A physical constraint optimization apparatus, method and system based on in-situ monitoring of semiconductor material defects

Abstract

The application discloses a kind of physical constraint optimization device, method and system based on semiconductor material defect in-situ monitoring, belongs to semiconductor epitaxial manufacturing technical field.The present application includes twin growth chamber, sample transmission component, in-situ cathodoluminescence monitoring unit, multi-source in-situ monitoring system, data collector and edge computing node.First chamber integrated CL unit is used for dynamic monitoring and process exploration, and second chamber is used for lag batch production.Edge computing node deploys physical constraint neural network and double-cavity collaborative control module, processes double-cavity data in real time and issues optimization instructions to deposition execution unit, and constructs millisecond-level closed-loop regulation and control link.By physical equation constraint and double-cavity timing cooperation, millisecond-level accurate regulation and control of low dislocation density epitaxial layer is realized, and traditional epitaxial process monitoring lag and pure AI model prediction distortion defects are effectively overcome.

Description

Technical Field

[0001] This invention belongs to the field of semiconductor epitaxial manufacturing technology, specifically relating to a physical constraint optimization device, method and system based on in-situ monitoring of semiconductor material defects. Background Technology

[0002] Metal-organic chemical vapor deposition (MOCVD) is a core process for fabricating wide-bandgap semiconductor devices such as GaN and SiC. However, existing technologies face the following technical bottlenecks in achieving real-time monitoring and precise control of low-dislocation-density epitaxial growth: First, monitoring methods are inherently lagging. Traditional techniques such as offline XRD and wet etching cannot provide real-time feedback during the process. Although in-situ monitoring technology for MOCVD has been publicly disclosed, it is currently only applicable to defect densities higher than 10. 8 In scenarios with a resolution of / cm²; facing the requirement of less than 10 for 5G RF devices. 6 With a low dislocation density index of / cm², the monitoring sensitivity of this technology is clearly insufficient.

[0003] Second, the accuracy of in-situ monitoring and prediction is insufficient. Existing technologies that use AI to optimize epitaxial growth suffer from prediction errors as high as 38.7% in small sample scenarios (<100 sets of data), which is difficult to meet the accuracy requirements of actual semiconductor production.

[0004] Third, the process control lacks physical mechanism constraints. Existing dual-chamber MOCVD systems fail to effectively address the dynamic mapping of process parameters between the two chambers. In actual operation, when chamber pressure fluctuations exceed 5 Pa, the fixed-weight control strategy results in dislocation density fluctuations as high as ±40%, severely impacting the stability of epitaxial layer crystal quality. Summary of the Invention

[0005] In order to overcome the shortcomings of the prior art, the present invention aims to provide a physical constraint optimization device, method and system based on in-situ monitoring of defects in semiconductor materials.

[0006] To achieve the above objectives, the present invention employs the following technical solution: In a first aspect, this invention discloses a physical constraint optimization device based on in-situ defect monitoring of semiconductor materials, comprising: a chemical vapor deposition system, the chemical vapor deposition system including a first growth chamber, a second growth chamber, a sample transfer assembly, a first in-situ monitoring unit, a second in-situ monitoring unit, an in-situ cathodoluminescence monitoring unit, a high vacuum unit, a first data acquisition unit, a second data acquisition unit, an edge computing node, and a vapor deposition execution unit; the sample transfer assembly is connected to the first growth chamber and the second growth chamber, and the first in-situ monitoring unit, the in-situ cathodoluminescence monitoring unit, and the high vacuum unit are all connected to the first growth chamber. The second in-situ monitoring unit is connected to the second growth chamber; the first data acquisition unit is connected to the first in-situ monitoring unit and the in-situ cathodoluminescence monitoring unit, and the second data acquisition unit is connected to the second in-situ monitoring unit; the edge computing node includes a physical constraint neural network module and a dual-chamber collaborative control module, the input terminals of the physical constraint neural network module and the dual-chamber collaborative control module are both connected to the first data acquisition unit and the second data acquisition unit; the output terminal of the physical constraint neural network module is connected to the vapor deposition execution unit and the dual-chamber collaborative control module respectively, and the output terminal of the dual-chamber collaborative control module is connected to the vapor deposition execution unit.

[0007] Furthermore, the edge computing node also includes a physical consistency verification module, a lightweight finite element module, and an electromagnetic interference noise utilization module. The input end of the physical consistency verification module is connected to the physical constraint neural network module, the first data acquisition unit, and the second data acquisition unit. The output end of the physical consistency verification module is connected to the physical constraint neural network module and the vapor deposition execution unit, respectively. The lightweight finite element module is connected to the physical consistency verification module. The input end of the electromagnetic interference noise utilization module is connected to the first growth chamber and the second growth chamber, and the output end is connected to the physical constraint neural network module and the physical consistency verification module.

[0008] Furthermore, the in-situ cathodoluminescence monitoring unit includes a high-temperature resistant electron optical column, a differential vacuum electron gun, and a spectrometer. The differential vacuum electron gun includes a multi-stage vacuum transition chamber and a microchannel array module connected to the multi-stage vacuum transition chamber. The microchannel array module is disposed inside the first growth chamber. The multi-stage vacuum transition chamber is connected between the microchannel array module and the high-temperature resistant electron optical column. The high-temperature resistant electron optical column is connected to the spectrometer, and the output end of the spectrometer is connected to a first data acquisition unit.

[0009] Furthermore, the first in-situ monitoring unit includes a first optical monitoring component, a first thermal monitoring component, and a first chemical monitoring component; the monitoring ends of the first optical monitoring component, the first thermal monitoring component, and the first chemical monitoring component are respectively disposed in or connected to the first growth chamber, and their output ends are all connected to the first data acquisition device; the second in-situ monitoring unit includes a second optical monitoring component, a second thermal monitoring component, and a second chemical monitoring component; the monitoring ends of the second optical monitoring component, the second thermal monitoring component, and the second chemical monitoring component are respectively disposed in or connected to the second growth chamber, and their output ends are all connected to the second data acquisition device; The sample transfer assembly includes a sample delivery chamber and a central transfer chamber. A magnetic levitation manipulator is installed in the central transfer chamber. The sample delivery chamber, the first growth chamber, and the second growth chamber are all connected to the magnetic levitation manipulator in the central transfer chamber.

[0010] Secondly, the present invention provides a physical constraint optimization method based on in-situ monitoring of semiconductor material defects, applied to the physical constraint optimization device based on in-situ monitoring of semiconductor material defects described in any of the above claims, comprising the following steps: S1: Epitaxial growth is performed in the first growth chamber, and dislocation feature vectors and process parameters are collected in real time through the in-situ cathode fluorescence monitoring unit and the first in-situ monitoring unit. S2: Combine the actual dislocation density obtained by the dislocation detection to construct a prediction training set containing the dislocation feature vector and the actual dislocation density, and construct an optimization training set containing the target dislocation density, the predicted dislocation density, the initial process parameters and the optimized process parameters. S3: Based on the physical constraint neural network module, combined with the prediction training set and the optimization training set, train the dislocation density prediction model and the process parameter optimization model; during the model training process, use the residual terms of the semiconductor thermodynamic control equation and the residual terms of the mechanical equilibrium equation as the physical constraint embedding loss function to train the physical constraint neural network model. S4: Input the dislocation feature vector and process parameters collected in real time into the trained physical constraint neural network model, and dynamically adjust the weight of the physical constraint according to the fluctuation characteristics of the gas pressure in the first growth chamber within the preset process coordination time window. When the gas pressure fluctuation exceeds the disturbance judgment standard, increase the weight to enhance the constraint strength of physical laws on the neural network, and output the preliminary predicted dislocation density and optimized process parameters. S5: The second growth chamber is started by controlling the dual-chamber collaborative control module. Its start time is delayed by the preset process collaborative time window of the first growth chamber, and batch epitaxial growth is performed in the second growth chamber based on the optimized process parameters. S6: During the growth process in the second growth chamber, the current growth status data of the second growth chamber is collected in real time and compared with the optimized process parameters and corresponding dislocation feature vectors of the first growth chamber; when the difference between the two exceeds the first preset threshold, a closed-loop adjustment mechanism is triggered to compensate the process parameters of the second growth chamber in real time.

[0011] Furthermore, the dislocation feature vector in S1 is composed of the following feature parameters as components, including: dark spot density, YLB / YLO intensity ratio, local emission peak half width at half maximum, dark spot size distribution entropy, emission oscillation ring radius, spatial correlation function decay coefficient, and temperature quenching activation energy; the in-situ cathodoluminescence monitoring unit is configured to trigger a scan to acquire cathodoluminescence data containing the above dislocation feature vector after the pressure of the first growth chamber is reduced to a preset detection pressure range during the growth interval.

[0012] Furthermore, the physical constraint neural network model in S3 includes a cathodoluminescence image encoder based on a convolutional neural network, a process timing processor based on a recurrent neural network, a multimodal fusion layer based on an attention mechanism, and a regression output layer with physical constraints; the physical constraints are implemented by embedding the weighted residual terms of the thermodynamic control equation and the mechanical equation as regularization terms into the loss function.

[0013] Furthermore, the output of the preliminary predicted dislocation density and optimized process parameters in S4 also includes the following parallel verification steps: First verification path: The physical consistency verification module integrates an independent physical model. The theoretical dislocation density calculated by the independent physical model is compared with the preliminary predicted dislocation density. When the relative error between the two exceeds the second preset threshold, the lightweight finite element model is triggered to make corrections and the corrected dislocation density value is output as the basis for process adjustment. The second verification path is as follows: During the calculation of the independent physical model, the defect confidence level obtained by the independent physical model based on the multi-physics field consistency assessment is obtained simultaneously. When the defect confidence level is lower than the third preset threshold, an adversarial perturbation that conforms to the symmetry direction of the intrinsic slip system of the target material is injected into the input of the physical constraint neural network model.

[0014] Furthermore, the closed-loop adjustment mechanism in S6 specifically includes: The edge computing node determines whether the dislocation density is abnormally increasing by a first preset threshold. If so, it issues control commands through the real-time industrial communication protocol deployed between the edge computing node and the vapor deposition execution unit. The control commands include temperature setpoint adjustment or gas flow rate setpoint adjustment. The temperature adjustment adopts a feedforward-feedback composite control strategy, and the gas flow rate adjustment adopts an adaptive smooth transition curve to avoid process step disturbances.

[0015] Thirdly, the present invention also provides a physical constraint optimization system based on in-situ monitoring of defects in semiconductor materials, comprising: In-situ data acquisition module: used for epitaxial growth in the first growth chamber, and to acquire dislocation feature vectors and process parameters in real time through the in-situ cathode fluorescence monitoring unit and the first in-situ monitoring unit; Training set construction module: used to combine the actual dislocation density obtained by dislocation detection to construct a prediction training set containing the dislocation feature vector and the actual dislocation density, and to construct an optimization training set containing the target dislocation density, the predicted dislocation density, the initial process parameters and the optimized process parameters; Model building and training module: used to train the dislocation density prediction model and the process parameter optimization model based on the physical constraint neural network module, combined with the prediction training set and the optimization training set; during the model training process, the residual terms of the semiconductor thermodynamic control equation and the residual terms of the mechanical equilibrium equation are used as physical constraint embedding loss functions to train the physical constraint neural network model. Online prediction and optimization module: It is used to input the dislocation feature vector and process parameters collected in real time into the trained physical constraint neural network model, dynamically map and adjust the weight of the physical constraint according to the fluctuation characteristics of the gas pressure in the first growth chamber within the preset process coordination time window, and increase the weight when the gas pressure fluctuation exceeds the disturbance judgment standard to enhance the constraint strength of physical laws on the neural network, and output the preliminary predicted dislocation density and optimized process parameters. Dual-chamber collaborative execution module: The dual-chamber collaborative control module controls the start-up of the second growth chamber, the start-up time of which lags behind the preset process collaboration time window of the first growth chamber, and performs batch epitaxial growth in the second growth chamber based on the optimized process parameters; Closed-loop feedback iteration module: During the growth process of the second growth chamber, it collects the current growth status data of the second growth chamber in real time and compares it with the optimized process parameters and corresponding dislocation feature vectors of the first growth chamber; when the difference between the two exceeds the first preset threshold, it triggers the closed-loop adjustment mechanism to compensate the process parameters of the second growth chamber in real time.

[0016] Compared with the prior art, the present invention has the following beneficial effects: This invention utilizes a first growth chamber dedicated to in-situ CL feature acquisition and process exploration, while a second growth chamber employs a time-lag strategy for mass production. A physical constraint neural network embedded in the edge computing node embeds the residuals of semiconductor thermodynamics and mechanics equations into the data-driven model, ensuring that small-sample predictions conform to physical mechanisms. A physical consistency verification module compares the predicted dislocation density with the theoretical values ​​of an independent physical model in real time, triggering anti-disturbance injection or lightweight finite element correction when the deviation exceeds a threshold. A dual-chamber collaborative control module compares the current growth state of the second growth chamber with historical data from the first growth chamber in real time, issuing dynamic compensation commands to the vapor deposition execution unit when the difference exceeds a preset dynamic threshold. This scheme achieves deep coupling between in-situ defect perception, physical constraint reasoning, and millisecond-level closed-loop process control, significantly improving the epitaxial layer crystal quality and the stability of dual-chamber collaborative production. Attached Figure Description

[0017] Figure 1 A structural diagram of a physical constraint optimization device based on in-situ monitoring of semiconductor material defects provided by the present invention; Figure 2 The structural diagram of the edge computing node provided by this invention; Figure 3 This is a structural diagram of the in-situ cathodic fluorescence monitoring unit provided by the present invention; Figure 4 A flowchart of a physical constraint optimization method based on in-situ monitoring of defects in semiconductor materials provided by the present invention; Figure 5 The flowchart of a physical constraint optimization system based on in-situ monitoring of defects in semiconductor materials is provided for this invention.

[0018] Figure 6 This is a cathode fluorescence micrograph of gallium nitride single-crystal dislocations provided by the present invention.

[0019] Wherein: 1-First growth chamber; 2-Second growth chamber; 3-First in-situ monitoring unit; 4-In-situ cathodoluminescence monitoring unit; 5-Second in-situ monitoring unit; 6-High vacuum unit; 7-Sample delivery chamber; 8-Central transmission chamber; 9-First data acquisition unit; 10-Second data acquisition unit; 11-Edge computing node; 12-Vacuum deposition execution unit; 13-Physically constrained neural network module; 14-Dual-chamber collaborative control module; 15-Physical consistency verification module; 16-Lightweight finite element module; 17-Electromagnetic interference noise utilization module; 18-High temperature resistant electron optical column; 19-Differential vacuum electron gun; 20-Spectrometer; 21-Multi-stage vacuum transition chamber; 22-Microchannel array module. Detailed Implementation

[0020] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0021] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0022] The present invention will now be described in further detail with reference to the accompanying drawings: Example 1 See Figure 1-3 As shown: This embodiment discloses a physical constraint optimization device based on in-situ monitoring of semiconductor material defects, including: a chemical vapor deposition system, the chemical vapor deposition system including a first growth chamber 1, a second growth chamber 2, a sample transfer component, a first in-situ monitoring unit 3, a second in-situ monitoring unit 5, a set of in-situ cathodoluminescence monitoring units 4, a high vacuum unit 6, a first data acquisition unit 9, a second data acquisition unit 10, an edge computing node 11, and a vapor deposition execution unit 12.

[0023] The sample transfer assembly includes a sample delivery chamber 7 and a central transfer chamber 8. A magnetic levitation manipulator is installed in the central transfer chamber 8. The sample delivery chamber 7, the first growth chamber 1, and the second growth chamber 2 are all connected to the magnetic levitation manipulator in the central transfer chamber 8. The first in-situ monitoring unit 3, a set of in-situ cathodoluminescence monitoring units 4, and the high vacuum unit 6 are all connected to the first growth chamber 1, and the second in-situ monitoring unit 5 is connected to the second growth chamber 2. Graphite disks are installed inside both the first growth chamber 1 and the second growth chamber 2. The graphite disks in the first growth chamber 1 are used to place semiconductor materials, and the second growth chamber 2 is equipped with multiple graphite disks for batch sample growth.

[0024] In this embodiment, the sample delivery chamber 7, the first growth chamber 1, and the second growth chamber 2 are all arranged around the central transmission chamber 8 through a vacuum-sealed interface. When using this device, the semiconductor material substrate placed in the sample delivery chamber 7 is transferred to the graphite disk in the first growth chamber 1 for epitaxial growth by a magnetic levitation manipulator. At the same time, three other substrates of the same material are sequentially loaded into the graphite disk in the second growth chamber 2 for batch growth. The entire process is carried out in a vacuum environment with no contact in the sample transfer.

[0025] The in-situ cathodoluminescence monitoring unit 4 is used to monitor dislocation dynamics and is integrated into the top window of the first growth chamber 1. The in-situ cathodoluminescence monitoring unit 4 includes a high-temperature resistant electron optical column 18, a differential vacuum electron gun 19, and a spectrometer 20. The differential vacuum electron gun 19 includes a multi-stage vacuum transition chamber 21 and a microchannel array module 22. The first growth chamber 1, the microchannel array module 22, the multi-stage vacuum transition chamber 21, the high-temperature resistant electron optical column 18, the spectrometer 20, and the first data acquisition unit 9 are connected in sequence to form a gradient vacuum between the high-pressure environment of the first growth chamber 1 and the low-pressure environment of the high-temperature resistant electron optical column 18.

[0026] The multi-stage vacuum transition chamber 21 achieves step-by-step pressure reduction through a series connection of multiple chambers, ensuring the differential vacuum electron gun 19 operates normally under varying high pressure differentials. The number of stages in the multi-stage vacuum transition chamber 21 of the differential vacuum electron gun 19 is determined based on the following physical factors: from the pressure of the first growth chamber (50-200 Torr) to the detection pressure (10... - ³~10 -6 The pressure difference range (Pa), the flow conductance C of the microchannel array module 22 (pore size <1μm) limiting the gas throughput of each stage, and the configuration of an independent vacuum pump for each stage chamber; the number of stages of the multi-stage vacuum transition chamber 21 is usually 2 to 5 stages, and 3 to 4 stages are preferred in this embodiment.

[0027] The differential vacuum electron gun 19 achieves directional electron beam emission through the microchannel array module 22, maintaining a 10-degree electron beam within the microchannel. -6 Pa to 10 - ¹ Pa gradient vacuum environment; the inner surface of the microchannel is coated with a Pt(70%)-Ir(30%) alloy layer, which is designed to resist the erosion of MO source gas in a gradient vacuum environment.

[0028] The high-temperature resistant electron optical column 18 uses a silicon carbide-based composite material shell or a sapphire shell, and integrates a micro parabolic reflector and fiber bundle inside. The reflector surface is coated with Pt / Ir alloy to resist MO source gas corrosion. The surface coating is a Pt(70%)-Ir(30%) alloy with a surface roughness Ra≤0.05μm.

[0029] The in-situ cathodoluminescence monitoring unit 4 also includes a sample cooling stage for cooling the sample in the first growth chamber.

[0030] The first in-situ monitoring unit 3 is used to collect process parameters in the first growth chamber 1 in real time. The first in-situ monitoring unit 3 includes a first optical monitoring component, a first thermal monitoring component and a first chemical monitoring component; the first optical monitoring component includes a first laser interferometer and a first in-situ elliptic spectrometer.

[0031] The first laser interferometer is integrated into the side wall window of the first growth chamber 1 at an angle of 30° to 60°. In this embodiment, a 45° angle is preferred to avoid interference from vapor deposition. The wavelength of the first laser interferometer is 632.8 nm, which is used to measure the thickness of the epitaxial layer in real time.

[0032] The first in-situ elliptic spectrometer is set in the top window of the first growth chamber 1; the wavelength is 250–1000 nm, and it is used for the analysis of the composition of the multilayer structure. In this embodiment, the "multilayer structure" refers to the semiconductor multilayer thin film structure formed during the MOCVD epitaxial growth process.

[0033] The first thermal monitoring component includes a first infrared pyrometer, integrated into the bottom window of the first growth chamber 1; with a response wavelength of 1.5–2.2 μm and a temperature measurement accuracy of ±1℃, it is used for real-time, non-contact measurement of the actual temperature of the substrate / epitaxy wafer surface inside the first growth chamber 1.

[0034] The first chemical monitoring component includes a first quadrupole mass spectrometer, which is connected to the first growth chamber 1 via a bypass of the exhaust gas pipeline. It has a mass resolution of M / ΔM = 300 and a sampling frequency of 10 Hz, and is used for real-time, online quantitative analysis of MOCVD exhaust gas components to monitor the gas-phase chemical reaction process and process anomalies. Specifically, a sampling port is provided on the main exhaust gas pipeline of the first growth chamber 1. A branch is led out from the sampling port and connected to the inlet of the first quadrupole mass spectrometer via a stainless steel capillary or quartz capillary to split a portion of the exhaust gas for real-time component analysis. The analyzed gas is returned to the main exhaust gas pipeline or discharged separately. The capillary acts as a flow limiter and pressure transition valve; the online mass spectrometer injection pressure is typically in the range of 100 mbar to 2 bar. A differential vacuum transition chamber can be configured inside the quadrupole mass spectrometer to ensure the high vacuum environment required for its operation.

[0035] The second in-situ monitoring unit 5 includes a second optical monitoring component, a second thermal monitoring component, and a second chemical monitoring component. The structure and connection relationship of the second optical monitoring component, the second thermal monitoring component, and the second chemical monitoring component are respectively referenced to the first optical monitoring component, the first thermal monitoring component, and the first chemical monitoring component. The difference is that the second optical monitoring component is integrated into the side wall window of the second growth chamber 2, the second thermal monitoring component is integrated into the bottom window of the second growth chamber 2, and the second chemical monitoring component is connected to the second growth chamber 2 through the exhaust gas pipeline bypass. It is used to collect process parameters in the second growth chamber 2 in real time. The process parameters include the growth status data of the batch samples.

[0036] The first data acquisition unit 9 is connected to the first in-situ monitoring unit 3 and the in-situ cathodoluminescence monitoring unit 4, and the second data acquisition unit 10 is connected to the second in-situ monitoring unit 5.

[0037] The high vacuum unit 6 uses a turbomolecular pump with a pumping speed ≥2000 L / s.

[0038] The edge computing node 11 includes a physical constraint neural network module 13 and a dual-chamber collaborative control module 14. The input terminals of the physical constraint neural network module 13 and the dual-chamber collaborative control module 14 are both connected to the first data acquisition unit 9 and the second data acquisition unit 10. The output terminal of the physical constraint neural network module 13 is connected to the vapor deposition execution unit 12 and the dual-chamber collaborative control module 14, respectively. The output terminal of the dual-chamber collaborative control module 14 is connected to the vapor deposition execution unit 12.

[0039] The edge computing node 11 also includes a physical consistency verification module 15, a lightweight finite element module 16, and an electromagnetic interference noise utilization module 17. The input of the physical consistency verification module 15 is connected to the physical constraint neural network module 13, the first data acquisition unit 9, and the second data acquisition unit 10. The output of the physical consistency verification module 15 is connected to the physical constraint neural network module 13 and the vapor deposition execution unit 12, respectively. The lightweight finite element module 16 is connected to the physical consistency verification module 15. The lightweight finite element module 16 feeds back the corrected data to the physical consistency verification module 15, and the corrected data is output to the vapor deposition execution unit 12 through the physical consistency verification module 15. The input of the electromagnetic interference noise utilization module 17 is connected to the first growth chamber 1 and the second growth chamber 2, and the output is connected to the physical constraint neural network module 13 and the physical consistency verification module 15.

[0040] Example 2 Combination Figure 4 As shown, this embodiment provides a physical constraint optimization method based on in-situ monitoring of defects in semiconductor materials, including the following steps: S1: Epitaxial growth is carried out in the first growth chamber, and dislocation feature vectors and process parameters are collected in real time through the in-situ cathode fluorescence monitoring unit and the first in-situ monitoring unit.

[0041] The dislocation feature vector in S1 includes dark spot density, YLB / YLO intensity ratio, local emission peak half width at half maximum, dark spot size distribution entropy, emission oscillation ring radius, spatial correlation function decay coefficient, and temperature quenching activation energy. The in-situ cathodoluminescence monitoring unit is configured to trigger a scan to acquire cathodoluminescence data containing the above dislocation feature vectors after the pressure in the first growth chamber is reduced to a preset detection pressure range during the growth interval.

[0042] The relationship between the YLB / YLO intensity ratio and the dislocation density D is as follows:

[0043] In this embodiment, the coefficient is 1.2 × 10⁻⁶. 7 The coefficients 3.8 and 3.8 are applicable to GaN, and can be adjusted according to the material type.

[0044] The initial epitaxial growth in the first growth chamber includes: timing the activation of the high vacuum unit to reduce the pressure in the first growth chamber from 50-200 Torr to 10 Torr within 3-5 seconds. - ³–10 -6 Pa, and triggers the in-situ CL monitoring unit to acquire the dislocation feature vector.

[0045] S2: Combine the actual dislocation density obtained from the dislocation detection to construct a prediction training set containing dislocation feature vectors and actual dislocation density, and construct an optimization training set containing target dislocation density, predicted dislocation density, initial process parameters and optimized process parameters.

[0046] S3: Based on the physical constraint neural network module, the dislocation density prediction model and the process parameter optimization model are trained by combining the prediction training set and the optimization training set. During the model training process, the residual terms of the semiconductor thermodynamic control equation and the residual terms of the mechanical equilibrium equation are used as physical constraint embedding loss functions to train the physical constraint neural network model.

[0047] The physically constrained neural network model in S3 includes a cathodoluminescence (CL) image encoder based on a convolutional neural network, a process timing processor based on a recurrent neural network, a multimodal fusion layer based on an attention mechanism, and a regression output layer with physical constraints.

[0048] In this embodiment, the CL image encoder specifically adopts a ResNet-based architecture to extract the spatial distribution features of dislocations in the CL image; the process timing processor specifically adopts an LSTM-based architecture to capture the temporal evolution features of process parameters; the multimodal fusion layer specifically adopts a Transformer-based architecture to fuse CL features and process parameters; and the regression output layer with physical constraints is used to generate dislocation density predictions that satisfy semiconductor physical laws.

[0049] Furthermore, physical constraints are implemented by weighting the residual terms of the semiconductor thermodynamic control equations and the residual terms of the mechanical equilibrium equations and embedding them as regularization terms into the loss function to ensure that the model output conforms to the physical mechanism.

[0050] Furthermore, the edge computing node also includes an electromagnetic interference noise utilization module, which is used to collect the electromagnetic interference (EMI) spectrum of the first growth chamber and the second growth chamber, couple the EMI spectrum characteristics with the lattice distortion equation, generate supplementary constraint terms and input them into the physical constraint neural network model to enhance the physical reality of the residual terms of the mechanical equation.

[0051] The relationship between EMI spectral characteristics and lattice distortion ε is as follows:

[0052] Where ε is the lattice distortion degree, characterizing the degree of lattice distortion of the material caused by electromagnetic interference, and S EMI ( f ) represents the EMI power spectral density. f1 and f2 The characteristic frequency range is determined based on the electromagnetic environment characteristics of the MOCVD equipment. β For GaN materials, the material coefficient is... β Preferably 1.2×10 -8 m² / V².

[0053] The supplementary constraint term, as a strengthening factor of the residual term of the mechanical equation, transforms electromagnetic environment information into quantifiable lattice distortion degree and serves as a supplement or strengthening term of the residual constraint of the mechanical equation in the physical constraint neural network model. When embedded in the physical constraint loss function, the neural network model can perceive the influence of the cavity electromagnetic environment on the micro-stress state of the material, thereby improving the prediction robustness of the model under complex electromagnetic interference scenarios.

[0054] In this step, the expression for the physical constraints is:

[0055] in, This indicates the severity of the penalty imposed by the physical constraint. κ Thermal conductivity, T For temperature,Q For internal heat source items, ρ For density, Cp For specific heat capacity, σ For stress tensor, t For time, f Let be the body force vector, where the first term is the residual of the thermodynamic equation and the second term is the residual of the mechanical equation.

[0056] S4: Input the real-time collected dislocation feature vector and process parameters into the trained physical constraint neural network model. Dynamically map and adjust the weight of physical constraints according to the fluctuation characteristics of the gas pressure in the first growth chamber within the preset process coordination time window. When the gas pressure fluctuation exceeds the disturbance judgment standard, increase the weight to enhance the constraint strength of physical laws on the neural network. Output the preliminary predicted dislocation density and optimized process parameters.

[0057] In this step, the weighting coefficient λ of the physical constraints is based on the standard deviation σ of the gas pressure in the first growth chamber within a preset process coordination time window. p Dynamic adjustment, and satisfying λ and σ p The relationship is negatively correlated. The preset process coordination time window is preferably 10 seconds, and the dynamic adjustment of the weight coefficient λ of the physical constraint satisfies the following relationship: λ = 0.75 + 0.04×(10 - σ p When the process disturbance is large (i.e., σ), p When the pressure is >5 Pa, λ automatically increases to above 0.95 to enhance the constraint strength of physical constraints on the neural network, thereby effectively suppressing prediction distortion caused by air pressure fluctuations.

[0058] When outputting the preliminary predicted dislocation density and optimized process parameters in S4, the following parallel verification steps are also included: First verification path: The physical consistency verification module integrates an independent physical model. The theoretical dislocation density calculated by the independent physical model is compared with the preliminary predicted dislocation density. When the relative error between the two exceeds the second preset threshold, the lightweight finite element model is triggered to make corrections and outputs the corrected dislocation density value as the basis for process adjustment. The second preset threshold is preferably 15%. After the lightweight finite element model is triggered, the thermo-mechanical coupled finite element model obtained by order reduction through intrinsic orthogonal decomposition (POD) performs fast finite element simulation based on the current sensor data. The calculation delay is <0.5ms. The corrected dislocation density value and the corresponding temperature / stress field distribution are generated. The corrected dislocation density value will be used as the basis for the edge computing node to output control commands to the vapor deposition execution unit, and used to adjust process parameters such as temperature and gas flow rate in real time.

[0059] The second verification path is to simultaneously acquire the defect confidence level obtained by the independent physical model based on the multi-physics field consistency assessment during the calculation process. When the defect confidence level is lower than the third preset threshold, an adversarial perturbation that conforms to the symmetry direction of the intrinsic slip system of the target material is injected into the input of the physical constraint neural network model.

[0060] The third preset threshold is preferably 0.5%; when this threshold is triggered, the amplitude of the anti-disturbance is set to |b| / 10, where b is the Burgers vector of the target material; the injection direction of the anti-disturbance is determined according to the lattice symmetry of the target material to match the symmetry direction of the intrinsic slip system of the lattice.

[0061] Specifically, for hexagonal GaN materials, the perturbation directions are injected along the [1-100] and [11-20] crystal orientations. This sub-angstrom level perturbation aims to utilize the intrinsic lattice symmetry of the material as a priori constraint to force the neural network to learn feature representations that conform to crystallographic laws when the confidence of the physical model is insufficient, thereby improving the feature robustness and prediction stability of the model under boundary conditions.

[0062] S5: The second growth chamber is started by controlling the dual-chamber collaborative control module. Its start-up time is delayed by the preset process collaborative time window of the first growth chamber, and batch epitaxial growth is performed in the second growth chamber based on the optimized process parameters.

[0063] In this step, the start-up time of the second growth chamber lags behind the preset process coordination time window of the first growth chamber by Δt, where 5min≤Δt≤30min.

[0064] S6: During the growth process in the second growth chamber, the current growth status data of the second growth chamber is collected in real time and compared with the optimized process parameters and corresponding dislocation feature vectors of the first growth chamber. When the difference between the two exceeds the first preset threshold, a closed-loop adjustment mechanism is triggered to compensate the process parameters of the second growth chamber in real time.

[0065] The closed-loop adjustment mechanism in S6 specifically includes: The edge computing node determines whether the dislocation density is abnormally increasing by a first preset threshold. If so, it issues control commands through a real-time industrial communication protocol deployed between the edge computing node and the vapor deposition execution unit. The real-time industrial communication protocol is preferably the OPC UA protocol.

[0066] In this step, the first preset threshold is a dynamic value, dynamically set according to the material properties and the target dislocation density. For example, for GaN epitaxial layers, the threshold is related to the target dislocation density. D target The relationship is: The dynamic compensation adopts an incremental parameter adjustment strategy as shown in the following formula:

[0067] in, For the adjusted parameters, Let α be the target parameter and α be the adaptive coefficient. For measured features, The predictive features are determined by the adaptive coefficient α, which is dynamically adjusted according to the growth stage: α=0.3 in the initial stage of extension, α=0.1 in the steady-state growth stage, and α=0.25 in the growth termination stage.

[0068] The control commands include adjustments to the temperature setpoint or the gas flow rate setpoint. The specific adjustment step depends on the actual situation. If temperature adjustment is involved, a feedforward-feedback composite control strategy is employed. The feedforward part predicts the temperature field distribution based on the thermodynamic control equation, while the feedback part compensates in real time for measured temperature deviations to ensure the temperature field conforms to semiconductor physics. If gas flow rate adjustment is involved, an S-curve smooth transition is used to avoid process step disturbances. The transition time t... s Satisfying the relation:

[0069] in, t s The time required for gas flow rate adjustment k Δ is a material constant. F The change in flow rate is represented by F0, where F0 is the initial flow rate. d This represents the characteristic dimension of the reaction chamber. t s The calculation formula defines the transition time required for the gas flow rate to adjust from one steady-state value to another. This adaptive transition time enables smooth adjustment of the gas flow rate, avoiding process disturbances caused by abrupt changes and ensuring the stability of epitaxial growth. The formula comprehensively considers the magnitude of flow rate changes, chamber size, and material properties, and has a clear physical basis.

[0070] Example 3 This embodiment uses the growth of a GaN epitaxial layer on a 6-inch SiC substrate as an example, with a target dislocation density of ≤5×10⁻⁶. 5 cm - ², suitable for the manufacture of 5G RF power amplifiers, including the following steps: S1: A 6-inch 4H-SiC (0001) substrate is placed in the sample delivery chamber and vacuum baked (300℃, 1 h) to remove surface adsorbates. Then, it is transferred to the graphite disk in the first growth chamber by a magnetic levitation robot in the central transfer chamber. At the same time, three other SiC substrates of the same specifications are sequentially loaded into the four graphite disks in the second growth chamber. The MOCVD program was started, and GaN epitaxial growth was performed in the first growth chamber. The process parameters were as follows: substrate temperature 1050℃, reaction chamber pressure 200 Torr, TMGa flow rate 20 μmol / min, NH3 flow rate 8000 sccm, V / III ratio 400. During growth, in-situ CL detection is performed every 10 minutes, specifically: In this embodiment, a high vacuum pump is used to reduce the chamber pressure from 200 Torr to 8 × 10⁻⁶ within 3.2 seconds. -4 Pa; A differential vacuum electron gun emits a 2.5 keV electron beam to bombard the sample surface; The CL signal is collected by a mirror, guided by an optical fiber bundle to a spectrometer, and then generated as shown in the image. Figure 6 The full-spectrum cathodoluminescence image shown depicts dark areas corresponding to dislocation defects in the crystal and bright areas representing high-quality epitaxial regions. The cathodoluminescence image was analyzed using edge computing nodes, and dislocation feature vectors composed of multi-dimensional components were extracted. In this embodiment, the specific quantization result of these multi-dimensional components is: dark spot density 1.2 × 10⁻⁶. 6 cm - ², YLB / YLO intensity ratio 0.38, local emission peak full width at half maximum (FWHM) 85 meV, dark spot size distribution entropy 1.72, emission oscillation ring radius 2.1 μm, spatial correlation function attenuation coefficient 0.043 μm. - ¹ and the temperature quenching activation energy is 180 meV. Subsequently, this feature vector is used as the standardized input of the physical constraint neural network model, and after being regularized by the physical constraint loss function, the predicted dislocation density is output.

[0071] S2: Combine the actual dislocation density obtained by high-precision detection of dislocations to construct a prediction training set containing dislocation feature vectors and actual dislocation densities, as well as an optimization training set containing target dislocation density, predicted dislocation density, initial process parameters, and optimized process parameters.

[0072] S3: A hybrid neural network model that takes the above feature vectors as input and deploys them on edge computing nodes. The model structure includes: CL image encoder (ResNet-18, input 256×256 CL image); Process timing processor (2-layer LSTM, input 10-dimensional process parameter sequence). Multimodal fusion layer (Transformer encoder, fusing CL and process features); Regression output layer (with physical constraints).

[0073] The weighting coefficient λ of the physical constraints in the loss function is dynamically adjusted based on the standard deviation of the chamber pressure over 10 seconds. In this embodiment, σp =6.2 Pa, therefore λ=0.95.

[0074] The model outputs a predicted dislocation density of 8.7 × 10⁻⁶. 5 cm - ², higher than the target value by 5 × 10 5 cm - ². The process optimization model was then initiated, with [target dislocation density = 5 × 10⁻⁶] as the target. 5 Predicted dislocation density = 8.7 × 10⁻⁶ 5 [Initial process parameters] are the inputs, and the outputs are the optimized parameters: Among these improvements, the substrate temperature was increased to 1065℃, the V / III ratio was increased to 450, and the base rotation speed was increased from 80 rpm to 100 rpm.

[0075] S4: Input the real-time collected dislocation feature vectors and process parameters into the trained physical constraint neural network model. Dynamically map and adjust the weight of physical constraints according to the fluctuation characteristics of the gas pressure in the first growth chamber within a 10-second time window. When the gas pressure fluctuation exceeds the disturbance judgment standard, increase the weight to enhance the constraint strength of physical laws on the neural network. Output the preliminary predicted dislocation density and optimized process parameters.

[0076] S5: The second growth chamber begins growth Δt=15 minutes after the first growth chamber is started, and the above optimized parameters are directly used for batch epitaxial growth.

[0077] S6: During the growth process, the second in-situ monitoring unit collects data from the second growth chamber in real time and compares it with the historical data of the first chamber.

[0078] When a film thickness deviation >2% is detected, closed-loop adjustment is triggered: the TMGa mass flow controller is adjusted via the OPC UA protocol; temperature is controlled using a feedforward-feedback composite control, with the feedforward term predicted by thermodynamic equations and the feedback term compensating for measured deviations; gas flow rate changes are transitioned using an S-shaped curve with a transition time t. s =1.8 s (satisfying t) s = k·ΔF / F0·d², k=0.025).

[0079] Effect Verification: After growth was completed, the 6-inch GaN / SiC epitaxial wafer in the second growth chamber was subjected to in-situ characterization tests: Dislocation CL mapping test: The 6-inch wafer is scanned in a grid pattern. Five test points are selected for each die, including the center and four corners. The scanning area of ​​each test point is 100×100 μm², and a total of ≥500 sampling points are collected for the entire wafer. Based on the dark spot count in the CL image and the YLB / YLO intensity ratio conversion, the dislocation density of each test point is calculated, and the average value of the entire wafer is taken as the final dislocation density.

[0080] Test results: The average dislocation density is 4.3 × 10⁻⁶. 5 cm - ², satisfying the target value ≤ 5 × 10 5 cm - The requirements of ².

[0081] TEM cross-section calibration: Sample preparation: Cross-sectional TEM samples were prepared by cutting along the [11-20] crystal orientation of the GaN epitaxial layer. After mechanical thinning to approximately 10 μm, Ar ion thinning was performed using a Gatan PIPS II ion thinner under liquid nitrogen cooling. + Ions are thinned to electronic transparency.

[0082] Dislocation type and density calibration; Calibration was performed using weak-beam dark-field imaging. The distributions of screw dislocations, edge dislocations, and mixed dislocations were observed under dual-beam conditions of g =

[0002] and g = [11-20]. Ten fields of view (each with an area of ​​approximately 2 × 2 μm²) were randomly selected, and the number of dislocations of each type was manually counted. The dislocation density was calculated and cross-validated with the CL mapping results. The TEM-calibrated dislocation density was 4.1 × 10⁻⁶. 5 cm - ², with a relative deviation of <5% from the CL mapping result.

[0083] Final technical effect: Average dislocation density 4.3 × 10 5 cm - ² (Meets the target); 6-inch wafer uniformity (CV) 3.1%; prediction error 5.2% (compared to 38.7% for pure AI models); physical constraints reduce process window width by ±3.8% (compared to ±1.2% for traditional methods); detection latency reduced from 2.1s to 0.35s.

[0084] This embodiment demonstrates that the present invention achieves high-precision and high-efficiency fabrication of low-dislocation GaN epitaxial layers by combining a physical constraint neural network with dual-chamber timing coordination.

[0085] Example 4 Combination Figure 5 As shown, this embodiment provides a physical constraint optimization system based on in-situ monitoring of defects in semiconductor materials, including: In-situ data acquisition module: Epitaxial growth is carried out in the first growth chamber, and dislocation feature vectors and process parameters are acquired in real time through the in-situ cathodoluminescence monitoring unit and the first in-situ monitoring unit; Training set construction module: Combine the actual dislocation density obtained by dislocation detection to construct a prediction training set containing dislocation feature vectors and actual dislocation density, and construct an optimization training set containing target dislocation density, predicted dislocation density, initial process parameters and optimized process parameters. Model building and training module: Based on the physical constraint neural network module, combined with the prediction training set and the optimization training set, the dislocation density prediction model and the process parameter optimization model are trained; during the model training process, the residual terms of the semiconductor thermodynamic control equation and the residual terms of the mechanical equilibrium equation are used as physical constraint embedding loss functions to train the physical constraint neural network model. Online prediction and optimization module: Input the dislocation feature vectors and process parameters collected in real time into the trained physical constraint neural network model. Based on the fluctuation characteristics of the gas pressure in the first growth chamber within the preset process coordination time window, dynamically map and adjust the weight of the physical constraint. When the gas pressure fluctuation exceeds the disturbance judgment standard, increase the weight to enhance the constraint strength of the physical law on the neural network. Output the preliminary predicted dislocation density and optimized process parameters. Dual-chamber collaborative execution module: The dual-chamber collaborative control module controls the start-up of the second growth chamber, whose start-up time lags behind the preset process collaboration time window of the first growth chamber, and performs batch epitaxial growth in the second growth chamber based on the optimized process parameters; Closed-loop feedback iteration module: During the growth process in the second growth chamber, the current growth status data of the second growth chamber is collected in real time and compared with the optimized process parameters and corresponding dislocation feature vectors of the first growth chamber; when the difference between the two exceeds the first preset threshold, the closed-loop adjustment mechanism is triggered to compensate the process parameters of the second growth chamber in real time.

[0086] Example 5 In conjunction with the foregoing embodiments, this embodiment further provides a cloud optimization method for semiconductor epitaxial processes, applicable to collaborative optimization scenarios involving multiple geographically dispersed MOCVD equipment, specifically including the following steps: Step 1: Multi-source data acquisition and local preprocessing In-situ CL feature vectors (including dark spot density, YLB / YLO intensity ratio, and full width at half maximum of local emission peaks) and corresponding process parameters (substrate temperature, reaction chamber pressure, gas flow rate, etc.) are collected in real time from multiple geographically dispersed MOCVD equipment (such as epitaxial production lines located in different factories or foundries). After each equipment completes data standardization at its local edge computing node, only the model parameters or gradient information are uploaded to the cloud server, while the original process data is kept locally, ensuring the privacy of the enterprise's core data.

[0087] Step 2: Deployment of the Cloud-based Federated Learning Framework and Local Training A federated learning coordination framework is built on a cloud server, with each MOCVD device connected as a federated learning client. Each device independently trains its own physical constraint neural network model based on local historical batch data and real-time acquired data. During local training, each device uses the physical constraint embedding mechanism of Example 2, that is, the residual terms of the semiconductor thermodynamic control equation and the residual terms of the mechanical equilibrium equation are weighted and then embedded as regularization terms into the loss function to ensure that each local model has basic physical rationality.

[0088] Step 3: Model Aggregation and Introduction of Global Physical Constraints During the cloud-based model aggregation phase, semiconductor physics equations are introduced as global constraints to correct biases in purely data-driven models. The specific constraint expressions are as follows:

[0089] in, N The total number of MOCVD devices participating in federated learning; w i For the first i The aggregation weight of each device can be dynamically allocated based on the sample size, data quality score, or process maturity of that device. For the first i The squared value of the physical equation residuals (i.e., the sum of the thermodynamic control equation residuals and the mechanical equilibrium equation residuals) output by the local model of each device, after taking the L2 norm, is used to quantify the degree to which the prediction results of the device model deviate from the physical laws of semiconductors. The convergence threshold represents the maximum acceptable average physical deviation in the cloud.

[0090] Furthermore, to reduce the deployment load of physically constrained neural network models on edge computing nodes and the communication overhead during federated learning, after global model aggregation is completed in the cloud, lightweight processing is performed on the models. Lightweighting is achieved through parameter pruning and knowledge distillation, specifically including: (1) Parametric pruning based on physical constraint sensitivity: Calculate the sensitivity S of each parameter of the neural network to physical constraints. j， Sensitivity is defined as:

[0091] in, S j Representing the j The sensitivity of each neural network parameter to physical constraints. θ j For the first j Each model parameter For the physical loss function with respect to parameters θ jThe rate of change; (2) Knowledge distillation including physical residual preservation terms: The high-performance global model aggregated in the cloud is used as the teacher model, and the pruned sparse model is used as the student model. During the distillation training process, a composite loss function is constructed, which includes a cross-entropy loss term and a physical residual preservation term.

[0092] The lightweight model is deployed to the edge computing nodes of geographically dispersed MOCVD equipment for operation. This significantly reduces inference latency and memory usage while ensuring the accuracy and physical rationality of dislocation density prediction, thus meeting the requirements for millisecond-level closed-loop control.

[0093] The cloud optimization method in this embodiment enables process knowledge sharing and model co-evolution across multiple devices in different regions. While ensuring data privacy for each manufacturing unit, it utilizes semiconductor physics equations as global constraints, effectively avoiding the common "rich data but physics-distorted" problem in federated learning. This significantly improves the model's generalization ability and prediction robustness under different device hardware differences and material systems. For newly built production lines or small-sample equipment, the globally validated model can be directly loaded, greatly shortening the process debugging cycle and reducing trial-and-error costs.

[0094] The above content is only for illustrating the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solution based on the technical concept proposed in this invention shall fall within the scope of protection of the claims of this invention.

Claims

1. A physical constraint optimization device based on in-situ monitoring of defects in semiconductor materials, characterized in that: include: A chemical vapor deposition (CVD) system includes a first growth chamber, a second growth chamber, a sample transfer assembly, a first in-situ monitoring unit, a second in-situ monitoring unit, an in-situ cathodoluminescence monitoring unit, a high vacuum unit, a first data acquisition unit, a second data acquisition unit, an edge computing node, and a CVD execution unit. The sample transfer assembly connects the first and second growth chambers. The first, second, and third in-situ monitoring units are all connected to the first growth chamber, and the second in-situ monitoring unit is connected to the second growth chamber. The first data acquisition unit connects to both the first and second in-situ monitoring units, and the second data acquisition unit connects to the second in-situ monitoring unit. The edge computing node includes a physically constrained neural network (PCN) module and a dual-chamber collaborative control module. The inputs of both the PCN module and the dual-chamber collaborative control module are connected to the first and second data acquisition units, respectively. The output of the PCN module is connected to both the CVD execution unit and the dual-chamber collaborative control module, and the output of the dual-chamber collaborative control module is connected to the CVD execution unit.

2. The physical constraint optimization device based on in-situ monitoring of semiconductor material defects according to claim 1, characterized in that: The edge computing node further includes a physical consistency verification module, a lightweight finite element module, and an electromagnetic interference noise utilization module. The input of the physical consistency verification module is connected to the physical constraint neural network module, the first data acquisition unit, and the second data acquisition unit. The output of the physical consistency verification module is connected to the physical constraint neural network module and the vapor deposition execution unit, respectively. The lightweight finite element module is connected to the physical consistency verification module. The input of the electromagnetic interference noise utilization module is connected to the first growth chamber and the second growth chamber, and the output is connected to the physical constraint neural network module and the physical consistency verification module.

3. The physical constraint optimization device based on in-situ monitoring of semiconductor material defects according to claim 1, characterized in that: The in-situ cathodoluminescence monitoring unit includes a high-temperature resistant electron optical column, a differential vacuum electron gun, and a spectrometer. The differential vacuum electron gun includes a multi-stage vacuum transition chamber and a microchannel array module connected to the multi-stage vacuum transition chamber. The microchannel array module is disposed inside the first growth chamber. The multi-stage vacuum transition chamber is connected between the microchannel array module and the high-temperature resistant electron optical column. The high-temperature resistant electron optical column is connected to the spectrometer, and the output of the spectrometer is connected to a first data acquisition unit.

4. The physical constraint optimization device based on in-situ monitoring of semiconductor material defects according to claim 1, characterized in that: The first in-situ monitoring unit includes a first optical monitoring component, a first thermal monitoring component, and a first chemical monitoring component; the monitoring ends of the first optical monitoring component, the first thermal monitoring component, and the first chemical monitoring component are respectively disposed in or connected to the first growth chamber, and the output ends of each component are connected to a first data acquisition device; the second in-situ monitoring unit includes a second optical monitoring component, a second thermal monitoring component, and a second chemical monitoring component; the monitoring ends of the second optical monitoring component, the second thermal monitoring component, and the second chemical monitoring component are respectively disposed in or connected to the second growth chamber, and the output ends of each component are connected to a second data acquisition device; The sample transfer assembly includes a sample delivery chamber and a central transfer chamber. A magnetic levitation manipulator is installed in the central transfer chamber. The sample delivery chamber, the first growth chamber, and the second growth chamber are all connected to the magnetic levitation manipulator in the central transfer chamber.

5. A physical constraint optimization method based on in-situ monitoring of defects in semiconductor materials, characterized in that: The physical constraint optimization device based on in-situ monitoring of semiconductor material defects according to any one of claims 1-4 includes the following steps: S1: Epitaxial growth is performed in the first growth chamber, and dislocation feature vectors and process parameters are collected in real time through the in-situ cathode fluorescence monitoring unit and the first in-situ monitoring unit. S2: Combine the actual dislocation density obtained from the dislocation detection to construct a prediction training set containing dislocation feature vectors and actual dislocation density, and construct an optimization training set containing target dislocation density, predicted dislocation density, initial process parameters and optimized process parameters. S3: Based on the physical constraint neural network module, combined with the prediction training set and the optimization training set, train the dislocation density prediction model and the process parameter optimization model; during the model training process, use the residual terms of the semiconductor thermodynamic control equation and the residual terms of the mechanical equilibrium equation as the physical constraint embedding loss function to train the physical constraint neural network model. S4: Input the dislocation feature vector and process parameters collected in real time into the trained physical constraint neural network model, and dynamically adjust the weight of the physical constraint according to the fluctuation characteristics of the gas pressure in the first growth chamber within the preset process coordination time window. When the gas pressure fluctuation exceeds the disturbance judgment standard, increase the weight to enhance the constraint strength of physical laws on the neural network, and output the preliminary predicted dislocation density and optimized process parameters. S5: The second growth chamber is started by controlling the dual-chamber collaborative control module. Its start time is delayed by the preset process collaborative time window of the first growth chamber, and batch epitaxial growth is performed in the second growth chamber based on the optimized process parameters. S6: During the growth process in the second growth chamber, the current growth status data of the second growth chamber is collected in real time and compared with the optimized process parameters and corresponding dislocation feature vectors of the first growth chamber; when the difference between the two exceeds the first preset threshold, a closed-loop adjustment mechanism is triggered to compensate the process parameters of the second growth chamber in real time.

6. The physical constraint optimization method based on in-situ monitoring of semiconductor material defects according to claim 5, characterized in that: The dislocation feature vector in S1 is composed of the following feature parameters as components, including: dark spot density, YLB / YLO intensity ratio, local emission peak half width at half maximum, dark spot size distribution entropy, emission oscillation ring radius, spatial correlation function decay coefficient, and temperature quenching activation energy; the in-situ cathodoluminescence monitoring unit is configured to trigger a scan to acquire cathodoluminescence data containing the above dislocation feature vector after the pressure of the first growth chamber is reduced to a preset detection pressure range during the growth interval.

7. The physical constraint optimization method based on in-situ monitoring of semiconductor material defects according to claim 5, characterized in that: The physical constraint neural network model in S3 includes a cathodoluminescence image encoder based on a convolutional neural network, a process timing processor based on a recurrent neural network, a multimodal fusion layer based on an attention mechanism, and a regression output layer with physical constraints. The physical constraints are implemented by embedding the weighted residual terms of the thermodynamic control equation and the mechanical equation as regularization terms into the loss function.

8. The physical constraint optimization method based on in-situ monitoring of semiconductor material defects according to claim 5, characterized in that: The S4 step, which outputs the preliminary predicted dislocation density and optimized process parameters, also includes the following parallel verification steps: First verification path: The physical consistency verification module integrates an independent physical model. The theoretical dislocation density calculated by the independent physical model is compared with the preliminary predicted dislocation density. When the relative error between the two exceeds the second preset threshold, the lightweight finite element model is triggered to make corrections and the corrected dislocation density value is output as the basis for process adjustment. The second verification path is as follows: During the calculation of the independent physical model, the defect confidence level obtained by the independent physical model based on the multi-physics field consistency assessment is obtained simultaneously. When the defect confidence level is lower than the third preset threshold, an adversarial perturbation that conforms to the symmetry direction of the intrinsic slip system of the target material is injected into the input of the physical constraint neural network model.

9. The physical constraint optimization method based on in-situ monitoring of semiconductor material defects according to claim 5, characterized in that: The closed-loop adjustment mechanism in S6 specifically includes: The edge computing node determines whether the dislocation density is abnormally increasing by a first preset threshold. If so, it issues control commands through the real-time industrial communication protocol deployed between the edge computing node and the vapor deposition execution unit. The control commands include temperature setpoint adjustment or gas flow rate setpoint adjustment. The temperature adjustment adopts a feedforward-feedback composite control strategy, and the gas flow rate adjustment adopts an adaptive smooth transition curve to avoid process step disturbances.

10. A physical constraint optimization system based on in-situ monitoring of defects in semiconductor materials, characterized in that: include: In-situ data acquisition module: used for epitaxial growth in the first growth chamber, and to acquire dislocation feature vectors and process parameters in real time through the in-situ cathode fluorescence monitoring unit and the first in-situ monitoring unit; Training set construction module: used to combine the actual dislocation density obtained by dislocation detection to construct a prediction training set containing dislocation feature vectors and actual dislocation density, and to construct an optimization training set containing target dislocation density, predicted dislocation density, initial process parameters and optimized process parameters; Model building and training module: used to train the dislocation density prediction model and the process parameter optimization model based on the physical constraint neural network module, combined with the prediction training set and the optimization training set; during the model training process, the residual terms of the semiconductor thermodynamic control equation and the residual terms of the mechanical equilibrium equation are used as physical constraint embedding loss functions to train the physical constraint neural network model. Online prediction and optimization module: It is used to input the dislocation feature vector and process parameters collected in real time into the trained physical constraint neural network model, dynamically map and adjust the weight of the physical constraint according to the fluctuation characteristics of the gas pressure in the first growth chamber within the preset process coordination time window, and increase the weight when the gas pressure fluctuation exceeds the disturbance judgment standard to enhance the constraint strength of physical laws on the neural network, and output the preliminary predicted dislocation density and optimized process parameters. Dual-chamber collaborative execution module: The dual-chamber collaborative control module controls the start-up of the second growth chamber, the start-up time of which lags behind the preset process collaboration time window of the first growth chamber, and performs batch epitaxial growth in the second growth chamber based on the optimized process parameters; Closed-loop feedback iteration module: used to collect the current growth status data of the second growth chamber in real time during the growth process of the second growth chamber, and compare it with the optimized process parameters and corresponding dislocation feature vectors of the first growth chamber; When the difference between the two exceeds the first preset threshold, a closed-loop adjustment mechanism is triggered to compensate the process parameters of the second growth chamber in real time.

Images