structural parameters to the training data; at the same time, it supports incremental learning strategies, allowing the model to be continuously optimized as new data arrives, without the need for a complete overhaul, thus protecting initial investments and adapting to iterative R&D processes. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the overall process of dataset construction, model training, and net current gain prediction in this invention. Figure 2 This is a schematic diagram of the overall architecture of the hybrid prediction model of the present invention; Figure 3 This is a cross-sectional view of a typical planar PIN structure avalanche photodiode of the present invention; Figure 4 Design a flowchart for parameter space sampling for full factorial experiments; Figure 5 The training set of current gain curves used in the standard modeling of Embodiment 1 of this invention is the complete set of such curves. Figure 6The diagram shows the convergence curves of the loss function of the hybrid prediction model in Embodiment 1 of this invention on the training and test sets. Figure 7 This is a comparison curve between the prediction results of the hybrid prediction model and the TCAD simulation results in Embodiment 1 of the present invention; Figure 8 This refers to the set of current gain curves for each batch of training sets in the incremental learning process of Embodiment 2 of the present invention. Figure 9 This is a comparison curve of the prediction and simulation results of the initial model, the incremental learning intermediate model, and the final model in Embodiment 2 of the present invention.

[0021] Figure labeling: Ⅰ, Current gain sequence input; Ⅱ, Jump connection; Ⅲ, Prediction result output; Ⅳ, Bottleneck; Ⅴ, Variable parameter input to parameterized physics module; Ⅵ, Encoded current gain feature input to parameterized physics module; Ⅶ, Parameterized physics module output to bottleneck. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will be described in conjunction with the accompanying drawings. Figures 1-9 The technical solutions of the embodiments of the present invention are clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention are within the scope of protection of the present invention.

[0023] Example 1: As Figures 1-4 As shown, this invention discloses a hybrid modeling method for improving the prediction of current gain in avalanche photodiodes. The hybrid prediction model is based on constructing a physically guided improved one-dimensional U-Net neural network architecture. The hybrid modeling method includes the following steps: (I) Device Structure Selection and Data Preparation: The avalanche photodiode can have any one of the following structures: planar, mesa, guard ring, or etch-fill. Planar avalanche photodiodes include planar PIN and planar N-on-P structures; mesa avalanche photodiodes include mesa PIN and mesa N-on-P structures. This embodiment uses... Figure 3 The planar PIN structure HgCdTe APD shown is used as an example. From top to bottom, this structure consists of a passivation layer, an HgCdTe layer, and a CdZnTe substrate; wherein the HgCdTe layer consists of n... + District, n - District, p +For avalanche photodiodes with different structures, they share certain similarities in current gain mechanism and core input parameters, and the basic simulation process is consistent. Since the planar PIN avalanche photodiode has a simple structure and is easy to simulate and verify, this embodiment prefers to use this structure for TCAD simulation to obtain various parameters and current gain sequences. The simulation process is as follows: First, the SDE module of Sentaurus is used to design the device geometry, define the doping of each region, and generate the mesh; second, the TCLSH module is used to set the physical parameters of the mercury cadmium telluride material (dielectric constant, band gap, carrier mobility, etc.) and generate the parameter file for HgCdTe; then, the SDEVICE module is used to solve the physical model by applying boundary conditions to simulate the characteristics of the APD device; finally, the SVISUAL module is used to generate the net current gain curve.

[0024] Multiple sets of variable parameters of avalanche photodiodes and their corresponding current gain simulation data are obtained to construct a training dataset. The variable parameters include at least one of geometric parameters, process parameters, material parameters, and external operating conditions. Geometric parameters include the thickness of each region, the width of each region, the mesa sidewall tilt angle, the mesa diameter, the guard ring width, and the guard ring spacing. Process parameters include the doping concentration of each region, impurity spatial distribution, etched surface roughness, and filler layer thickness. Material parameters include the cadmium composition of HgCdTe material, defect density, and dielectric constant of the filler material. External operating conditions include operating temperature, incident light angle, and incident light wavelength. It should be noted that any valid combination of the above variable parameters falls within the scope of protection of this invention and is applicable to avalanche photodiode devices of planar, mesa, guard ring, and other structures.

[0025] This embodiment defines nine key variable parameters: operating temperature (T), cadmium composition (x), incident light wavelength (λ), and P-region doping concentration (N). a ) and thickness (W) p Intrinsic region doping concentration (N) i ) and thickness (W) i ), N-region doping concentration (N) d ) and thickness (W) n The input is a third-order polynomial regression module, used to inject physical constraints. The combination of input parameters can be selected according to the actual device structure and requirements. A skip connection structure is used to ensure that the microscopic details of the input sequence are transmitted losslessly during encoding and decoding, thereby achieving high-fidelity reconstruction of the gain curve. This is achieved using... Figure 4 The full factorial experimental design method shown systematically selects 2180 parameter combinations within the parameter space, uses Sentaurus TCAD to simulate each parameter combination, and obtains its corresponding current-gain-voltage curve (sampling 150 points), constructing a model as shown below. Figure 5 The high-fidelity training dataset shown.

[0026] (II) Model Construction: like Figure 2 As shown, a hybrid prediction model based on a one-dimensional U-Net architecture is constructed. This model includes a feature extraction encoder, a feature reconstruction decoder, and a parametric physics module integrated at the bottleneck between the encoder and decoder. The parametric physics module is a multinomial regression (PR) model. This module explicitly receives variable parameters as input, maps them to physical prior information, and fuses them with the data features extracted by the encoder at the bottleneck, injecting physical constraints into the model. The inherent skip connection II structure of U-Net is preserved and utilized to concatenate the high-resolution features of each layer of the encoder with the features of the corresponding layers of the decoder, ensuring that the microscopic details of the current gain curve are preserved during reconstruction. The feature extraction encoder consists of multiple layers of convolution and downsampling. Its head is connected to the current gain sequence input I, and its tail is connected to the feature reconstruction decoder. The connection point is called the bottleneck IV. The feature reconstruction decoder consists of multiple layers of upsampling and convolution. Its head is connected to the feature extraction encoder, and its tail is connected to the prediction result output III. The feature reconstruction decoder and the feature extraction encoder are connected through a residual network and the bottleneck at the bottom. The role of the bottleneck IV is to connect the U-Net network and the third-order polynomial regression algorithm. The solid arrow path in the PR part represents the input of the encoded current gain feature to the parameterized physical module VI and the input of the variable parameter to the parameterized physical module V during training. The dashed path represents the output of the parameterized physical module to the bottleneck VII after the new variable parameter input to the parameterized physical module V is transformed by the trained PR. The aforementioned modules are part of a U-Net-like architecture in a 1D-CNN. The residual network simply copies the features symmetrical to the feature extraction encoder in the feature reconstruction decoder and connects them to the corresponding parts of the feature reconstruction decoder. Subsequent convolutions can process the sequences of both the feature extraction encoder and the feature reconstruction decoder simultaneously, avoiding information loss. The bottleneck connects the feature reconstruction decoder and the feature extraction encoder, completing the transition from downsampling to doubling the channels to upsampling to reducing the channels. This invention introduces a parameterized physics module, explicitly embedding key physical parameters affecting gain into the model. This subjects the model's prediction process to soft constraints from physical laws, effectively avoiding non-physical interpretations and maintaining a reasonable prediction trend even in areas not covered by training data. Utilizing the skip connection structure of U-Net, it achieves direct transmission of shallow detail information from the encoder to the decoder, enabling the predicted gain curve to highly reproduce the fluctuations, inflection points, and other key details of the real curve, significantly improving prediction accuracy. Third-order polynomial regression is used to input the variable parameters corresponding to the original current gain into the hybrid prediction model. It also addresses the known strong nonlinear relationship between the input features and the target, reducing the burden on a single model to learn these fundamental nonlinear transformations and allowing it to focus on learning more complex, higher-order interaction patterns based on these fundamental features. Third-order polynomial regression uses the variable parameters and the original input sequence encoded at the bottleneck in U-Net as training parameters. New variable parameters are transformed by the trained third-order polynomial regression model and then input into the bottleneck position at the bottom of U-Net for decoding, outputting a predicted sequence.

[0027] In this embodiment, the construction is as follows: Figure 2 The hybrid prediction model shown has two 4-layer feature extraction encoders and feature reconstruction decoders, with channel numbers of [2, 4, 8, 16] and [32, 16, 8, 4] respectively. The convolutional kernel size is primarily 3×1; max pooling and downsampling both use 2×1 kernels. The encoding part, from the 36×1 to the 32×1 sequence, uses 5×1 convolutional layers, with a ReLU activation function applied element-wise to each sequence after each convolutional layer to increase the model's non-linearity and expressive power of the input sequence. Between each stage of the feature extraction encoder, a 2×1 max pooling operation is used to downsample the sequence, achieving a balance between information preservation and compression. Between each stage of the feature reconstruction decoder, a 2×1 convolutional layer is used to upsample the sequence to recover features lost during the encoding stage. In the prediction process, the third-order polynomial is used as the input of the only external parameter. The input geometric parameters, process parameters, material parameters and external working conditions are converted into an encoded current gain sequence and input to the bottleneck position. After decoding, it is then convolved into a 3-channel 50×1 sequence, and finally expanded into a single-channel sequence of length 150.

[0028] (III) Model Training: The hybrid prediction model is trained using a training dataset. During training, current gain data is input to the feature extraction encoder, and the corresponding variable parameters are input to the parameterization physics module. The parameterization physics module learns the mapping relationship between the variable parameters and the gain features processed by the encoder, and fuses them with the encoder output at the bottleneck. The decoder then reconstructs the predicted current gain. The feature extraction encoder and the feature reconstruction decoder are connected by a skip connection structure, which is used to stitch together the high-resolution feature maps extracted by each layer of the encoder with the feature maps of the corresponding layers of the decoder, so as to preserve the detailed features of the current gain curve during reconstruction.

[0029] The parameterized physics module employs third-order polynomial regression, using all 2180 samples for training. It utilizes a smoothed L1 loss function and the Adam optimizer, then optimizes both the U-Net weights and the coefficients of the polynomial regression module simultaneously through backpropagation. The training process is as follows: Figure 6 As shown, the hybrid prediction model converges rapidly on both the training and test sets, demonstrating its excellent learning ability and generalization capacity. After training, the model of this invention can perform a single prediction in just seconds, a speed improvement of over a thousand times compared to TCAD simulations which take several hours, significantly accelerating the design iteration and parameter scanning process of the APD.

[0030] (iv) Gain Prediction and Verification: The target variable parameters of the avalanche photodiode to be predicted are input into the trained hybrid prediction model. The parameterized physics module outputs the physical prior information, which is then processed by the decoder to output a high-fidelity and physically reasonable current gain prediction result.

[0031] In this embodiment of the invention, five sets of parameter combinations that were not trained are randomly selected within the parameter range, and TCAD simulation and model prediction are performed respectively.

[0032] like Figure 7 As shown, the solid line represents the simulation result, and the dashed line represents the prediction result. The comparison results show a high degree of agreement between the predicted and simulated curves. Quantitative calculations show that the hybrid prediction model achieves a prediction determination coefficient R² of 0.984 across the entire test set, with an average relative error of only 2.5%. The single prediction time is less than 2 seconds, representing a more than 1000-fold improvement in efficiency compared to TCAD simulation. The core advantage of this method lies in its ability to successfully encapsulate the complex nonlinear mapping from nine key physical parameters to complete gain characteristics into a single, offline model. In practical applications, designers can input any set of parameters, and the model can complete the prediction within 1-2 seconds, replacing traditional numerical simulations that take hours or even days. This provides a powerful tool for the initial global design and parameter scanning of APDs. This embodiment demonstrates that, under ideal conditions with sufficient data, this invention can establish a high-performance standard prediction model, providing a stable and reliable gain prediction tool for APD design.

[0033] This invention utilizes a U-Net architecture as its core and incorporates a third-order PR hybrid prediction model to achieve rapid prediction of avalanche photodiode (APD) gain. The feature extraction encoder and feature reconstruction decoder in the hybrid prediction model are connected via a residual network, avoiding the loss of original input information and thus improving prediction accuracy. The multinomial regression model introduced in this invention transforms the "black box" functional relationship between input and output in the hybrid prediction model into an understandable form, ensuring the physical rationality of the prediction results and reducing overfitting caused by insufficient training datasets. Applying the U-Net convolutional neural network used for processing medical images to APD device performance prediction, the established prediction model significantly shortens the time required for APD device net gain prediction compared to TCAD simulation, effectively reducing R&D costs.

[0034] Example 2: The hybrid modeling method of this invention employs an incremental learning strategy to optimize the trained model. When new training data is obtained, training and fine-tuning are continued using all existing data, based on the original model parameters. This example simulates a real-world scenario where data arrives in batches during the R&D process, demonstrating the model's continuous learning capability through incremental learning-based model optimization and evolution.

[0035] (1) Incremental training process Figure 8 In this embodiment, a batch-arriving model training dataset was constructed to incrementally learn the current gain curve sets of each batch of training sets. First, the initial model M1 was trained using the first batch of 1180 samples.

[0036] Subsequently, 500 new sets of sample data were introduced in two batches (e.g. Figure 8 As shown in the figure, each time new data is introduced, the training does not start from scratch. Instead, based on the parameters of the previous stage model (M1 or M2), the training and fine-tuning are carried out using all the existing data (1680 sets in the first time and 2180 sets in the second time) to finally obtain the hybrid prediction model M3.

[0037] (2) Performance evolution comparison: Similarly, five sets of parameters were randomly selected for simulation and model prediction, and the prediction results of the hybrid prediction models M1, M2, and M3 were compared with those of the TCAD simulation. Figure 9 As shown. Performance evaluation shows that the initial model M1 has a test set R² of 0.961; after the first round of incremental learning, the R² of the hybrid prediction model M2 is... 2 Increased to 0.979; the final R-value of the mixed prediction model M3 was improved. 2 The accuracy reached 0.988; simultaneously, the total computational cost of incremental training was reduced by approximately 40% compared to training a new model from scratch. The above embodiments fully demonstrate that the hybrid modeling method provided by this invention can not only construct high-performance prediction models under ideal conditions, but also flexibly adapt to iterative R&D processes, possessing extremely high engineering practical value and broad application prospects. The method of this invention is applicable to various APD structures such as planar and tabletop types, requiring only the addition of corresponding structural parameters to the training data; furthermore, it supports incremental learning strategies, allowing the model to continuously optimize as new data arrives, eliminating the need for a complete overhaul, protecting initial investments, and adapting to iterative R&D processes.

[0038] Example 3: This invention also discloses a fast gain prediction system for avalanche photodiodes, used to implement a hybrid modeling method to improve the prediction of current gain in avalanche photodiodes. The system includes a data acquisition module, a preprocessing module, a gain prediction module, and a result output module. The data acquisition module acquires variable parameter and current gain data. The preprocessing module standardizes the variable parameter and current gain data. The gain prediction module internally stores a hybrid prediction model trained using the hybrid modeling method, which receives the processed target variable parameter and outputs the predicted gain value. The result output module displays or transmits the predicted gain value.

[0039] Example 4: The present invention also discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement a hybrid modeling method.

[0040] Example 5: The present invention also discloses a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described hybrid modeling method.

[0041] The above are preferred embodiments of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A hybrid modeling method for improving the prediction of current gain in avalanche photodiodes, characterized in that, The hybrid prediction model is based on a physically guided, improved one-dimensional U-Net neural network architecture. The hybrid modeling method includes the following steps: Step 1, Data Preparation: Obtain multiple sets of variable parameters of the avalanche photodiode and their corresponding current gain simulation data to construct a training dataset; wherein, the variable parameters include at least one of geometric parameters, process parameters, material parameters and external operating conditions; Step 2, Model Construction: Construct a hybrid prediction model based on a one-dimensional U-Net architecture. The hybrid prediction model includes a feature extraction encoder, a feature reconstruction decoder, and a parameterized physical module integrated at the bottleneck between the encoder and decoder. Step 3, Model Training: The hybrid prediction model is trained using the training dataset; wherein, during the training process, the current gain data is input to the feature extraction encoder, and the corresponding variable parameters are input to the parameterized physics module. The parameterized physics module learns the mapping relationship between the variable parameters and the gain features processed by the encoder, and merges them with the encoder output at the bottleneck, and then reconstructs the predicted current gain through the decoder. Step 4, Gain Prediction: Input the target variable parameters of the avalanche photodiode to be predicted into the trained hybrid prediction model. The parameterized physics module outputs the physical prior information, which is then processed by the decoder to output a high-fidelity and physically reasonable current gain prediction result.

2. The hybrid modeling method for improving avalanche photodiode current gain prediction according to claim 1, characterized in that, The parameterized physics module is a multinomial regression model, preferably a third-order multinomial regression model.

3. The hybrid modeling method for improving avalanche photodiode current gain prediction according to claim 1 or 2, characterized in that, The feature extraction encoder and the feature reconstruction decoder are connected by a skip connection structure, which is used to stitch together the high-resolution feature maps extracted by each layer of the encoder with the feature maps of the corresponding layers of the decoder, so as to preserve the detailed features of the current gain curve during the reconstruction process.

4. The hybrid modeling method for improving avalanche photodiode current gain prediction according to claim 1, characterized in that, The variable parameter includes at least one of the following: Geometric parameters include the thickness of each zone, the width of each zone, the inclination angle of the table sidewall, the table diameter, the width of the protective ring, and the spacing between the protective rings; Process parameters include doping concentration in each region, spatial distribution of impurities, surface roughness of the etched area, and thickness of the filling layer; Material parameters include the cadmium composition, defect density, and dielectric constant of the filler material in the HgCdTe material; External operating conditions include operating temperature, incident light angle, and incident light wavelength.

5. The hybrid modeling method for improving avalanche photodiode current gain prediction according to claim 1, characterized in that, The avalanche photodiode has a structure that is planar, mesa, guard ring, or etch-fill type; wherein, the planar avalanche photodiode structure includes planar PIN type and planar N-on-P type; and the mesa avalanche photodiode structure includes mesa PIN type and mesa N-on-P type.

6. The hybrid modeling method for improving avalanche photodiode current gain prediction according to claim 1, characterized in that, The hybrid modeling method employs an incremental learning strategy to optimize the trained model. When new training data is obtained, the model is trained and fine-tuned using all existing data, based on the original model parameters.

7. A fast gain prediction system for avalanche photodiodes, characterized in that, A hybrid modeling method for improving the prediction of avalanche photodiode current gain as described in any one of claims 1-6 includes: The data acquisition module is used to acquire the variable parameters and current gain data; The preprocessing module is used to standardize the variable parameter and current gain data; The gain prediction module internally stores a hybrid prediction model trained according to any one of the methods in claims 1-6, and is used to receive the processed target variable parameters and output the gain prediction value. The result output module is used to display or transmit the predicted gain value.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the hybrid modeling method as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the hybrid modeling method as described in any one of claims 1-6.

Images