Intelligent identification optimization method and system for interface defect characteristics of third-generation semiconductor material

Abstract

The invention provides a third-generation semiconductor material interface defect feature intelligent identification optimization method and system, and relates to the technical field of semiconductor defect identification, and the method comprises the steps: collecting a semiconductor material cross section image and electrical data; constructing a morphology and stress distribution feature extraction channel, extracting interface defect multi-dimensional features, inputting the extracted interface defect multi-dimensional features into a heterogeneous graph attention network, identifying defect types, density and distribution, constructing a performance loss evaluation model, determining an optimal migration path by adopting a layered progressive defect regulation and control strategy in combination with a temperature gradient auxiliary mechanism, and optimizing defect regulation and control parameters in real time. Accurate identification and intelligent regulation and control of interface defects are realized, and the electrical performance of the semiconductor material is improved.

Description

Technical Field

[0001] The present invention relates to the field of semiconductor defect identification technology, and in particular to a method and system for intelligently identifying and optimizing interface defect characteristics of third-generation semiconductor materials. Background Art

[0002] Third-generation semiconductor materials have excellent properties such as wide bandgap, high breakdown field strength, and high thermal conductivity, showing great application prospects in the fields of high-frequency, high-temperature, and high-power electronic devices. However, interface defects are inevitably generated during the preparation process, which significantly affect the electrical properties of the materials and device performance. Traditional methods for characterizing interface defects rely primarily on techniques such as electron microscopy, X-ray diffraction, and spectral analysis. Defect identification and characterization are primarily based on manual experience and judgment, lacking systematic quantitative analysis methods. This makes it difficult to extract high-precision three-dimensional defect morphology and stress field distribution information, and to quantitatively assess the impact of different types of defects on material performance. This restricts the precise implementation of defect control strategies. Furthermore, during the control process, control parameters cannot be optimized in a timely manner based on changes in material properties, resulting in low defect control efficiency and difficulty in optimizing material performance. Therefore, a solution is urgently needed to solve the problems existing in the prior art. Summary of the Invention

[0003] The embodiments of the present invention provide a method and system for intelligently identifying and optimizing interface defect characteristics of third-generation semiconductor materials, which can at least solve some of the problems existing in the prior art.

[0004] A first aspect of an embodiment of the present invention provides a method for intelligently identifying and optimizing interface defect characteristics of third-generation semiconductor materials, comprising: Collecting a cross-sectional image of the third-generation semiconductor material to be tested, and obtaining electron mobility distribution data and carrier concentration data corresponding to the cross-sectional image; A high-resolution morphology feature extraction channel and a local stress distribution feature extraction channel are respectively constructed based on the cross-sectional image. The electron mobility distribution data and carrier concentration data are used as auxiliary parameters to extract the three-dimensional morphology information and stress field distribution information of the interface defect and perform a multi-scale feature joint analysis. A multi-dimensional feature vector of the interface defect is generated and input into a pre-trained heterogeneous graph attention network. The type, density, and distribution information of the interface defect are output through an adaptive weight allocation mechanism. Based on the type, density and distribution information of the interface defects, a material performance loss evaluation model is constructed to calculate the influence of material performance on the interface defects and determine the loss weight coefficient; A hierarchical and progressive defect control strategy is adopted. The defect diffusion kinetic equation is established based on the material lattice structure. The barrier height and position parameters of the multiple defect barriers are determined in combination with the loss weight coefficient. A temperature gradient auxiliary control mechanism is introduced to form a controllable temperature gradient field to obtain the optimal migration path. Based on the optimal migration path, the changes in the material electrical properties during the defect control process are monitored in real time, the defect control parameters corresponding to the optimal migration path are optimized and performance verified in real time, and the performance improvement indicators and control parameters are stored in the optimization parameter database.

[0005] In an optional embodiment, Collecting a cross-sectional image of the third-generation semiconductor material to be tested and obtaining electron mobility distribution data and carrier concentration data corresponding to the cross-sectional image includes: Performing cross-sectional scanning imaging of the third-generation semiconductor material using a transmission electron microscope, focusing the electron beam to the cross-sectional position of the third-generation semiconductor material, collecting interaction signals between the electron beam and the third-generation semiconductor material, and generating a cross-sectional image of the third-generation semiconductor material; Four detection electrodes are set on the surface of the third-generation semiconductor material. By applying a scanning voltage between adjacent detection electrodes, Hall voltage data is collected under the action of a perpendicular magnetic field and the electron mobility of each area of ​​the third-generation semiconductor material is calculated to obtain electron mobility distribution data; The current data between the four detection electrodes is measured, and the resistivity of the surface of the third-generation semiconductor material is calculated. Combined with the electron mobility distribution data, the carrier concentration data is calculated.

[0006] In an optional embodiment, A high-resolution morphology feature extraction channel and a local stress distribution feature extraction channel are respectively constructed based on the cross-sectional image. The electron mobility distribution data and carrier concentration data are used as auxiliary parameters to extract the three-dimensional morphology information and stress field distribution information of the interface defect and perform multi-scale feature joint analysis to generate a multi-dimensional feature vector of the interface defect. Perform multi-scale morphological feature enhancement on cross-sectional images of third-generation semiconductor materials. A morphological feature enhancement function is constructed based on a Gaussian kernel function and a Laplace operator. The enhanced cross-sectional image is then generated and its boundary extracted using an adaptive threshold segmentation algorithm, where the adaptive threshold is dynamically calculated based on the statistical characteristics of pixel values ​​within a local window. A high-resolution morphology feature extraction channel and a local stress distribution feature extraction channel are respectively constructed according to the cross-sectional image, the boundary extraction result is used as input data, and the electron mobility distribution data and the carrier concentration data are input as auxiliary parameters into the high-resolution morphology feature extraction channel and the local stress distribution feature extraction channel to extract three-dimensional morphology information of the interface defect and stress field distribution information of the interface defect; performing a multi-scale feature joint analysis on the three-dimensional morphology information and the stress field distribution information, performing weighted fusion on the three-dimensional morphology information and the stress field distribution information using a feature fusion function, and dynamically adjusting feature weights based on an adaptive weight optimization equation; The fused features are input into the feature space mapping function for feature transformation, and feature extraction is performed through a nonlinear feature extraction operator to generate a multidimensional feature vector of the interface defect. The multidimensional feature vector contains three-dimensional morphology information and stress field distribution information of the interface defect.

[0007] In an optional embodiment, The input is fed into the pre-trained heterogeneous graph attention network, which outputs the type, density, and distribution information of interface defects through an adaptive weight distribution mechanism, including: The multidimensional feature vector is input into the pre-trained heterogeneous graph attention network. The neighborhood nodes within the preset radius of the multidimensional feature vector are weighted and summed to obtain the local density feature. The local density feature is averaged and weighted with the local density feature variance to obtain the global density feature. The local density feature and the global density feature are concatenated with the multidimensional feature vector and subjected to linear transformation and activation function to obtain the dynamic adjacency matrix. Calculate the density gradient of the local density feature under different feature extraction radii, obtain the scale weight coefficient by linear transformation of the density gradient, and add the product of the density gradient and the multidimensional feature vector weighted by the scale weight coefficient to the residual feature of the multidimensional feature vector to obtain the density flow feature; Based on the dynamic adjacency matrix and density flow features, the multidimensional feature vector is updated through the pre-trained heterogeneous graph attention network to obtain the updated feature vector. The updated feature vector is linearly transformed to obtain the current density distribution. The KL divergence between the current density distribution and the density distribution at the previous moment and the Euclidean distance between the current density distribution and the target density distribution are calculated to obtain the evolution constraint. The updated feature vector, density flow feature, local density feature and global density feature are spliced ​​and input into a multilayer perceptron to obtain the feature importance score. The adaptive weight coefficient is calculated based on the feature importance score. The product of the adaptive weight coefficient and the updated feature vector is input into the pre-trained heterogeneous graph attention network. The type, density and distribution information of the interface defects are output in combination with the evolutionary constraints.

[0008] In an optional embodiment, Based on the type, density and distribution information of the interface defects, a material performance loss evaluation model is constructed to calculate the influence of material performance on interface defects and determine the loss weight coefficient, including: The type, density, and distribution characteristics of interface defects are input into a loss assessment model based on a multi-layer neural network, and the defect characteristics are mapped to the performance loss space through nonlinear transformation. Residual connections are used to enhance the output features of the loss assessment model, and the influence of interface defects on material properties is calculated in combination with pre-set material performance evaluation indicators. The influence of interface defects is weighted and normalized according to the importance distribution of material performance indicators to obtain the loss weight coefficient.

[0009] In an optional embodiment, A hierarchical and progressive defect control strategy is adopted. Based on the material lattice structure, the defect diffusion kinetic equation is established. The barrier height and position parameters of the multiple defect barriers are determined in combination with the loss weight coefficient. A temperature gradient auxiliary control mechanism is introduced to form a controllable temperature gradient field. The optimal migration path is obtained, including: Collect the defect type and defect distribution information of the material lattice structure, determine the diffusion coefficient matrix according to the defect type, construct the defect source term function based on the defect distribution information, and substitute the diffusion coefficient matrix and the defect source term function into the diffusion kinetics equation to calculate the defect concentration distribution; The strain tensor is obtained by calculating the lattice distortion energy through the elastic constant tensor based on the defect concentration distribution, and the defect formation energy is corrected according to the strain tensor and the local density of defects to obtain the corrected defect formation energy; The modified defect formation energy is input into the Gaussian distribution function and combined with the loss weight coefficient to calculate the barrier height of the multiple defect barriers. The gradient of the strain tensor and the gradient of the local defect density are used to determine the barrier position parameters. The interaction energy of the multiple defect barriers is calculated based on the coupling coefficient between the barriers to obtain the total barrier energy distribution. A temperature gradient-assisted control mechanism is constructed based on the total barrier energy distribution. The reference temperature, the spatial distribution function of the total barrier energy, and the time modulation function are input into the pre-set heat diffusion equation to obtain the evolution characteristics of the controllable temperature gradient field. A transition probability function is constructed based on the evolution characteristics of the total potential barrier energy distribution and the controllable temperature gradient field. The transition probability function is solved under the dual constraints of the temperature gradient restriction condition and the potential barrier energy threshold, and the optimal migration path is obtained by combining the energy gradient iterative calculation.

[0010] In an optional embodiment, Based on the evolution characteristics of the total barrier energy distribution and the controllable temperature gradient field, a transition probability function is constructed. The transition probability function is solved under the dual constraints of the temperature gradient restriction condition and the barrier energy threshold. The optimal migration path obtained by combining the energy gradient iterative calculation includes: A transition probability function is constructed based on the total barrier energy distribution and the evolution characteristics of the controllable temperature gradient field. The maximum temperature gradient in the controllable temperature gradient field is extracted as the equivalent temperature parameter. A set of candidate positions is generated within a preset radius of the current path position. The total barrier energy of each candidate position in the candidate position set is calculated based on the total barrier energy distribution. Substitute the difference between the total barrier energy of each candidate position and the total barrier energy of the current path position and the Laplace operator of the total barrier energy distribution into the transition probability function to obtain the transition probability, select the candidate position with a transition probability greater than a preset threshold as the transition position, and update to the transition position with the maximum transition probability if a transition position exists. If no transition position exists, update along the negative gradient direction of the total barrier energy; Under the dual constraints of temperature gradient restriction and potential barrier energy threshold, the transition position is repeatedly solved until convergence to obtain the optimal migration path.

[0011] In an optional embodiment, Real-time monitoring of material electrical property changes during defect control based on the optimal migration path, real-time optimization and performance verification of defect control parameters corresponding to the optimal migration path, and storing performance improvement indicators and control parameters in an optimization parameter database include: Real-time monitoring of material electrical property changes during defect control based on the optimal migration path. This includes collecting real-time data on material conductivity, carrier mobility, and carrier concentration distribution, and establishing a quantitative mapping relationship between defect density, defect type, and electrical property changes. Real-time optimization and performance verification of defect control parameters corresponding to the optimal migration path are performed. The performance loss gradient is calculated based on the electrical characteristic change data. The migration rate, barrier height, and temperature gradient in the defect control parameters are optimized and adjusted online according to the performance loss gradient. The electrical performance improvement index before and after the control is calculated. The performance improvement indicators and control parameters are stored in the optimization parameter database and classified according to the corresponding relationship between defect type, defect density and electrical characteristics.

[0012] A second aspect of an embodiment of the present invention provides a system for intelligently identifying and optimizing interface defect characteristics of third-generation semiconductor materials, comprising: The first unit is used to collect a cross-sectional image of the third-generation semiconductor material to be tested, and obtain electron mobility distribution data and carrier concentration data corresponding to the cross-sectional image; The second unit is used to construct a high-resolution morphology feature extraction channel and a local stress distribution feature extraction channel based on the cross-sectional image, respectively, and use the electron mobility distribution data and carrier concentration data as auxiliary parameters to extract the three-dimensional morphology information and stress field distribution information of the interface defect and perform multi-scale feature joint analysis, generate a multi-dimensional feature vector of the interface defect, and input it into a pre-trained heterogeneous graph attention network, and output the type, density and distribution information of the interface defect through an adaptive weight allocation mechanism; The third unit is used to construct a material performance loss evaluation model based on the type, density and distribution information of the interface defects, calculate the influence of material performance on interface defects and determine the loss weight coefficient; The fourth unit is used to adopt a hierarchical progressive defect control strategy, establish a defect diffusion kinetic equation based on the material lattice structure, determine the barrier height and position parameters of the multiple defect barriers in combination with the loss weight coefficient, introduce a temperature gradient auxiliary control mechanism to form a controllable temperature gradient field, and obtain the optimal migration path; The fifth unit is used to monitor the changes in the electrical properties of the material during the defect control process in real time based on the optimal migration path, perform real-time optimization and performance verification on the defect control parameters corresponding to the optimal migration path, and store the performance improvement indicators and control parameters in the optimization parameter database.

[0013] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including: A processor and a memory for storing processor-executable instructions, wherein the processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0014] In the present invention, high-resolution image analysis is combined with electronic property data to construct a comprehensive defect characterization system, which can accurately capture key information such as the morphology and stress distribution of microscale interface defects. The adaptive weight distribution mechanism is used to achieve accurate identification and classification of different types of defects, which solves the limitations of traditional methods in the identification of complex interface defects. The material performance loss assessment model can quantify the degree of influence of defects on material properties, providing a scientific basis for subsequent defect control. The layered progressive defect control strategy combined with the temperature gradient assisted control mechanism realizes effective control of interface defects, greatly improving the electrical properties and stability of third-generation semiconductor materials. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a flow chart of a method for intelligently identifying and optimizing interface defect characteristics of third-generation semiconductor materials according to an embodiment of the present invention; Figure 2 This is a comparison chart of the interface defect density distribution prediction accuracy of the third-generation semiconductor material interface defect feature intelligent identification optimization method according to an embodiment of the present invention; Figure 3 This is a flow chart for constructing an optimal migration path based on total barrier energy distribution and controllable temperature gradient field according to an embodiment of the present invention. DETAILED DESCRIPTION

[0016] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0017] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0018] Figure 1 FIG. 1 is a flow chart of an intelligent identification and optimization method for interface defect characteristics of third-generation semiconductor materials according to an embodiment of the present invention. Figure 1 As shown, the method includes: Collecting a cross-sectional image of the third-generation semiconductor material to be tested, and obtaining electron mobility distribution data and carrier concentration data corresponding to the cross-sectional image; A high-resolution morphology feature extraction channel and a local stress distribution feature extraction channel are respectively constructed based on the cross-sectional image. The electron mobility distribution data and carrier concentration data are used as auxiliary parameters to extract the three-dimensional morphology information and stress field distribution information of the interface defect and perform a multi-scale feature joint analysis. A multi-dimensional feature vector of the interface defect is generated and input into a pre-trained heterogeneous graph attention network. The type, density, and distribution information of the interface defect are output through an adaptive weight allocation mechanism. Based on the type, density and distribution information of the interface defects, a material performance loss evaluation model is constructed to calculate the influence of material performance on the interface defects and determine the loss weight coefficient; A hierarchical and progressive defect control strategy is adopted. The defect diffusion kinetic equation is established based on the material lattice structure. The barrier height and position parameters of the multiple defect barriers are determined in combination with the loss weight coefficient. A temperature gradient auxiliary control mechanism is introduced to form a controllable temperature gradient field to obtain the optimal migration path. Based on the optimal migration path, the changes in the material electrical properties during the defect control process are monitored in real time, the defect control parameters corresponding to the optimal migration path are optimized and performance verified in real time, and the performance improvement indicators and control parameters are stored in the optimization parameter database.

[0019] In an optional embodiment, Collecting a cross-sectional image of the third-generation semiconductor material to be tested and obtaining electron mobility distribution data and carrier concentration data corresponding to the cross-sectional image includes: Performing cross-sectional scanning imaging of the third-generation semiconductor material using a transmission electron microscope, focusing the electron beam to the cross-sectional position of the third-generation semiconductor material, collecting interaction signals between the electron beam and the third-generation semiconductor material, and generating a cross-sectional image of the third-generation semiconductor material; Four detection electrodes are set on the surface of the third-generation semiconductor material. By applying a scanning voltage between adjacent detection electrodes, Hall voltage data is collected under the action of a perpendicular magnetic field and the electron mobility of each area of ​​the third-generation semiconductor material is calculated to obtain electron mobility distribution data; The current data between the four detection electrodes is measured, and the resistivity of the surface of the third-generation semiconductor material is calculated. Combined with the electron mobility distribution data, the carrier concentration data is calculated.

[0020] To acquire cross-sectional images of the third-generation semiconductor material under test (such as gallium nitride, silicon carbide, or gallium oxide), a slice is sliced ​​to a thickness of 100 nanometers and further thinned to approximately 30 nanometers using ion thinning technology to ensure effective electron beam penetration. The prepared sample is secured to the sample stage. The microscope's accelerating voltage is adjusted to 200 kV, and the beam current density is controlled at 20 mA / cm². The electron beam is focused onto the sample cross section, with the focal spot diameter precisely controlled to within 1 nanometer. The scanning range is set to 10 μm × 10 μm, with an acquisition resolution of 4096 × 4096 pixels. A high-angle annular dark field (HAADF) detector is used to detect scattered electron signals. Signals generated by the interaction between the electron beam and the sample, including transmitted electrons, secondary electrons, and characteristic X-ray signals, are collected in real time and digitally converted to produce a high-resolution image of the sample cross section. For the gallium nitride sample, the hexagonal lattice structure is clearly observed, with a lattice constant of approximately 0.318 nanometers. For the silicon carbide sample, polymorphic structural features can be identified.

[0021] To acquire electron mobility distribution data, four metal detector electrodes were deposited on the sample surface, spaced 500 microns apart and arranged in a square pattern. The electrodes consisted of a titanium / gold (Ti / Au, 20 nm / 150 nm) bilayer structure deposited by electron beam evaporation to ensure ohmic contact with the semiconductor material. The sample was placed in a uniform 0.5 Tesla perpendicular magnetic field, directed perpendicular to the sample surface. A high-precision current source was used to apply a DC current between adjacent electrodes (e.g., between electrodes 1 and 2) starting at 0.1 mA in steps of 0.1 mA to a maximum of 1.0 mA, generating 10 different sets of measurement data. A nanovolt-precision voltmeter was used to measure the Hall voltage generated between the remaining two electrodes (e.g., electrodes 3 and 4). The sample surface was scanned in a dot matrix pattern with a 50-micron spacing, forming a 20×20 measurement grid, allowing for Hall voltage data acquisition at 400 locations on the sample surface. The control system automatically switches the power supply and measurement mode of the electrode pairs, collecting eight sets of data at each point (forward and reverse measurements are performed for each of the four electrode configurations). The typical Hall voltage values ​​measured for gallium nitride samples range from 100 to 500 microvolts, and for silicon carbide samples from 200 to 800 microvolts. Based on the collected Hall voltage data, combined with the known magnetic field strength, applied current value, and sample geometry, the electron mobility at each point is calculated to generate an electron mobility distribution data matrix covering the entire sample surface. The measured electron mobility of gallium nitride materials varies from 1200 to 1800 cm2 / V·s in different regions, and for silicon carbide materials from 700 to 950 cm2 / V·s.

[0022] Carrier concentration measurements are made using a four-probe resistivity technique. Using the same four electrodes, a semiconductor parameter analyzer applies a constant current (e.g., 1 mA) between the first pair of adjacent electrodes while simultaneously measuring the voltage drop between the second pair of adjacent electrodes. To reduce measurement error, the four electrodes are rotated, acquiring four sets of current-voltage data. For gallium nitride samples, a typical voltage drop of approximately 25-40 mV at a current of 1 mA is measured; for silicon carbide samples, the voltage drop is approximately 45-65 mV at the same current. The sheet resistance of the sample is calculated based on the van der Pauw equation, taking into account the sample thickness (typically 2 microns), and the bulk resistivity is converted. The calculated resistivity for gallium nitride samples is approximately 0.015-0.025 ohm·cm, and for silicon carbide samples, it is approximately 0.05-0.08 ohm·cm. Combined with the previously obtained electron mobility distribution data, the carrier (electron) concentration is calculated for each measurement point based on the relationship between resistivity and electron mobility. The calculated carrier concentration of gallium nitride samples ranges from 2×1017cm-3 to 3×1018cm-3, and the carrier concentration of silicon carbide samples ranges from 1×1016cm-3 to 5×1017cm-3.

[0023] The measured data were visualized to generate a two-dimensional distribution map of electron mobility and carrier concentration. The color gradient represents the numerical variation in different regions, with red representing high values ​​and blue representing low values. For the GaN sample, a decrease in electron mobility of approximately 15% and an increase in carrier concentration of approximately 20% were observed near grain boundaries. In defect-dense regions, the decrease in electron mobility can reach over 30%. For the SiC sample, the polytype transition region exhibits significant electron mobility fluctuations, with a variation of approximately 25%.

[0024] In this embodiment, cross-sectional scanning imaging is performed through a transmission electron microscope, which can intuitively observe the microstructural characteristics inside the third-generation semiconductor material, accurately obtain the position, type and distribution of interface defects, and provide a structural characterization basis for subsequent defect regulation. The four-point detection method is combined with Hall effect measurement to achieve two-dimensional spatial distribution measurement of electron mobility under the action of a vertical magnetic field, avoiding the limitations of traditional single-point measurement methods and improving the spatial resolution of electrical property characterization. Through the correlation calculation of resistivity and electron mobility, indirect characterization of carrier concentration is achieved, which is simpler to operate than direct measurement methods and can obtain the spatial distribution characteristics of carrier concentration, providing accurate experimental data support for studying the influence of defect regulation on the electrical properties of materials.

[0025] In an optional embodiment, A high-resolution morphology feature extraction channel and a local stress distribution feature extraction channel are respectively constructed based on the cross-sectional image. The electron mobility distribution data and carrier concentration data are used as auxiliary parameters to extract the three-dimensional morphology information and stress field distribution information of the interface defect and perform multi-scale feature joint analysis to generate a multi-dimensional feature vector of the interface defect. Perform multi-scale morphological feature enhancement on cross-sectional images of third-generation semiconductor materials. A morphological feature enhancement function is constructed based on a Gaussian kernel function and a Laplace operator. The enhanced cross-sectional image is then generated and its boundary extracted using an adaptive threshold segmentation algorithm, where the adaptive threshold is dynamically calculated based on the statistical characteristics of pixel values ​​within a local window. A high-resolution morphology feature extraction channel and a local stress distribution feature extraction channel are respectively constructed according to the cross-sectional image, the boundary extraction result is used as input data, and the electron mobility distribution data and the carrier concentration data are input as auxiliary parameters into the high-resolution morphology feature extraction channel and the local stress distribution feature extraction channel to extract three-dimensional morphology information of the interface defect and stress field distribution information of the interface defect; performing a multi-scale feature joint analysis on the three-dimensional morphology information and the stress field distribution information, performing weighted fusion on the three-dimensional morphology information and the stress field distribution information using a feature fusion function, and dynamically adjusting feature weights based on an adaptive weight optimization equation; The fused features are input into the feature space mapping function for feature transformation, and feature extraction is performed through a nonlinear feature extraction operator to generate a multidimensional feature vector of the interface defect. The multidimensional feature vector contains three-dimensional morphology information and stress field distribution information of the interface defect.

[0026] The cross-sectional images of semiconductor materials are enhanced with multi-scale morphological features, and high-resolution morphological feature extraction channels and local stress distribution feature extraction channels are constructed. Multi-dimensional feature vectors of interface defects are generated through joint analysis of multi-scale features.

[0027] When performing multi-scale morphological feature enhancement, a Gaussian kernel function and Laplacian operator are used to construct the morphological feature enhancement function. The input silicon carbide semiconductor cross-sectional image is first smoothed using Gaussian kernel functions of different scales. The kernel width parameter can be set to 3, 5, or 7 pixels to accommodate feature extraction at different scales. The smoothed image is then edge-enhanced using the Laplacian operator, with an enhancement factor set to 1.5. The enhanced image significantly improves the edge definition of the interface defect region. In experiments, a 1024×768 pixel cross-sectional image was processed. The contrast of the defect edge region was 0.25 before enhancement, but increased to 0.78 after enhancement.

[0028] The enhanced cross-sectional image is segmented using an adaptive thresholding algorithm for boundary extraction. A threshold is dynamically calculated based on the statistical characteristics of the pixel values ​​within a local window. The window size can be set to 25×25 pixels. The threshold calculation considers the mean and standard deviation of the pixels within the window, and the threshold coefficient can be set to 0.9. For example, for micron-scale defects in a silicon carbide cross-sectional image, when the local pixel mean is 120 (grayscale value range 0-255) and the standard deviation is 15, the calculated local threshold is 133.5. Segmentation using this threshold achieves boundary definition exceeding 95%.

[0029] When constructing a high-resolution morphological feature extraction pipeline, edge extraction results are used as input data, with electron mobility distribution data (ranging from 800-2000 cm² / V·s) and carrier concentration data (ranging from 1×10¹⁶ to 5×10¹⁸ cm⁻³) serving as auxiliary parameters. The pipeline employs a multi-layered architecture: the first layer uses 16 3×3 convolution kernels to extract edge features, the second layer uses 32 3×3 convolution kernels to extract texture features, and the third layer uses 64 3×3 convolution kernels to extract shape features. During processing, feature extraction accuracy reaches 92.5% at a carrier concentration of 2×10¹⁷ cm⁻³. For silicon carbide interface defect regions, morphological feature extraction yields high-precision three-dimensional morphological information, including defect depth distribution (ranging from 0.1 to 5.0 microns), width distribution (ranging from 0.2 to 10.0 microns), and area distribution (ranging from 0.03 to 50 square microns).

[0030] The local stress distribution feature extraction channel uses the boundary extraction results as input, but focuses on extracting stress field distribution information. It includes a stress field estimation layer and a stress feature analysis layer. The stress field estimation layer calculates the stress field distribution in the interface defect area based on electron mobility and carrier concentration data. When the electron mobility is 1500cm2 / V·s, the stress field estimation accuracy can reach 89%. The stress feature analysis layer further extracts features such as the direction distribution, intensity distribution, and gradient distribution of the stress field. In practical applications, for interface defects with a depth of 2.5 microns, the measured stress field intensity distribution ranges from 0.5 to 3.0 GPa, and the stress directions are mainly concentrated in the two intervals of 45°±15° and 135°±15°. The maximum stress gradient occurs at the defect boundary and can reach 0.8 GPa / micron.

[0031] During the multi-scale feature joint analysis phase, a feature fusion function is used to weightedly fuse the three-dimensional morphology information and stress field distribution information. The initial weight of the morphology feature is set to 0.6, and the initial weight of the stress field feature is set to 0.4. The feature weights are dynamically adjusted based on an adaptive weight optimization equation. When the defect morphology complexity exceeds the threshold of 0.75, the morphology feature weight automatically increases to 0.7-0.8; when the stress field gradient exceeds 0.5 GPa / micron, the stress field feature weight automatically increases to 0.5-0.6. When processing dislocation defects in silicon carbide, after 50 iterative optimizations, the morphology feature weight was 0.65, and the stress field feature weight was 0.35. The accuracy of the fused features increased to 96.3%.

[0032] The fused features are input into a feature space mapping function for feature transformation, and key features are extracted using a nonlinear feature extraction operator. Feature space mapping employs dimensionality reduction techniques to reduce the original high-dimensional features (approximately 500 dimensions) to 64 dimensions. The nonlinear feature extraction operator applies an activation function for nonlinear transformation, extracting more discriminative features. The resulting multidimensional feature vector contains both the three-dimensional morphology and stress field distribution information of the interface defect. The vector has 64 dimensions, with the first 32 dimensions representing morphological features and the last 32 representing stress field features.

[0033] In this embodiment, multi-scale morphological feature enhancement processing and adaptive threshold segmentation algorithm are adopted to significantly improve the imaging quality and boundary recognition accuracy of interface defects. The dynamic calculation strategy of the adaptive threshold ensures the local adaptability of boundary extraction, effectively overcoming the limitations of the traditional fixed threshold method in processing non-uniform images. By constructing a dual-channel feature extraction architecture, the coordinated characterization of interface defect morphology information and stress field distribution is realized. The electrical characteristic data is input into the feature extraction channel as an auxiliary parameter, and a correlation mapping between the defect structural characteristics and the electrical properties is established, providing more complete defect characterization information. The multi-dimensional feature vector generated by the nonlinear feature extraction operator not only retains the complete feature information of the defect, but also establishes a quantitative characterization system of the defect characteristics, providing a reliable data basis for subsequent defect evolution analysis and regulation optimization.

[0034] In an optional embodiment, The input is fed into the pre-trained heterogeneous graph attention network, which outputs the type, density, and distribution information of interface defects through an adaptive weight distribution mechanism, including: The multidimensional feature vector is input into the pre-trained heterogeneous graph attention network. The neighborhood nodes within the preset radius of the multidimensional feature vector are weighted and summed to obtain the local density feature. The local density feature is averaged and weighted with the local density feature variance to obtain the global density feature. The local density feature and the global density feature are concatenated with the multidimensional feature vector and subjected to linear transformation and activation function to obtain the dynamic adjacency matrix. Calculate the density gradient of the local density feature under different feature extraction radii, obtain the scale weight coefficient by linear transformation of the density gradient, and add the product of the density gradient and the multidimensional feature vector weighted by the scale weight coefficient to the residual feature of the multidimensional feature vector to obtain the density flow feature; Based on the dynamic adjacency matrix and density flow features, the multidimensional feature vector is updated through the pre-trained heterogeneous graph attention network to obtain the updated feature vector. The updated feature vector is linearly transformed to obtain the current density distribution. The KL divergence between the current density distribution and the density distribution at the previous moment and the Euclidean distance between the current density distribution and the target density distribution are calculated to obtain the evolution constraint. The updated feature vector, density flow feature, local density feature and global density feature are spliced ​​and input into a multilayer perceptron to obtain the feature importance score. The adaptive weight coefficient is calculated based on the feature importance score. The product of the adaptive weight coefficient and the updated feature vector is input into the pre-trained heterogeneous graph attention network. The type, density and distribution information of the interface defects are output in combination with the evolutionary constraints.

[0035] Obtain a multidimensional feature vector representing the interface defect. This vector contains multiple dimensions of information characterizing the interface material properties, defect characteristics, and environmental factors. In practical applications, multidimensional feature vectors are extracted from preprocessed high-resolution microscopy images or spectral data and typically have 64 to 256 dimensions. For metal interface defects, the feature vector includes lattice strain, element distribution, and dislocation density; for semiconductor interfaces, the feature vector includes band structure and carrier concentration.

[0036] For the input multidimensional feature vector, a weighted summation operation is performed on neighboring nodes within a preset radius (in practice, this can be set to 5nm to 100nm, depending on the interface type). The Euclidean distance between each node and its surrounding nodes is calculated, and a Gaussian kernel function is applied to assign weights to generate local density features. For example, for a 50nm×50nm interface region, a 32-dimensional local density feature can be extracted at each point, effectively characterizing the aggregation state of defects within the microscopic region.

[0037] The average value of the local density signature is calculated, representing the overall defect density level at the interface. The variance of the local density signature is also calculated, reflecting the unevenness of the defect distribution. The global density signature is then weighted and summed with the average and variance in a ratio of 0.7:0.3. In practice, for evenly distributed defects, the global density signature typically ranges from 0.1 to 0.3; for highly concentrated defect areas, the local density signature can reach values ​​exceeding 0.8.

[0038] The local density features, global density features, and the original multidimensional feature vector are concatenated to form an enhanced feature vector. This enhanced feature vector is processed through a linear transformation layer (with a weight matrix dimension of input dimension × output dimension, typically 256 × 256) and a ReLU activation function to generate a dynamic adjacency matrix. The dynamic adjacency matrix describes the strength of associations between nodes in the feature space, with values ​​ranging from 0 to 1. Elements greater than a preset threshold (e.g., 0.5) indicate valid connections between nodes.

[0039] The density gradient of local density features at different feature extraction radii (e.g., 10nm, 20nm, and 50nm) is calculated. The density gradient is obtained by dividing the difference between local density features at adjacent radii by the radius difference, reflecting the spatial variation of defect density. The density gradient is input into a linear transformation layer to generate a scale weight coefficient, which is used to adjust the importance of features at different scales. The element-wise product of the density gradient and the multidimensional feature vector is calculated and weighted by the scale weight coefficient. The weighted result is added to the residual features of the multidimensional feature vector (obtained through the residual network structure) to generate the density flow feature.

[0040] Based on the dynamic adjacency matrix and density flow features, the multidimensional feature vector is input into a pre-trained heterogeneous graph attention network. The heterogeneous graph attention network consists of three to five graph attention layers, each with eight attention heads and a hidden layer dimension of 128. The network aggregates and updates node features through a message passing mechanism, outputting an updated feature vector. A linear transformation (dimension of 128×64) is applied to the updated feature vector and normalized using the Softmax function to obtain the current density distribution.

[0041] The KL divergence between the current density distribution and the previous density distribution is calculated. The KL divergence reflects the temporal evolution of the defect distribution. The Euclidean distance between the current density distribution and the preset target density distribution (the ideal defect-free state) is also calculated. The KL divergence and Euclidean distance are combined in a 1:2 ratio to form an evolution constraint. This evolution constraint guides the model in learning the formation and evolution of defects.

[0042] The updated feature vector, density flow features, local density features, and global density features are concatenated and fed into a three-layer fully connected network (with layer dimensions of 512, 256, and 128, respectively). Feature importance scores are then processed using a sigmoid function and converted into adaptive weight coefficients ranging from 0 to 1. The element-wise product of the adaptive weight coefficients and the updated feature vector is calculated to produce a weighted feature vector. This weighted feature vector is then fed into the final classification layer of a pre-trained heterogeneous graph attention network. Combined with evolutionary constraints, it outputs the interface defect type (e.g., point defects, line defects, surface defects), density (number of defects per square micron), and spatial distribution information (two-dimensional or three-dimensional coordinate mapping).

[0043] In this embodiment, by introducing a heterogeneous graph attention network and combining it with a dynamic adjacency matrix construction method, a multi-scale dynamic characterization of interface defect features is achieved. A multi-scale density characterization method based on local density features and global density features accurately depicts the evolution of defect density distribution through the calculation of density flow features. An adaptive weight calculation mechanism based on feature importance scores, combined with an optimization strategy with evolutionary constraints, enables dynamic adjustment of feature weights. Existing technologies typically use single feature extraction and static network structures to characterize interface defects, making it difficult to capture the dynamic evolution of defect characteristics. Characterization relies solely on local features, resulting in insufficient accuracy in defect evolution prediction. Fixed weight strategies are often used, making it difficult to adapt to the characteristic differences of different defect types. This embodiment realizes the precise extraction and dynamic characterization of interface defect features, has higher characterization accuracy and stronger adaptability, and provides important technical support for the precise regulation of interface defects. The adaptive feature extraction mechanism significantly improves the versatility of the characterization method and can be widely used in the characterization of different types of interface defects.

[0044] Figure 2 This is a comparison chart of the interface defect density distribution prediction accuracy of the third-generation semiconductor material interface defect feature intelligent identification optimization method according to an embodiment of the present invention, showing the comparison of the prediction accuracy of the three methods under different defect density ranges. This technical solution shows obvious advantages in all defect density ranges, and the prediction results are more stable and reliable. When the defect density is 0.5×1010 / cm2, the prediction accuracy of this technical solution is 92.3%, the dynamic graph convolutional network (DGCN) is 82.7%, and the boundary element method (BEM) is 79.5%; when the defect density is 1.0×1010 / cm2, the prediction accuracy of the three methods are 94.7%, 85.2% and 80.3% respectively; in the defect density range of 2.0×1010 / cm2, this technical solution achieved a maximum prediction accuracy of 96.8%, while the dynamic graph convolutional network and boundary element method were 87.6% and 81.1% respectively; the defect density is 5.0 ×1010 / cm2, the prediction accuracies of the three methods were 96.2%, 88.4% and 80.4%, respectively; when the defect density increased to 10.0×1010 / cm2, the prediction accuracies became 95.3%, 86.9% and 78.9%, respectively; when the defect density further increased to 15.0×1010 / cm2, the prediction accuracies were 93.1%, 84.5% and 77.8%, respectively; at the highest defect density of 20.0×1010 / cm2, the proposed technical solution still maintained a high accuracy of 91.4%, while the accuracy of DGCN and BEM dropped to 82.1% and 76.2%, respectively.

[0045] This technical solution integrates local and global density features, capturing the spatial correlation of defect distribution through a dynamic adjacency matrix and density flow features. Particularly at high densities, the adaptive weighting coefficients of this technical solution effectively suppress noise and highlight key features, maintaining high prediction accuracy. The curve shape also shows that this technical solution performs more evenly and has greater adaptability across different density ranges, which is of great significance for solving the problem of uneven defect distribution commonly encountered in actual material interface analysis.

[0046] In summary, this technical solution demonstrates stable and high-precision prediction capabilities under various defect density conditions, providing a reliable tool for the accurate characterization of interface defect density distribution. The medium defect density range (2.0-5.0×1010 / cm2) is the area where this technical solution performs best, which is highly consistent with the defect density range that is most commonly concerned in actual materials science research, further demonstrating the value of this technical solution in practical applications.

[0047] In an optional embodiment, Based on the type, density and distribution information of the interface defects, a material performance loss evaluation model is constructed to calculate the influence of material performance on interface defects and determine the loss weight coefficient, including: The type, density, and distribution characteristics of interface defects are input into a loss assessment model based on a multi-layer neural network, and the defect characteristics are mapped to the performance loss space through nonlinear transformation. Residual connections are used to enhance the output features of the loss assessment model, and the influence of interface defects on material properties is calculated in combination with pre-set material performance evaluation indicators. The influence of interface defects is weighted and normalized according to the importance distribution of material performance indicators to obtain the loss weight coefficient.

[0048] Information on the type, density, and distribution characteristics of interface defects is collected. Interface defect type characteristics include vacancies, dislocations, grain boundaries, and impurities; density characteristics include the number of defects per unit area or volume; and distribution characteristics include information such as the spatial uniformity and degree of clustering of defects. These characteristics are acquired through characterization techniques such as electron microscopy and X-ray diffraction, and digitized to form feature vectors. For example, for an aluminum-steel composite interface, a dislocation density of 3.2 × 109 / cm2 and a grain boundary density of 1.5 × 105 / cm2 may be detected, with the distribution exhibiting central clustering.

[0049] The collected interface defect features are input into the loss assessment model, which is built on a multi-layer neural network consisting of an input layer, multiple hidden layers, and an output layer. The number of nodes in the input layer matches the feature dimensions. For example, an input layer containing 15-dimensional defect type features, 5-dimensional density features, and 10-dimensional distribution features has a total of 30 nodes. The model contains three hidden layers, with 64, 32, and 16 nodes, respectively. The ReLU activation function introduces nonlinear transformation capabilities, allowing the model to map defect features to a performance loss space and capture the complex relationship between defect features and performance loss.

[0050] The neural network is trained using a supervised learning approach, utilizing a dataset containing known interface defect characteristics and corresponding measured material properties. For example, for aluminum-copper composites, 500 sets of data on performance indicators such as tensile strength and conductivity under different interface defect conditions can be collected. Training is performed using the Adam optimizer with a learning rate of 0.001, a batch size of 32, and 500 training epochs. To prevent overfitting, a weight constraint with an L2 regularization coefficient of 0.0001 and a node random dropout strategy with a dropout ratio of 0.2 are introduced.

[0051] To further improve model performance, the output features of the loss assessment model are enhanced using residual connection technology. By adding cross-layer connections to the network, the deep network can directly learn the residual relationship with the shallow features, effectively alleviating the vanishing gradient problem of deep neural networks. In this embodiment, the residual connection is set between the first and third hidden layers, and feature fusion is achieved through element-wise addition operations. The features enhanced by residuals can retain more original input information, improving the model's ability to fit the relationship between interface defects and performance loss.

[0052] The impact of interface defects is quantitatively analyzed using pre-set material performance evaluation indicators. These performance evaluation indicators include mechanical properties (tensile strength, elastic modulus, fracture toughness), thermal properties (thermal conductivity, thermal expansion coefficient), and electrical properties (electrical conductivity, dielectric constant). The model output represents the degree to which different performance indicators are affected by interface defects, ranging from 0 to 1, with larger values ​​indicating greater loss. For example, for an aluminum alloy composite material with a dislocation density of 3.8×10^8 / cm² at the interface, the model predicts a tensile strength loss of 0.23, a thermal conductivity loss of 0.15, and an electrical conductivity loss of 0.31.

[0053] Different importance weights need to be assigned to each performance indicator based on the different requirements for material performance in different application scenarios. For example, in high-temperature structural material applications, mechanical properties and thermal stability are more important; while in electronic packaging materials, electrical conductivity and thermal expansion matching are more critical. In this embodiment, the importance weights of each performance indicator are set using the hierarchical analysis method, such as a mechanical performance weight of 0.5, a thermal performance weight of 0.3, and an electrical performance weight of 0.2. The degree of loss of each performance indicator is multiplied by its importance weight, and the sum is calculated to obtain a comprehensive loss index.

[0054] The loss weight coefficient is obtained by normalizing the comprehensive loss index. Normalization uses the Min-Max scaling method to map the loss value to a range between 0.1 and 1, ensuring that the material retains at least a 10% safety margin in the case of complete loss. The normalization formula subtracts the minimum loss value from the original loss value, divides it by the difference between the maximum and minimum loss values, multiplies it by 0.9, and adds 0.1. For example, for a batch of composite material samples, the minimum comprehensive loss index is 0.12 and the maximum is 0.68. For a sample with a comprehensive loss index of 0.35, the normalized loss weight coefficient is 0.46.

[0055] In this embodiment, by constructing a loss assessment model based on a multi-layer neural network, a nonlinear mapping of interface defect characteristics to performance loss is achieved, and a quantitative correlation between defect characteristics and material properties is established, providing a quantifiable evaluation basis for defect regulation. The residual connection mechanism is introduced to enhance the output characteristics of the loss assessment model, effectively avoiding the gradient vanishing problem in the deep network training process, improving the feature extraction capability of the model, and ensuring the accuracy and reliability of the performance loss assessment results. A weighted calculation method based on the importance of material performance indicators is adopted to achieve differentiated characterization of the impact of different performance indicators on the comprehensive performance of the material. The loss weight coefficient obtained by normalization processing can accurately reflect the degree of influence of interface defects on material performance, providing an important reference for the subsequent optimization of defect regulation strategies.

[0056] In an optional embodiment, A hierarchical and progressive defect control strategy is adopted. Based on the material lattice structure, the defect diffusion kinetic equation is established. The barrier height and position parameters of the multiple defect barriers are determined in combination with the loss weight coefficient. A temperature gradient auxiliary control mechanism is introduced to form a controllable temperature gradient field. The optimal migration path is obtained, including: Collect the defect type and defect distribution information of the material lattice structure, determine the diffusion coefficient matrix according to the defect type, construct the defect source term function based on the defect distribution information, and substitute the diffusion coefficient matrix and the defect source term function into the diffusion kinetics equation to calculate the defect concentration distribution; The lattice distortion energy is calculated based on the elastic constant tensor based on the defect concentration distribution to obtain the strain tensor, and the defect formation energy is corrected according to the strain tensor and the local defect density to obtain the corrected defect formation energy; The modified defect formation energy is input into the Gaussian distribution function and combined with the loss weight coefficient to calculate the barrier height of the multiple defect barriers. The gradient of the strain tensor and the gradient of the local defect density are used to determine the barrier position parameters. The interaction energy of the multiple defect barriers is calculated based on the coupling coefficient between the barriers to obtain the total barrier energy distribution. A temperature gradient-assisted control mechanism is constructed based on the total barrier energy distribution. The reference temperature, the spatial distribution function of the total barrier energy, and the time modulation function are input into the pre-set heat diffusion equation to obtain the evolution characteristics of the controllable temperature gradient field. A transition probability function is constructed based on the evolution characteristics of the total potential barrier energy distribution and the controllable temperature gradient field. The transition probability function is solved under the dual constraints of the temperature gradient restriction condition and the potential barrier energy threshold, and the optimal migration path is obtained by combining the energy gradient iterative calculation.

[0057] In silicon crystals, a scanning tunneling microscope scans the surface at nanometer-scale resolution to identify defect types, including vacancies, interstitials, and dislocations, and to determine the three-dimensional coordinates of the defects within the lattice. Vacancies are represented by missing points in the silicon lattice, interstitials by extra atoms at non-lattice locations, and dislocations by atoms offset from their ideal lattice positions. A diffusion coefficient matrix is ​​determined based on the defect type: the diffusion coefficient for vacancies is 2.3×10⁻¹⁴ square centimeters per second, for interstitials by 1.7×10⁻¹⁴ square centimeters per second, and for dislocations by 0.9×10⁻¹⁴ square centimeters per second. A defect source function is constructed based on the spatial distribution of defects in the lattice, taking values ​​of 3.6×10⁻¹⁸ per cubic centimeter in high-density regions and 5.2×10⁻¹⁶ per cubic centimeter in low-density regions. The diffusion coefficient matrix and the defect source function are substituted into the diffusion kinetics equation, and the defect concentration distribution field is calculated using a finite difference method with a time step of 10 nanoseconds and 100 iterations.

[0058] The lattice distortion energy was calculated based on the defect concentration distribution. The elastic constant tensor for silicon crystals (C11 = 165.7 GPa, C12 = 63.9 GPa, and C44 = 79.6 GPa) was used in conjunction with the defect concentration distribution to calculate the local strain tensor. In the high defect density region (1019 cm3), the principal components of the strain tensor are ε11 = 0.0075, ε22 = 0.0068, and ε33 = 0.0082. In the low defect density region (1017 cm3), the principal components of the strain tensor are ε11 = 0.0012, ε22 = 0.0009, and ε33 = 0.0014. The lattice distortion energy calculated from the strain tensor is used to correct the defect formation energy. The original defect formation energies are: 3.6 eV for vacancies, 4.2 eV for interstitials, and 2.8 eV for dislocations. After strain field correction, the corrected defect formation energy in the high defect density region becomes: vacancy 3.85eV, interstitial atom 4.43eV, dislocation 2.96eV; the corrected defect formation energy in the low defect density region becomes: vacancy 3.67eV, interstitial atom 4.26eV, dislocation 2.84eV.

[0059] The barrier heights and position parameters of the multiple defect barriers were calculated by inputting the modified defect formation energy into a Gaussian distribution function and combining the loss weight coefficients (0.85 for vacancies, 0.92 for interstitials, and 0.78 for dislocations). The barrier heights for the primary barrier region were 4.65 eV, the secondary barrier region was 3.82 eV, and the edge barrier region was 2.71 eV. The barrier position parameters were determined using the gradient of the strain tensor and the gradient of the local defect density. The barrier position parameters were 25.6 nm in the x-direction, 18.3 nm in the y-direction, and 30.2 nm in the z-direction. The coupling coefficients between barriers are determined by the distance between them: the coupling coefficient between nearest-neighbor barriers is 0.65, the coupling coefficient between next-nearest-neighbor barriers is 0.32, and the coupling coefficient between distant barriers is 0.11. By calculating the interaction energy between potential barriers, the total barrier energy distribution is obtained, where the total barrier energy of the main barrier region is 5.2eV, the total barrier energy of the secondary barrier region is 4.1eV, and the total barrier energy of the edge region is 2.9eV.

[0060] A temperature gradient-assisted control mechanism was constructed. Based on the total barrier energy distribution, a temperature gradient-assisted control mechanism was constructed. The baseline temperature was set at 800 K, with the local temperature increased to 950 K in high-barrier regions and decreased to 650 K in low-barrier regions. The spatial distribution function of the total barrier energy was mapped into three-dimensional space. The high-energy region (5.0-5.2 eV) was located in the center of the crystal, covering a volume of 125 cubic nanometers; the intermediate-energy region (3.8-4.2 eV) was located in the middle layer of the crystal, covering a volume of 340 cubic nanometers; and the low-energy region (2.8-3.0 eV) was located on the surface of the crystal, covering a volume of 530 cubic nanometers. The temporal modulation function employed a periodic heating and cooling strategy, with a heating phase duration of 200 nanoseconds and a cooling phase duration of 150 nanoseconds, for a total cycle duration of 350 nanoseconds, and 10 complete cycles. The base temperature, the spatial distribution function of the total barrier energy, and the temporal modulation function were input into the heat diffusion equation and solved using the alternating direction implicit finite difference method with a spatial step of 5 nanometers and a temporal step of 2 nanoseconds. The evolution characteristics of the controllable temperature gradient field were obtained through iterative calculations. The maximum gradient of the temperature gradient field in the x-direction is 12 K / nm, the maximum gradient in the y-direction is 9 K / nm, and the maximum gradient in the z-direction is 15 K / nm.

[0061] A transition probability function was constructed and the optimal migration path was solved. The transition probability function was constructed based on the evolutionary characteristics of the total barrier energy distribution and the controllable temperature gradient field. The temperature gradient constraint was set to no more than 20K / nm, and the barrier energy threshold was set to no more than 4.5eV. The transition probability function was sampled using the Monte Carlo method, with each grid point sampled 1000 times. The direction with the highest transition probability was selected as the local optimal direction. Combined with iterative calculations of the energy gradient, the direction with the fastest energy decrease was prioritized. The starting point was set in the high-energy region (5.1eV) and the ending point was set in the low-energy region (2.85eV). The optimal migration path was obtained through iterative calculations. This path traverses 17 key nodes, has a total length of 86.5 nanometers, an average barrier height of 3.45eV, a maximum barrier point of 4.3eV located at 25.6 nanometers along the path, and a minimum barrier point of 2.85eV located at the end of the path.

[0062] In this embodiment, the defect concentration distribution is calculated by the diffusion kinetics equation, combined with the precise characterization of the defect source term function and the diffusion coefficient matrix, to achieve an accurate description of the defect diffusion behavior in the material, providing a reliable theoretical basis for subsequent defect regulation. The elastic constant tensor is used to calculate the lattice distortion energy and correct the defect formation energy. The influence of the local defect density on the formation energy is taken into account, which improves the accuracy of the defect formation energy calculation and makes the barrier energy calculation more consistent with the actual physical process. A multiple defect barrier model is constructed based on the corrected defect formation energy and the loss weight coefficient, and the interaction energy is calculated by the coupling coefficient between the barriers, which achieves an accurate description of the complex interactions of multiple defect barriers and provides complete barrier information for optimizing the migration path.

[0063] In an optional embodiment, Based on the evolution characteristics of the total barrier energy distribution and the controllable temperature gradient field, a transition probability function is constructed. The transition probability function is solved under the dual constraints of the temperature gradient restriction condition and the barrier energy threshold. The optimal migration path obtained by combining the energy gradient iterative calculation includes: A transition probability function is constructed based on the total barrier energy distribution and the evolution characteristics of the controllable temperature gradient field. The maximum temperature gradient in the controllable temperature gradient field is extracted as the equivalent temperature parameter. A set of candidate positions is generated within a preset radius of the current path position. The total barrier energy of each candidate position in the candidate position set is calculated based on the total barrier energy distribution. Substitute the difference between the total barrier energy of each candidate position and the total barrier energy of the current path position and the Laplace operator of the total barrier energy distribution into the transition probability function to obtain the transition probability, select the candidate position with a transition probability greater than a preset threshold as the transition position, and update to the transition position with the maximum transition probability if a transition position exists. If no transition position exists, update along the negative gradient direction of the total barrier energy; Under the dual constraints of temperature gradient restriction and potential barrier energy threshold, the transition position is repeatedly solved until convergence to obtain the optimal migration path.

[0064] Figure 3 The flow chart for constructing the optimal migration path based on the total barrier energy distribution and the controllable temperature gradient field according to the embodiment of the present invention is as follows: Figure 3 As shown, the method includes: Obtain the system's total barrier energy distribution data and controllable temperature gradient field data. The total barrier energy distribution can be obtained through discrete point sampling. For example, barrier energy measurements are performed at 5 nm intervals on a two-dimensional plane, forming a 200 × 200 barrier energy distribution matrix. Controllable temperature gradient field data can be obtained by arranging heat sources and temperature measurement devices. For example, 10 × 10 temperature measurement points are arranged on the same two-dimensional plane, with a measurement range of 20°C to 80°C, and a continuous temperature distribution is obtained through interpolation.

[0065] Based on the acquired data, a transition probability function is constructed. This construction takes into account two key factors: the total barrier energy difference and the temperature gradient. The maximum temperature gradient within the controllable temperature gradient field is extracted as the equivalent temperature parameter, Teff, which characterizes the maximum rate of temperature change in the system. In practical applications, the maximum temperature gradient can reach 5°C / mm. A higher equivalent temperature parameter indicates a more dramatic temperature change in the system, making it easier for particles to overcome the potential barrier and achieve transitions.

[0066] Generate a set of candidate locations S within a preset radius R around the current path position P(x, y). The preset radius R can be set to 10nm, and candidate locations are generated within this range in steps of 1nm. For example, if the current position is P(100nm, 100nm), approximately 314 candidate locations are generated within a circle with a radius of 10nm centered at that point. For each candidate position Pi, calculate its total barrier energy E(Pi).

[0067] Calculate the total barrier energy difference ΔE = E(Pi) - E(P) between each candidate position Pi and the current position P. Simultaneously, calculate the Laplace operator value L(Pi) of the total barrier energy distribution at the candidate position Pi. The Laplace operator can be calculated using the finite difference method, that is, taking the second-order partial derivative of the barrier energy distribution. In practical applications, when the barrier energy is 5 eV, the possible values ​​of the Laplace operator range from -0.01 to 0.01 eV / nm².

[0068] The energy difference ΔE, the Laplace operator value L(Pi), and the equivalent temperature parameter Teff are substituted into the transition probability function to calculate the transition probability P(i→j). The transition probability function uses an exponential form, reflecting the relationship between the energy difference and temperature. A negative energy difference indicates an increased transition probability; a negative Laplace operator indicates a concave region of the potential barrier, where the transition probability increases. In practice, a preset threshold of 0.6 is set for the transition probability, meaning that only candidate locations with a transition probability greater than 0.6 are considered as possible transition locations.

[0069] The transition probabilities of all candidate positions are sorted, and the candidate position with the highest transition probability that is greater than a preset threshold is selected as the next transition position. If a transition position that meets the requirements exists, the current path position is updated to that transition position. If no transition position meets the requirements, the position is updated along the negative gradient of the total barrier energy. This negative gradient update is performed using gradient descent with a step size of 5 nm.

[0070] During the iteration process, temperature gradient constraints and potential barrier energy thresholds are introduced as dual constraints. The temperature gradient constraint ensures that the path does not enter areas with excessive temperature changes, for example, by setting the temperature gradient upper limit to 3°C / mm; the potential barrier energy threshold ensures that the path does not cross excessively high energy barriers, for example, by setting the barrier energy upper limit to 10eV.

[0071] During the iterative calculation process, the convergence condition is checked after each position update. The convergence condition can be set to a position change of less than 1 nm for 10 consecutive iterations or to a maximum number of 500 iterations. When the convergence condition is met, the current path is considered the optimal migration path.

[0072] In a specific application case, the starting position is set at (50nm, 50nm) and the target position is (150nm, 150nm). There are two high-potential barrier regions in the system, located at (80nm-100nm, 80nm-120nm) and (120nm-140nm, 100nm-140nm), with a barrier height of 8eV. The temperature field increases linearly along the x-axis from 25°C to 75°C. Using the above method, the optimal migration path, obtained after 137 iterations, avoids the high-potential barrier regions and completes the migration along the path (50, 50) → (65, 70) → (75, 75) → (110, 90) → (145, 135) → (150, 150). The total path length is 152.3nm, and the average barrier energy is 3.2eV.

[0073] In existing technologies, a single temperature field or fixed potential barrier calculation method is usually used to determine the defect migration path. This method is difficult to adapt to complex defect migration environments and is prone to falling into local optimal solutions. The fixed-step path update strategy lacks the ability to dynamically respond to the local environment. In this embodiment, based on the dynamic selection mechanism of the candidate position set within a preset radius and combined with the calculation method of the transition probability, the adaptive update of the migration path is realized, which significantly improves the flexibility and accuracy of the path optimization. By introducing the comprehensive evaluation mechanism of the total barrier energy difference and the Laplace operator, and designing a dual-branch path update strategy, the path optimization problem in complex environments is effectively solved, the stability of the optimization results is ensured, and reliable theoretical guidance is provided for actual defect control, while improving the efficiency and accuracy of defect control.

[0074] In an optional embodiment, Real-time monitoring of material electrical property changes during defect control based on the optimal migration path, real-time optimization and performance verification of defect control parameters corresponding to the optimal migration path, and storing performance improvement indicators and control parameters in an optimization parameter database include: Real-time monitoring of material electrical property changes during defect control based on the optimal migration path. This includes collecting real-time data on material conductivity, carrier mobility, and carrier concentration distribution, and establishing a quantitative mapping relationship between defect density, defect type, and electrical property changes. Real-time optimization and performance verification of defect control parameters corresponding to the optimal migration path are performed. The performance loss gradient is calculated based on the electrical characteristic change data. The migration rate, barrier height, and temperature gradient in the defect control parameters are optimized and adjusted online according to the performance loss gradient. The electrical performance improvement index before and after the control is calculated. The performance improvement indicators and control parameters are stored in the optimization parameter database and classified according to the corresponding relationship between defect type, defect density and electrical characteristics.

[0075] A sensor array for real-time electrical property monitoring was set up, including distributed conductivity sensors, a Hall-effect carrier mobility measurement module, and a carrier concentration profile scanner. These sensors formed a measurement network at key locations on the material surface and within it, with a sampling frequency of 200 Hz to capture transient changes in electrical properties. Once the defect modulation process was initiated, conductivity data were collected at 10 ms intervals. In silicon-based semiconductor materials, the initial conductivity is typically 3.2 × 10⁻³ S / m. As the oxygen vacancy defect density decreases from 10⁻¹⁹ cm⁻³ to 10⁻¹⁹ cm⁻³, the conductivity increases to 4.7 × 10⁻³ S / m. Simultaneously, the carrier mobility measurement module recorded the change in electron mobility from an initial value of 1420 cm⁻² / (V·s) to 1650 cm⁻² / (V·s). The carrier concentration profile scanner mapped the evolution of the carrier distribution from an inhomogeneous state (coefficient of variation 0.42) to a uniform state (coefficient of variation 0.13).

[0076] Based on real-time data acquisition, an adaptive neural network was used to construct a quantitative mapping between defect density, defect type, and electrical properties. The network consists of a five-layer structure: the input layer receives defect density and type information, and the output layer generates predicted values ​​for conductivity, carrier mobility, and concentration distribution. The number of hidden layer nodes is 128, 256, and 128, respectively, and the activation function uses a modified Reluctant Unit (ReLU) function. The network was trained using a batch size of 64, and after 5000 iterations, the prediction accuracy reached 97.8%. For example, when the oxygen vacancy density is 8.5×1018cm⁻³, the mapping shows a conductivity of 3.8×103S / m, an electron mobility of 1520cm⁻² / (V·s), and a carrier distribution coefficient of variation of 0.25.

[0077] Based on the established mapping relationship, the performance loss gradient is calculated in real time to guide the optimization of defect control parameters. The performance loss gradient is expressed as the partial derivative of the defect control parameter with respect to the difference between the ideal value and the actual value of the electrical performance. The performance loss gradient is calculated every 50ms, and when the gradient value is greater than the preset threshold of 0.05, the parameter optimization process is triggered. For a specific defect migration rate parameter, the system starts from an initial value of 2.3×10-6m / s and adjusts it to 3.1×10-6m / s based on the performance loss gradient calculation results, increasing the conductivity improvement rate by 42%. The barrier height parameter is optimized from 0.78eV to 0.71eV, and the temperature gradient is adjusted from the initial setting of 25K / mm to 32K / mm. Together, they increase the carrier mobility improvement rate by 35%.

[0078] A closed-loop feedback mechanism was used to verify the effect of parameter optimization. By comparing the electrical performance data of 1,000 test points before and after optimization, the performance improvement indicators were quantified. In a complete optimization cycle, the average conductivity of silicon-based semiconductor materials increased by 46.2%, from 3.2×10^3S / m to 4.68×103S / m; the carrier mobility increased by 16.2%, from 1420cm2 / (V·s) to 1650cm2 / (V·s); the carrier distribution uniformity increased by 68.5%, and the coefficient of variation decreased from 0.42 to 0.13. During the performance verification process, a cross-validation method was used to divide the material into 9 areas for independent testing to ensure the uniformity and stability of the performance improvement.

[0079] The optimized performance improvement indicators and corresponding control parameters are automatically stored in a hierarchical optimization parameter database. The database uses a relational data structure, with a primary index based on defect type, including oxygen vacancies, silicon vacancies, and interstitial carbon atoms. A secondary index is established based on defect density range, categorized as high density (>1019cm-3), medium density (1017-1019cm-3), and low density (<1017cm-3).

[0080] In this embodiment, by calculating the electrical performance improvement index, dynamic optimization of key parameters such as migration rate, barrier height and temperature gradient is achieved, which significantly improves the accuracy and efficiency of defect control. Based on the optimization parameter database of multi-dimensional correlation relationships, a complete parameter optimization knowledge base is established by classifying and storing performance improvement indicators and control parameters, which provides reliable data support for the continuous improvement of defect control, improves the accuracy and efficiency of control parameter optimization, and at the same time establishes a knowledge accumulation mechanism for sustainable optimization, which provides a strong guarantee for the continuous improvement of material performance.

[0081] A second aspect of an embodiment of the present invention provides a system for intelligently identifying and optimizing interface defect characteristics of third-generation semiconductor materials, comprising: The first unit is used to collect a cross-sectional image of the third-generation semiconductor material to be tested, and obtain electron mobility distribution data and carrier concentration data corresponding to the cross-sectional image; The second unit is used to construct a high-resolution morphology feature extraction channel and a local stress distribution feature extraction channel based on the cross-sectional image, respectively, and use the electron mobility distribution data and carrier concentration data as auxiliary parameters to extract the three-dimensional morphology information and stress field distribution information of the interface defect and perform multi-scale feature joint analysis, generate a multi-dimensional feature vector of the interface defect, and input it into a pre-trained heterogeneous graph attention network, and output the type, density and distribution information of the interface defect through an adaptive weight allocation mechanism; The third unit is used to construct a material performance loss evaluation model based on the type, density and distribution information of the interface defects, calculate the influence of material performance on interface defects and determine the loss weight coefficient; The fourth unit is used to adopt a hierarchical progressive defect control strategy, establish a defect diffusion kinetic equation based on the material lattice structure, determine the barrier height and position parameters of the multiple defect barriers in combination with the loss weight coefficient, introduce a temperature gradient auxiliary control mechanism to form a controllable temperature gradient field, and obtain the optimal migration path; The fifth unit is used to monitor the changes in the electrical properties of the material during the defect control process in real time based on the optimal migration path, perform real-time optimization and performance verification on the defect control parameters corresponding to the optimal migration path, and store the performance improvement indicators and control parameters in the optimization parameter database.

[0082] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including: A processor and a memory for storing processor-executable instructions, wherein the processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0083] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0084] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent identification and optimization method for interface defect characteristics of third-generation semiconductor materials, characterized by: include: Collecting a cross-sectional image of the third-generation semiconductor material to be tested, and obtaining electron mobility distribution data and carrier concentration data corresponding to the cross-sectional image; A high-resolution morphology feature extraction channel and a local stress distribution feature extraction channel are respectively constructed based on the cross-sectional image. The electron mobility distribution data and carrier concentration data are used as auxiliary parameters to extract the three-dimensional morphology information and stress field distribution information of the interface defect and perform a multi-scale feature joint analysis. A multi-dimensional feature vector of the interface defect is generated and input into a pre-trained heterogeneous graph attention network. The type, density, and distribution information of the interface defect are output through an adaptive weight allocation mechanism. Based on the type, density and distribution information of the interface defects, a material performance loss evaluation model is constructed to calculate the influence of material performance on the interface defects and determine the loss weight coefficient; A hierarchical and progressive defect control strategy is adopted. The defect diffusion kinetic equation is established based on the material lattice structure. The barrier height and position parameters of the multiple defect barriers are determined in combination with the loss weight coefficient. A temperature gradient auxiliary control mechanism is introduced to form a controllable temperature gradient field to obtain the optimal migration path. Based on the optimal migration path, the changes in the material electrical properties during the defect control process are monitored in real time, the defect control parameters corresponding to the optimal migration path are optimized and performance verified in real time, and the performance improvement indicators and control parameters are stored in the optimization parameter database.

2. The method according to claim 1, characterized in that Collecting a cross-sectional image of the third-generation semiconductor material to be tested and obtaining electron mobility distribution data and carrier concentration data corresponding to the cross-sectional image includes: Performing cross-sectional scanning imaging of the third-generation semiconductor material using a transmission electron microscope, focusing the electron beam to the cross-sectional position of the third-generation semiconductor material, collecting interaction signals between the electron beam and the third-generation semiconductor material, and generating a cross-sectional image of the third-generation semiconductor material; Four detection electrodes are set on the surface of the third-generation semiconductor material. By applying a scanning voltage between adjacent detection electrodes, Hall voltage data is collected under the action of a perpendicular magnetic field and the electron mobility of each area of ​​the third-generation semiconductor material is calculated to obtain electron mobility distribution data; The current data between the four detection electrodes is measured, and the resistivity of the surface of the third-generation semiconductor material is calculated. Combined with the electron mobility distribution data, the carrier concentration data is calculated.

3. The method according to claim 1, characterized in that A high-resolution morphology feature extraction channel and a local stress distribution feature extraction channel are respectively constructed based on the cross-sectional image. The electron mobility distribution data and carrier concentration data are used as auxiliary parameters to extract the three-dimensional morphology information and stress field distribution information of the interface defect and perform multi-scale feature joint analysis to generate a multi-dimensional feature vector of the interface defect. Perform multi-scale morphological feature enhancement on cross-sectional images of third-generation semiconductor materials. A morphological feature enhancement function is constructed based on a Gaussian kernel function and a Laplace operator. The enhanced cross-sectional image is then generated and its boundary extracted using an adaptive threshold segmentation algorithm, where the adaptive threshold is dynamically calculated based on the statistical characteristics of pixel values ​​within a local window. A high-resolution morphology feature extraction channel and a local stress distribution feature extraction channel are respectively constructed according to the cross-sectional image, the boundary extraction result is used as input data, and the electron mobility distribution data and the carrier concentration data are input as auxiliary parameters into the high-resolution morphology feature extraction channel and the local stress distribution feature extraction channel to extract three-dimensional morphology information of the interface defect and stress field distribution information of the interface defect; performing a multi-scale feature joint analysis on the three-dimensional morphology information and the stress field distribution information, performing weighted fusion on the three-dimensional morphology information and the stress field distribution information using a feature fusion function, and dynamically adjusting feature weights based on an adaptive weight optimization equation; The fused features are input into the feature space mapping function for feature transformation, and feature extraction is performed through a nonlinear feature extraction operator to generate a multidimensional feature vector of the interface defect. The multidimensional feature vector contains three-dimensional morphology information and stress field distribution information of the interface defect.

4. The method according to claim 1, wherein The input is fed into the pre-trained heterogeneous graph attention network, which outputs the type, density, and distribution information of interface defects through an adaptive weight distribution mechanism, including: The multidimensional feature vector is input into the pre-trained heterogeneous graph attention network. The neighborhood nodes within the preset radius of the multidimensional feature vector are weighted and summed to obtain the local density feature. The local density feature is averaged and weighted with the local density feature variance to obtain the global density feature. The local density feature and the global density feature are concatenated with the multidimensional feature vector and subjected to linear transformation and activation function to obtain the dynamic adjacency matrix. Calculate the density gradient of the local density feature under different feature extraction radii, obtain the scale weight coefficient by linear transformation of the density gradient, and add the product of the density gradient and the multidimensional feature vector weighted by the scale weight coefficient to the residual feature of the multidimensional feature vector to obtain the density flow feature; Based on the dynamic adjacency matrix and density flow features, the multidimensional feature vector is updated through the pre-trained heterogeneous graph attention network to obtain the updated feature vector. The updated feature vector is linearly transformed to obtain the current density distribution. The KL divergence between the current density distribution and the density distribution at the previous moment and the Euclidean distance between the current density distribution and the target density distribution are calculated to obtain the evolution constraint. The updated feature vector, density flow feature, local density feature and global density feature are spliced ​​and input into a multilayer perceptron to obtain the feature importance score. The adaptive weight coefficient is calculated based on the feature importance score. The product of the adaptive weight coefficient and the updated feature vector is input into the pre-trained heterogeneous graph attention network. The type, density and distribution information of the interface defects are output in combination with the evolutionary constraints.

5. The method according to claim 1, wherein Based on the type, density and distribution information of the interface defects, a material performance loss evaluation model is constructed to calculate the influence of material performance on interface defects and determine the loss weight coefficient, including: The type, density, and distribution characteristics of interface defects are input into a loss assessment model based on a multi-layer neural network, and the defect characteristics are mapped to the performance loss space through nonlinear transformation. Residual connections are used to enhance the output features of the loss assessment model, and the influence of interface defects on material properties is calculated in combination with pre-set material performance evaluation indicators. The influence of interface defects is weighted and normalized according to the importance distribution of material performance indicators to obtain the loss weight coefficient.

6. The method according to claim 1, characterized in that A hierarchical and progressive defect control strategy is adopted. Based on the material lattice structure, the defect diffusion kinetic equation is established. The barrier height and position parameters of the multiple defect barriers are determined in combination with the loss weight coefficient. A temperature gradient auxiliary control mechanism is introduced to form a controllable temperature gradient field. The optimal migration path is obtained, including: Collect the defect type and defect distribution information of the material lattice structure, determine the diffusion coefficient matrix according to the defect type, construct the defect source term function based on the defect distribution information, and substitute the diffusion coefficient matrix and the defect source term function into the diffusion kinetics equation to calculate the defect concentration distribution; The strain tensor is obtained by calculating the lattice distortion energy through the elastic constant tensor based on the defect concentration distribution, and the defect formation energy is corrected according to the strain tensor and the local density of defects to obtain the corrected defect formation energy; The modified defect formation energy is input into the Gaussian distribution function and combined with the loss weight coefficient to calculate the barrier height of the multiple defect barriers. The gradient of the strain tensor and the gradient of the local defect density are used to determine the barrier position parameters. The interaction energy of the multiple defect barriers is calculated based on the coupling coefficient between the barriers to obtain the total barrier energy distribution. A temperature gradient-assisted control mechanism is constructed based on the total barrier energy distribution. The reference temperature, the spatial distribution function of the total barrier energy, and the time modulation function are input into the pre-set heat diffusion equation to obtain the evolution characteristics of the controllable temperature gradient field. A transition probability function is constructed based on the evolution characteristics of the total potential barrier energy distribution and the controllable temperature gradient field. The transition probability function is solved under the dual constraints of the temperature gradient restriction condition and the potential barrier energy threshold, and the optimal migration path is obtained by combining the energy gradient iterative calculation.

7. The method according to claim 6, characterized in that Based on the evolution characteristics of the total barrier energy distribution and the controllable temperature gradient field, a transition probability function is constructed. The transition probability function is solved under the dual constraints of the temperature gradient restriction condition and the barrier energy threshold. The optimal migration path obtained by combining the energy gradient iterative calculation includes: A transition probability function is constructed based on the total barrier energy distribution and the evolution characteristics of the controllable temperature gradient field. The maximum temperature gradient in the controllable temperature gradient field is extracted as the equivalent temperature parameter. A set of candidate positions is generated within a preset radius of the current path position. The total barrier energy of each candidate position in the candidate position set is calculated based on the total barrier energy distribution. Substitute the difference between the total barrier energy of each candidate position and the total barrier energy of the current path position and the Laplace operator of the total barrier energy distribution into the transition probability function to obtain the transition probability, select the candidate position with a transition probability greater than a preset threshold as the transition position, and update to the transition position with the maximum transition probability if a transition position exists. If no transition position exists, update along the negative gradient direction of the total barrier energy; Under the dual constraints of temperature gradient restriction and potential barrier energy threshold, the transition position is repeatedly solved until convergence to obtain the optimal migration path.

8. The method according to claim 1, characterized in that Real-time monitoring of material electrical property changes during defect control based on the optimal migration path, real-time optimization and performance verification of defect control parameters corresponding to the optimal migration path, and storing performance improvement indicators and control parameters in an optimization parameter database include: Real-time monitoring of material electrical property changes during defect control based on the optimal migration path. This includes collecting real-time data on material conductivity, carrier mobility, and carrier concentration distribution, and establishing a quantitative mapping relationship between defect density, defect type, and electrical property changes. Real-time optimization and performance verification of defect control parameters corresponding to the optimal migration path are performed. The performance loss gradient is calculated based on the electrical characteristic change data. The migration rate, barrier height, and temperature gradient in the defect control parameters are optimized and adjusted online according to the performance loss gradient. The electrical performance improvement index before and after the control is calculated. The performance improvement indicators and control parameters are stored in the optimization parameter database and classified according to the corresponding relationship between defect type, defect density and electrical characteristics.

9. A third-generation semiconductor material interface defect feature intelligent identification and optimization system, used to implement the method according to any one of claims 1 to 8, characterized in that: include: The first unit is used to collect a cross-sectional image of the third-generation semiconductor material to be tested, and obtain electron mobility distribution data and carrier concentration data corresponding to the cross-sectional image; The second unit is used to construct a high-resolution morphology feature extraction channel and a local stress distribution feature extraction channel based on the cross-sectional image, respectively, and use the electron mobility distribution data and carrier concentration data as auxiliary parameters to extract the three-dimensional morphology information and stress field distribution information of the interface defect and perform multi-scale feature joint analysis, generate a multi-dimensional feature vector of the interface defect, and input it into a pre-trained heterogeneous graph attention network, and output the type, density and distribution information of the interface defect through an adaptive weight allocation mechanism; The third unit is used to construct a material performance loss evaluation model based on the type, density and distribution information of the interface defects, calculate the influence of material performance on interface defects and determine the loss weight coefficient; The fourth unit is used to adopt a hierarchical progressive defect control strategy, establish a defect diffusion kinetic equation based on the material lattice structure, determine the barrier height and position parameters of the multiple defect barriers in combination with the loss weight coefficient, introduce a temperature gradient auxiliary control mechanism to form a controllable temperature gradient field, and obtain the optimal migration path; The fifth unit is used to monitor the changes in the electrical properties of the material during the defect control process in real time based on the optimal migration path, perform real-time optimization and performance verification on the defect control parameters corresponding to the optimal migration path, and store the performance improvement indicators and control parameters in the optimization parameter database.

10. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 8.

Images