A gain prediction method and evaluation system for a high-temperature mercury cadmium telluride avalanche photodetector

Abstract

This invention discloses a gain prediction method and evaluation system for a high-temperature mercury cadmium telluride avalanche photodetector, belonging to the field of semiconductor device modeling and performance prediction technology. The method first utilizes data transpose mapping to reconstruct the gain-voltage curve into a gain-temperature structure spanning multiple temperature zones, generating a trajectory path diagram of the gain peak changing with temperature. Based on this, peak trajectory analysis is introduced to track the gain peak drift trend, transforming physical laws into model constraint information, effectively improving the model's accuracy in identifying gain drift and nonlinear changes in the high-temperature range. Finally, a progressive data augmentation strategy is employed, dynamically supplementing subsequent temperature zone real samples based on the model's extrapolation boundary. Through iterative iteration, the effective training interval is gradually widened, achieving high-fidelity prediction of the target high-temperature range. This method achieves rapid and reliable extrapolation of high-temperature gain, providing effective support for the design optimization and high-temperature applications of mercury cadmium telluride avalanche photodetectors.

Description

Technical Field

[0001] This invention belongs to the field of semiconductor device modeling and performance prediction technology, and in particular relates to a gain prediction method and evaluation system for a high-temperature mercury cadmium telluride avalanche photodetector. Background Technology

[0002] Avalanche photodetectors (APDs), as photoelectric detection devices with extremely high internal gain, have broad application prospects in cutting-edge fields such as missile guidance, satellite remote sensing, gas monitoring, and lidar detection. As the application temperature increases, the nonlinearity of the internal physical processes of the device significantly intensifies, and the gain behavior in the high-temperature region becomes increasingly complex. Therefore, accurately obtaining the gain characteristics in the high-temperature range is of significant engineering importance and practical value in device design, performance prediction, and structural optimization.

[0003] The mainstream methods for predicting the gain of mercury cadmium telluride avalanche photodetectors are mainly divided into two categories: physical model simulation and data-driven methods. Physical model simulation relies on iterative solutions of carrier transport equations to simulate the mechanism, while data-driven methods achieve rapid prediction by establishing a mapping relationship between gain and characteristics such as bias voltage and temperature. However, existing technologies have significant drawbacks when predicting gain in the high-temperature range: (1) Complex physical mechanisms and low simulation efficiency: At high temperatures, enhanced lattice scattering leads to drastic fluctuations in the ionization coefficient, and the gain curve exhibits nonlinear drift. Traditional physical simulations are highly dependent on high-temperature material parameters that are difficult to obtain, and the calculation time is long, which cannot meet the rapid iteration requirements of device design; (2) Sparse sample data and poor extrapolation robustness: The cost of obtaining experimental data in the high-temperature range is high and the sample size is extremely small. In the absence of physical constraints, traditional data-driven models are prone to deviating from the actual physical process when predicting across temperature ranges, resulting in serious distortion of high-temperature extrapolation results; (3) Limitations of static modeling and difficulty in capturing dynamic laws: Existing methods are mostly static fitting, which makes it difficult to capture the evolution trajectory of the gain peak as it drifts with temperature. It is impossible to effectively utilize the gradual change characteristics of device performance, resulting in low data utilization and difficulty in broadening the effective prediction boundary. Summary of the Invention

[0004] Technical Problem Solved: To address the technical problems existing in the background art, this invention provides a gain prediction method and evaluation system for a high-temperature mercury cadmium telluride avalanche photodetector. Through three collaborative innovations—data transposition reconstruction, peak trajectory analysis, and progressive data enhancement—it can balance physical consistency, prediction accuracy, and extrapolation efficiency, achieving high-fidelity and highly robust intelligent gain prediction across temperature zones. Simultaneously, it provides a detector gain evaluation system adapted to this method, offering technical support for detector design optimization, high-temperature applications, and system-level dynamic monitoring.

[0005] Technical Solution: The present invention discloses a gain prediction method for a high-temperature mercury cadmium telluride avalanche photodetector. The prediction model is constructed based on a two-stream collaborative neural network architecture. The gain evaluation system and method include the following steps: Step 1, Data Transpose and Reconstruction: The original "gain-voltage" curve is mapped to a "gain-temperature" structure across temperature regions, eliminating the strong nonlinear characteristics dominated by temperature and generating a trajectory path diagram that quantitatively characterizes the gain evolution characteristics. Step 2, Peak Trajectory Analysis: Track the drift trend of the gain peak with temperature, extract the physical laws and transform them into model constraint information, and build a gray box prediction system driven by physics and data. Step 3: Progressive data augmentation: Based on the performance feedback of the model's current extrapolation boundary, dynamically supplement the real samples in the subsequent temperature range, and gradually widen the training range to the target high temperature range through iterative iteration.

[0006] Preferably, the data transpose reconstruction step further includes a preprocessing step: the original gain data is logarithmically and normally processed, and a Savitzky-Gore filter is used for smoothing and noise reduction while maintaining the shape and width of the gain peak.

[0007] Preferably, the specific steps of step 1 are as follows: the voltage curve is transposed into a temperature curve, and the physical consistency law of gain evolution under different temperature zones is used to transform the complex curve deformation into a continuous trajectory displacement, constraining the extrapolation behavior of the model in the high temperature range, and avoiding the numerical collapse caused by traditional direct extrapolation based on temperature.

[0008] Preferably, the specific execution steps of step 2 are as follows: Step 21, True Peak Identification and Search: Locate the true peak of each gain curve within the current data range through extreme value retrieval, and eliminate false peak interference at the edges by combining the confidence recognition mechanism; Step 22, Physical Prior Interval Freezing: Based on physical prior laws, low voltage interval data that do not contribute to the peak trajectory are ignored, and non-contributing points are pruned to improve the numerical stability of peak position identification. Step 23, Polynomial Trajectory Fitting: Extract the coordinates of the identified high-confidence peak points, and use the least squares method to perform multi-order polynomial fitting to establish a smooth function for the continuous evolution of peak temperature and peak gain with temperature. Step 24, Circuit Breaker and Diagnostic Evaluation: When there are too few effective samples participating in the fitting or the spatial distribution is too concentrated, the circuit breaker mechanism is automatically triggered, and a diagnostic map is exported simultaneously to judge the accuracy of peak extraction and fitting curve.

[0009] Preferably, the polynomial fitting uses a third-order polynomial, and the generated evolution function f(T) is used as a physical prior constraint and embedded into the output of the prediction model to force the numerical evolution of the model during high-temperature extrapolation.

[0010] Preferably, the gray box prediction system employs a two-stream collaborative neural network architecture, comprising: (1) Detail flow: The fluctuation characteristics of the temperature signal are extracted by using the frequency-learnable Fourier feature transform to capture the local nonlinear fluctuations in the gain evolution process. (2) Trend flow: A multi-layer fully connected network is used to learn complex physical mappings, and low-dimensional input features are directly transmitted to the fusion layer through a linear fast connection channel to prevent numerical drift of deep networks during high-temperature extrapolation. (3) Constraint embedding: At the output end after the dual-stream feature fusion, the evolution function f(T) generated by the peak trajectory analysis step is embedded as a physical boundary operator to force constraints on the peak coordinates of the predicted curve.

[0011] Preferably, it also includes an extrapolation correction step based on point displacement analysis: by calculating the gain difference between adjacent temperature points under a fixed bias voltage, the gain change law caused by temperature is explicitly introduced into the model to correct the target predicted value.

[0012] Preferably, in step 3, the initial training interval is set to 90-120K and initially extrapolated to 130K, based on R. 2 The extrapolation boundary was evaluated using MAPE and RMSLE performance indicators. After the indicators met the target, real samples were dynamically supplemented according to the temperature steps of 20-30K. Through iterative training, the extrapolation range of the model was gradually expanded from 130K to the target temperature range of 300K after multiple iterations.

[0013] The present invention also discloses a gain evaluation system for a detector based on the above prediction method, comprising: (1) Dynamic environment perception and mapping module, wherein the dynamic environment perception and mapping module is used to receive the environmental temperature sequence in real time, and use data transposition technology to map the temperature signal into a quasi-continuous input vector that the model can process, thereby realizing the quasi-continuous reconstruction of the input vector; (2) Physical trajectory calibration module, wherein the physical trajectory calibration module uses the evolution function generated by peak trajectory analysis as the physical benchmark. When the input temperature fluctuates, it automatically calculates the expected displacement of the gain peak corresponding to the target temperature point and applies it as a physical consistency constraint in the prediction process. (3) Bias compensation feedback module: The bias compensation feedback module uses the detail flow in the dual-flow cooperative neural network to capture the nonlinear fluctuations of the avalanche turn-on voltage with temperature drift in real time, and converts the predicted inflection point drift into an electrical signal to feed back to the back-end circuit to realize the automatic calibration of the detector working bias.

[0014] Preferably, the evaluation system utilizes the fast parallel computing characteristics of the linear fast connection channel and the parameterized physical bottleneck layer to achieve rapid forward inference after receiving the temperature change sequence, meeting the real-time requirements of the infrared detection system during the dynamic temperature change process; it uses a pre-trained gray box model to perform forward inference, accurately reducing the nonlinear evolution of gain caused by dark current disturbances due to thermal excitation in the 280-300K high temperature range, providing a basis for judging the gain stability under extreme thermal environments.

[0015] Compared with the prior art, the present invention has at least the following outstanding advantages: 1. Integrating white-box constraint mechanisms to ensure physical consistency of cross-temperature extrapolation: By constructing a peak trajectory constraint analysis module, semiconductor band evolution laws such as the Varshni effect are transformed into hard constraint criteria for the model, ensuring that the prediction curve is strictly constrained by the physical consistency boundary when extrapolating across temperature zones. This eliminates non-physical fluctuations and biases caused by pure data-driven approaches from the root, and solves the numerical collapse problem that traditional deep learning models are prone to when extrapolating at 300K high temperature. 2. Breaking through the time limitations of physical simulation and achieving efficient extrapolation of high-temperature characteristics: By using predictive models to replace traditional time-consuming TCAD physical simulation, the evolution trend prediction and reliable extrapolation of the gain characteristics of the detector in the high-temperature range can be achieved, which greatly accelerates the parameter scanning and structural optimization process in the early stage of detector design and provides a high-performance digital support tool for the rapid development and process tolerance analysis of mercury cadmium telluride detectors. 3. Innovative data processing strategies to solve the problem of scarce high-temperature samples: The cross-dimensional data transpose technology is used to reconstruct discrete response data into continuous global physical evolution trajectories. Combined with progressive data augmentation strategies, key samples are dynamically supplemented by extrapolation boundary feedback. This enables the model to robustly evolve the physical characteristics of the target high-temperature range using limited low-temperature data, effectively solving the problem of scarce training samples caused by the difficulty in obtaining experimental data and slow simulation convergence under high-temperature conditions. 4. Dual-stream architecture and high-fidelity filtering achieve high-fidelity restoration of local details: The constructed dual-stream collaborative architecture of detail stream and trend stream can simultaneously grasp the global temperature drift trend and capture the subtle nonlinear fluctuations caused by the electric field compression at the junction edge. Combined with the smoothing and noise reduction processing of the Savitzky-Gore filter, while filtering out numerical extrapolation glitches, it accurately preserves the inflection slope and peak shape of the gain curve's onset point, making the output result have extremely high physical fidelity. 5. Possesses excellent engineering practical value and supports system-level real-time performance evaluation: This invention can not only achieve accurate static prediction of high-temperature gain, but also be deeply integrated into the dynamic monitoring module of the infrared detection system. Under unsteady thermal environment or ambient temperature fluctuations, it can perform real-time mapping and restoration of dynamic temperature signals, providing accurate data support for automatic bias calibration and temperature compensation of back-end circuits, significantly improving the working stability and adaptability of infrared detectors in extreme environments. At the same time, the fast inference characteristics of the evaluation system meet the real-time requirements of industrial sites. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the entire process of the predictive gain method for high-fidelity extrapolation of gain characteristics under high-temperature environments according to the present invention. Figure 2 This is a cross-sectional view of a typical device structure of the mercury cadmium telluride avalanche photodetector involved in this invention. Figure 3 This is a schematic diagram illustrating the data transposition mapping and physical feature extraction principle in this invention. Figure 4 This is a graph showing the gain peak evolution effect extracted by the peak trajectory analysis module introduced in this embodiment of the invention; Figure 5 This is a schematic diagram illustrating the input feature mapping and association of the physical mechanism in the model of this invention; Figure 6 This is a schematic diagram of the detail flow extraction module based on learnable Fourier feature maps in an embodiment of the present invention; Figure 7 This is a diagram illustrating the iterative process of progressive dynamic data augmentation involved in this invention. Figure 8 This is a diagram showing the progressive extrapolation prediction effect of the cross-temperature zone gain prediction model in an embodiment of the present invention. Figure 9 This is a graph showing the performance evaluation of the model extrapolation under the progressive data augmentation strategy in this embodiment of the invention. Figure 10 The graph shows a comparison of gain extrapolation prediction before and after the introduction of the peak trajectory analysis module in this invention. Detailed Implementation

[0017] 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 1-10 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.

[0018] like Figures 1-10As shown, this invention provides a gain prediction method for a high-temperature mercury cadmium telluride avalanche photodetector. Based on a gray-box prediction model architecture that integrates physical prior driving and deep learning, it constructs a physically consistent data processing mechanism that introduces peak trajectory constraints, and coordinates cross-dimensional feature reconstruction and progressive data augmentation techniques. The core consists of three steps based on a two-stream collaborative neural network architecture: data transposition reconstruction, peak trajectory analysis, and progressive data augmentation. A prediction model is constructed based on the two-stream collaborative neural network architecture. A schematic diagram of the entire prediction process is shown below. Figure 1 As shown; the gain prediction method specifically includes the following steps: (I) Cross-dimensional data reconstruction and feature engineering: We used TCAD simulations to obtain raw gain-voltage datasets at different temperature ranges. The raw gain data underwent logarithmic and normalization preprocessing, and a Savitzky-Gore filter was used for smoothing and denoising while preserving the shape and width of the gain peaks. By employing data transpose, we extracted the translational displacement of fixed gain points on the temperature axis from the raw dataset, mapping the local response curve with voltage as the independent variable to a global physical evolution trajectory with temperature as the independent variable. This reconstructed the conventional gain curve into a cross-temperature gain-temperature curve, generating a trajectory path diagram of the gain peak changing with temperature, thus eliminating the strong nonlinear interference dominated by temperature. Based on this path diagram, we constructed physically driven input features (such as bandgap evolution and electric field gradient) and point dispersion driven input features, providing physically interpretable input support for subsequent models.

[0019] (II) Physical Constraint Analysis of Peak Trajectory: The system tracks the drift trend of the gain peak with temperature, extracts the physical laws, and transforms them into model constraint information to construct a physics-driven and data-driven gray-box prediction system. The specific execution steps are as follows: (1) True peak identification and search: Locate the true peak of each gain curve within the current data range by extreme value retrieval, and eliminate false peak interference at the edge by combining confidence identification mechanism; (2) Physical prior interval freezing: Based on the physical prior laws, low voltage interval data that do not contribute to the peak trajectory are ignored, and the non-contributing points are pruned to improve the numerical stability of peak position identification. (3) Polynomial trajectory fitting: Extract the coordinates of the identified high-confidence peak points, use the least squares method to perform multi-order polynomial fitting, and establish a smooth function for the continuous evolution of peak temperature and peak gain with temperature; the polynomial fitting adopts a third-order polynomial, and the generated evolution function f(T) is used as a physical prior constraint and embedded into the output of the prediction model to force the numerical evolution of the model when extrapolating at high temperature. (4) Circuit Breaking and Diagnostic Evaluation: When the number of effective samples participating in the fitting is too small or the spatial distribution is too concentrated, the circuit breaking mechanism is automatically triggered, and a diagnostic map is simultaneously exported to judge the accuracy of peak extraction and fitting curve. The trend line is further decoupled into explicit temperature features, including the evolution rate of the bandgap of mercury cadmium telluride material with temperature contraction obtained based on Varshni effect decoupling, and the carrier collision ionization rate coefficient that changes with temperature. The complex semiconductor physical evolution law is transformed into linear and nonlinear trend constraints that the model can understand, ensuring that the extrapolation process strictly follows the physical logic of avalanche multiplication and avoiding the numerical collapse caused by traditional direct temperature extrapolation.

[0020] (III) Construction of a dual-stream hybrid model and progressive data augmentation: Based on the model's current extrapolation boundary performance feedback, real samples for subsequent temperature ranges are dynamically added, and the training range is gradually widened to the target high-temperature range through iterative iteration. The gray-box prediction system adopts a two-stream collaborative neural network architecture, including: (1) Detail flow: The fluctuation characteristics of the temperature signal are extracted by using the frequency-learnable Fourier feature transform to capture the local nonlinear fluctuations in the gain evolution process. (2) Trend flow: A multi-layer fully connected network is used to learn complex physical mappings, and low-dimensional input features are directly transmitted to the fusion layer through a linear fast connection channel to prevent numerical drift of deep networks during high-temperature extrapolation. (3) Constraint embedding: At the output end after the dual-stream feature fusion, the evolution function f(T) generated by the peak trajectory analysis step is embedded as a physical boundary operator to force constraints on the peak coordinates of the predicted curve.

[0021] Multi-stage progressive data augmentation was conducted based on peak trajectories: the initial training interval was set at 90-120K, and extrapolated to 130K initially, according to R... 2 The extrapolation boundary is evaluated by the performance indicators MAPE and RMSLE. After the indicators meet the standards, real samples are dynamically supplemented according to the temperature steps of 20-30K. Through iterative training, the TCAD simulation sample supplementation mechanism is dynamically triggered by monitoring whether the prediction residual exceeds the preset threshold. After multiple iterations, the extrapolation range of the model is gradually expanded from 130K to the target temperature range of 300K, so that the model can reliably evolve around the physical trajectory in the high temperature range where samples are scarce.

[0022] (iv) Collaborative extrapolation prediction and high-fidelity output: A tensor containing peak trajectory constraints, explicit physical features, and dual physical constants is fed into a physically consistent hybrid prediction model. By fusing global evolution trends and local physical details through dual-stream coordinate mapping, the gain curve of the target high-temperature region (e.g., 300K) is extrapolated. Simultaneously, extrapolation correction is performed through point displacement analysis: the gain difference between adjacent temperature points under a fixed bias voltage is calculated, and the temperature-induced gain variation is explicitly introduced into the model to correct the target prediction value. The final output sequence is smoothed and denoised using a Savitzky-Gore filter. This filter has a convolution window length of 5 to 15 sampling points and a fitting order of 2 to 4. While filtering out numerical extrapolation glitches, it retains the inflection point slope and peak shape characteristics of the gain curve without loss, outputting high-fidelity, physically consistent high-temperature gain prediction results.

[0023] This invention also discloses a detector gain evaluation system based on the above prediction method, including a dynamic environment perception and mapping module, a physical trajectory calibration module, and a bias compensation feedback module. These modules work together to achieve real-time monitoring and dynamic calibration of the detector performance. (1) The dynamic environment perception and mapping module receives the ambient temperature sequence of the infrared detection system in real time during actual operation. It uses the aforementioned data transposition technology to map the temperature signal into a quasi-continuous input vector that the model can process, thereby realizing the quasi-continuous reconstruction of the input vector and supporting the performance prediction of non-sampling point temperature.

[0024] (2) The physical trajectory calibration module uses the peak trajectory function f(T) as the physical reference. When the input temperature fluctuates, it automatically calculates the expected displacement of the gain peak corresponding to the target temperature point and applies it as a physical consistency constraint in the prediction process to ensure that the predicted gain curve cluster remains highly consistent in physical logic when the ambient temperature fluctuates continuously.

[0025] (3) The bias compensation feedback module uses the detail flow in the dual-flow collaborative neural network to capture the nonlinear fluctuations of the avalanche turn-on voltage with temperature drift in real time, and converts the predicted inflection point drift into an electrical signal to feed back to the back-end circuit to realize the automatic calibration of the detector working bias voltage.

[0026] The gain evaluation system of this invention utilizes the fast parallel computing characteristics of linear fast connection channels and parameterized physical bottleneck layers to achieve rapid forward inference after receiving the temperature change sequence, reducing numerical redundancy calculations in deep networks and meeting the real-time requirements of infrared detection systems during dynamic temperature changes. By using a pre-trained gray-box model to perform forward inference, it can accurately reduce the nonlinear evolution of gain caused by dark current disturbances due to thermal excitation within the 280-300K high-temperature range, providing a basis for judging the gain stability of infrared detection systems under extreme thermal environments.

[0027] Example 1: This example uses a planar N-on-P structure mercury cadmium telluride (HgCdTe) avalanche photodetector (APD) applied in the mid-infrared band as the research object. Its structure is as follows: Figure 2 As shown; the device is grown on a CdZnTe substrate using molecular beam epitaxy, and p is grown on the substrate. + A type-n absorption layer is formed, upon which an n-type HgCdTe multiplication layer is constructed, and localized n-type layers are formed through ion implantation. + Type injection region, with p + The PN junction formed by the absorption layer is the source of photocurrent multiplication. The unique junction edge curvature effect of planar devices causes the electric field distribution to concentrate locally at the injection window edge. This non-uniform field strength is the physical root cause of the complex gain trajectory at high temperatures. By covering the device surface with a CdTe passivation layer and connecting it with metal electrodes, lossless extraction of bias injection and multiplication signal is achieved. The original gain-voltage response curves in the 80-280K temperature range are obtained through Sentaurus TCAD simulation.

[0028] (I) Cross-dimensional data reconstruction and feature engineering: Using the Sentaurus TCAD semiconductor physics simulation tool, the Poisson equation and continuity equation were solved through the SDEVICE module to simulate and obtain the raw gain-voltage (GV) response curves covering a temperature range of 80-280K. In this dimension, the gain curves at different temperatures intertwine and exhibit strong nonlinearity, and the physical correlation between temperature points is sparse in the data arrangement, making it difficult for deep learning models to directly capture the subtle details of the continuous temperature evolution. First, the raw data is preprocessed by logarithmic transformation and normalization, and noise is suppressed using Savitzky-Gore filters (window length 5-15, order 2-4). Second, this invention creatively extracts the translational displacement of each gain point on the temperature axis through cross-dimensional data transposition technology, reconstructing the conventional lateral bias response curve into a vertical cross-temperature-range gain-temperature (GT) curve (e.g., ...). Figure 3As shown in the diagram, after transposition, the originally discrete temperature sampling points are integrated into a continuous physical evolution trajectory, exhibiting obvious common fluctuation characteristics. This dimensionality transformation significantly reduces the difficulty of extracting high-dimensional nonlinear features, thus ensuring the model's high-precision stability in full-temperature range prediction. To further enhance the model's generalization ability, this invention introduces a multi-dimensional feature construction step based on the transposed data. By performing nonlinear combination and feature scaling on the original physical quantities and executing global normalization, physical parameters of different dimensions are mapped to a unified numerical range. This process not only eliminates model training bias caused by differences in voltage and temperature magnitudes but also enhances the sparse representation capability of the feature space through physical law correlation, laying a standardized data foundation for the efficient execution of Fourier feature mapping in the subsequent detail flow module, thereby ensuring the model's prediction stability across the entire temperature range. Through this coordinate space transformation, a gain peak trajectory path map that characterizes the evolution law of device performance is generated. Based on this path diagram, this embodiment further constructs physical driving input features (bandgap evolution, electric field gradient) and point dispersion driving input features to characterize the non-uniform evolution of the junction edge field strength with temperature rise, providing high-dimensional input support with clear physical interpretability for subsequent models.

[0029] (II) Physical Constraint Analysis of Peak Trajectory: As the core physical constraint construction stage of this invention, its essence is to set "physical fences" for the black-box logic of the neural network through white-box constraint analysis. This embodiment first retrieves the global trajectory path generated by data transposition, and then uses a third-order polynomial regression algorithm to perform nonlinear fitting on the discrete gain peak points, generating a continuous trend line of gain peak drift with temperature (e.g., ...). Figure 4 As shown in the figure, it not only mathematically realizes the continuous representation of discrete observations, but also establishes a reference coordinate system for cross-temperature evolution at the physical level.

[0030] To further enhance the physical consistency of the model, this embodiment constructs a physical-driven feature mapping process (such as...). Figure 5(As shown). First, the Varshni effect is used to analyze the underlying trend curves of the band gap (Eg) of cadmium telluride (HMt) material as a function of temperature (T) and the ionization coefficients (α, β) as a function of electric field. A white-box modeling feature mapper decouples the trend lines into explicit temperature features such as band gap and carrier collisional ionization rate. This mechanism transforms the complex evolution of semiconductor physics into linear and nonlinear trend constraint operators understandable by neural networks. In the input α feature construction region, the decoupled physical features (cadmium composition x, temperature T, band gap Eg, ionization coefficient, etc.) are structured to construct a two-dimensional master feature input vector based on the optimal physical structure, which is then injected into the model architecture. This ensures that when the model performs extrapolation to the 300K high-temperature range, the logic of the output result moving along the bias axis strictly follows the physical nature of avalanche multiplication, preventing numerical collapse and logical divergence that are prone to occur in purely data-driven models during high-temperature extrapolation from the underlying logic.

[0031] (III) Construction and Progressive Enhancement of the Dual-Stream Hybrid Model: This embodiment constructs a two-stream collaborative neural network that includes detail flow and trend flow (e.g., Figure 6 (As shown); The detail flow utilizes frequency-learnable Fourier feature extraction technology, and the detail flow extraction module structure of the Fourier feature mapping is as follows: The system first receives a time-domain gain curve containing complex fluctuations as input and introduces it into the core Fourier feature mapping module. In this module, the input signal undergoes nonlinear transformation via sine and cosine functions, while introducing learnable frequency and amplitude parameters, enabling the model to adaptively extract key frequency components based on data characteristics. Subsequently, the transformed data undergoes feature encoding, transforming the original single-point temperature data into a multi-dimensional frequency and amplitude feature vector. Combined with a hypercube-gated multilayer perceptron, this aims to accurately capture local nonlinear fluctuations and microscopic detail features of the gain near the turn-on voltage. The trend flow feature network deeply integrates the trend information extracted from the peak trajectory, ensuring that the weight of the global physical direction is not diluted by the deep nonlinear mapping through linear fast connection channels, and embedding the evolution function f(T) at the dual-stream fusion output as a physical boundary constraint.

[0032] Addressing the pain points of scarce samples and extremely difficult simulation convergence in the 300K target high-temperature range, this embodiment employs a multi-stage progressive data augmentation method based on the peak trajectory: using 30K as the temperature difference step, it progresses from the initial 90K training range towards higher temperatures (e.g., ...). Figure 7 (As shown). By Figure 8 It can be seen that the progressive enhancement of the extrapolation at each step shows a high degree of agreement with the simulation results. After the extrapolation at each stage is completed, the results are calculated based on the coefficient of determination (R²). 2 The performance indicators such as mean absolute percentage error (MAPE) and root mean square logarithmic error (RMSLE) are used to dynamically evaluate the extrapolation boundary (performance indicators such as...). Figure 9As shown), the coefficient of determination R, which measures the goodness of fit of the model, is... 2 The model maintains an accuracy above 0.999 across most temperature ranges, particularly approaching 1.0000 in the core temperature range of 180-220K, indicating its strong explanatory power for the evolution of physical gain characteristics. Meanwhile, the mean absolute percentage error (MAPE) and root mean square logarithmic error (RMSLE), reflecting prediction bias, show a significant decreasing trend with the deepening of the physical enhancement process, eventually converging to extremely low levels. Specifically, the RMSLE approaches 0 in the mid-to-high temperature range. 2 The synergistic effect of the curve running smoothly at a high level and the error curves (MAPE, RMSLE) converging stably at a low level fully verifies the significant technical effect of the physical enhancement mechanism in overcoming nonlinear fluctuations and improving global prediction consistency. If the residual index exceeds the preset threshold, TCAD simulation is automatically triggered to supplement samples and perform model feedback retraining. After multiple iterations, the extrapolation interval is widened to 300K (e.g., Figure 8 As shown in the figure, the model can still reliably evolve around the physical trajectory even in the high-temperature range where the sample is missing.

[0033] (iv) Collaborative extrapolation prediction and high-fidelity output: By fusing global evolution trends and local physical details through dual-stream coordinate mapping, high-precision extrapolation of the gain curve in the 300K high-temperature region is achieved. Simultaneously, the gain difference between adjacent temperature points under a fixed bias is calculated to correct the predicted values. The output gain sequence is fed into a Savitzky-Gore filter (convolution window length 10 sampling points, fitting order 3) for smoothing and denoising, eliminating extrapolation fluctuations while preserving the inflection point slope and peak shape. The prediction results before and after introducing the peak trajectory are compared (e.g.,...). Figure 10 As shown in the figure, the final output high-temperature gain curve is highly consistent with the full physical simulation results, and the single prediction time is greatly shortened, realizing high-fidelity and high-efficiency high-temperature gain prediction.

[0034] Example 2: This example is based on the hybrid prediction model trained in Example 1, and evaluates the performance of the planar N-on-P structure mercury cadmium telluride avalanche photodetector in real time under unsteady thermal conditions. The material and process parameters of the device are the same as in Example 1.

[0035] (1) Dynamic Environment Perception and Quasi-Continuous Input Mapping: Simulating the continuous temperature variation condition of 200-300K in the actual operation of the infrared detection system, the local response law of the discrete voltage independent variable is mapped to the quasi-continuous input vector of the temperature independent variable using the numerical transpose technique of Example 1. Although the original simulation data is discrete, through... Figure 1 The prediction architecture shown allows the model to receive temperature inputs of arbitrary precision and utilize encapsulated high-performance physical feature vectors to achieve performance prediction of non-sampling point temperatures.

[0036] (2) Dynamic stability calibration based on physical trajectory f(T): The physical trajectory calibration module inside the model uses the peak trajectory evolution function f(T) as the standard. When the input temperature deviates slightly, the physical trajectory calibration module automatically calculates the expected displacement of the gain peak corresponding to the temperature point and applies it as a physical constraint to the prediction process to ensure that the predicted gain curve cluster is physically logically consistent when the temperature fluctuates continuously, thus avoiding numerical random noise and non-physical fluctuations.

[0037] (3) Real-time capture and feedback of the gain inflection point using a dual-flow architecture: Addressing the dark current disturbance and nonlinear gain evolution caused by thermal excitation in the high-temperature range, the detail flow of the dual-flow architecture is utilized. A Fourier feature extractor is used to capture the minute influence of temperature changes on the slope of the gain curve's inflection point in real time. Even in the high-temperature range above 290K, the model has already passed the feedback mechanism... Figure 6 The progressive data augmentation process shown captures the physical characteristics of high temperatures. Its output gain sequence can still accurately reproduce the nonlinear evolution of gain caused by thermally excited dark current disturbances. It converts the predicted inflection point drift into an electrical signal and feeds it back to the back-end circuit to achieve automatic closed-loop calibration of the detector bias.

[0038] (4) Validation of prediction efficiency and high-fidelity output: The variable temperature sequence is input into the evaluation system, and the output sequence is processed by the Savitzky-Gore filter to obtain a high-fidelity gain sequence. Since the model transforms the complex physical simulation into efficient tensor operations, the calculation time of a single performance evaluation of the system is maintained at the second level, which meets the real-time requirements. Experimental verification shows that the trend of the predicted gain peak with temperature is highly consistent with the verification data of long-term TCAD simulation, which can provide an accurate basis for judging the gain stability of the infrared detection system under extreme thermal environment.

[0039] 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 gain prediction method for a high-temperature mercury cadmium telluride avalanche photodetector, characterized in that, The gain evaluation system and method, which constructs a prediction model based on a two-stream collaborative neural network architecture, include the following steps: Step 1, Data transpose and reconstruction: The original "gain-voltage" curve is mapped to a "gain-temperature" structure across the temperature range, eliminating the strong nonlinear characteristics dominated by temperature and generating a trajectory path diagram that quantitatively characterizes the gain evolution characteristics. Step 2, Peak Trajectory Analysis: Track the drift trend of the gain peak with temperature, extract the physical laws and transform them into model constraint information, and build a gray box prediction system driven by physics and data. Step 3: Progressive data augmentation: Based on the performance feedback of the model's current extrapolation boundary, dynamically supplement the real samples in the subsequent temperature range, and gradually widen the training range to the target high temperature range through iterative iteration.

2. The gain prediction method for a high-temperature mercury cadmium telluride avalanche photodetector according to claim 1, characterized in that, The data transpose reconstruction step is preceded by a preprocessing step: the original gain data is logarithmically and normally processed, and a Savitzky-Gore filter is used for smoothing and noise reduction while maintaining the shape and width of the gain peak.

3. The gain prediction method for a high-temperature mercury cadmium telluride avalanche photodetector according to claim 1, characterized in that, The specific steps of step 1 are as follows: transpose the voltage curve into a temperature curve, utilize the physical consistency law of gain evolution under different temperature zones, transform the complex curve deformation into a continuous trajectory displacement, constrain the extrapolation behavior of the model in the high temperature range, and avoid the numerical collapse caused by traditional direct extrapolation based on temperature.

4. The gain prediction method for a high-temperature mercury cadmium telluride avalanche photodetector according to claim 1, characterized in that, The specific steps for step 2 are as follows: Step 21, True Peak Identification and Search: Locate the true peak of each gain curve within the current data range through extreme value retrieval, and eliminate false peak interference at the edges by combining confidence recognition mechanism; Step 22, Physical Prior Interval Freezing: Based on physical prior laws, low voltage interval data that do not contribute to the peak trajectory are ignored, and non-contributing points are pruned to improve the numerical stability of peak position identification. Step 23, Polynomial Trajectory Fitting: Extract the coordinates of the identified high-confidence peak points, and use the least squares method to perform multi-order polynomial fitting to establish a smooth function for the continuous evolution of peak temperature and peak gain with temperature. Step 24, Circuit Breaker and Diagnostic Evaluation: When there are too few effective samples participating in the fitting or the spatial distribution is too concentrated, the circuit breaker mechanism is automatically triggered, and a diagnostic map is exported simultaneously to judge the accuracy of peak extraction and fitting curve.

5. The gain prediction method for a high-temperature mercury cadmium telluride avalanche photodetector according to claim 4, characterized in that, The polynomial fitting uses a third-order polynomial, and the generated evolution function f(T) is used as a physical prior constraint and embedded into the output of the prediction model to force the numerical evolution of the model under high-temperature extrapolation.

6. The gain prediction method for a high-temperature mercury cadmium telluride avalanche photodetector according to claim 5, characterized in that, The gray box prediction system employs a dual-stream collaborative neural network architecture, including: (1) Detail flow: The fluctuation characteristics of the temperature signal are extracted by using the frequency-learnable Fourier feature transform to capture the local nonlinear fluctuations in the gain evolution process. (2) Trend flow: A multi-layer fully connected network is used to learn complex physical mappings, and low-dimensional input features are directly transmitted to the fusion layer through a linear fast connection channel to prevent numerical drift of deep networks during high-temperature extrapolation. (3) Constraint embedding: At the output end after the dual-stream feature fusion, the evolution function f(T) generated by the peak trajectory analysis step is embedded as a physical boundary operator to force constraints on the peak coordinates of the predicted curve.

7. The gain prediction method for a high-temperature mercury cadmium telluride avalanche photodetector according to claim 6, characterized in that, It also includes an extrapolation correction step based on point displacement analysis: by calculating the gain difference between adjacent temperature points under a fixed bias voltage, the gain change law caused by temperature is explicitly introduced into the model to correct the target prediction value.

8. The gain prediction method for a high-temperature mercury cadmium telluride avalanche photodetector according to claim 1, characterized in that, The execution logic of the stepped enhancement layer in step 3 is as follows: The initial training interval is set to 90-120K, and then initially extrapolated to 130K, based on R... 2 The extrapolation boundary was evaluated using MAPE and RMSLE performance indicators. After the indicators met the target, real samples were dynamically supplemented according to the temperature steps of 20-30K. Through iterative training, the extrapolation range of the model was gradually expanded from 130K to the target temperature range of 300K after multiple iterations.

9. A gain evaluation system for a detector based on the prediction method according to any one of claims 1-8, characterized in that, include: (1) Dynamic environment perception and mapping module, wherein the dynamic environment perception and mapping module is used to receive the environmental temperature sequence in real time, and use data transposition technology to map the temperature signal into a quasi-continuous input vector that the model can process, thereby realizing the quasi-continuous reconstruction of the input vector; (2) Physical trajectory calibration module, wherein the physical trajectory calibration module uses the evolution function generated by peak trajectory analysis as the physical benchmark. When the input temperature fluctuates, it automatically calculates the expected displacement of the gain peak corresponding to the target temperature point and applies it as a physical consistency constraint in the prediction process. (3) Bias compensation feedback module: The bias compensation feedback module uses the detail flow in the dual-flow cooperative neural network to capture the nonlinear fluctuations of the avalanche turn-on voltage with temperature drift in real time, and converts the predicted inflection point drift into an electrical signal to feed back to the back-end circuit to realize the automatic calibration of the detector working bias.

10. The detector gain evaluation system according to claim 9, characterized in that, The evaluation system utilizes the fast parallel computing characteristics of linear fast connection channels and parameterized physical bottleneck layers to achieve rapid forward inference after receiving the temperature change sequence, meeting the real-time requirements of the infrared detection system during dynamic temperature changes. It uses a pre-trained gray box model to perform forward inference, accurately reducing the nonlinear evolution of gain caused by dark current disturbances due to thermal excitation in the 280-300K high temperature range, providing a basis for judging the gain stability under extreme thermal environments.

Images