Molecular beam epitaxial growth process control method based on artificial intelligence

Abstract

The invention relates to the field of infrared detection, and discloses a molecular beam epitaxial growth process control method based on artificial intelligence, which comprises the following steps: S1, growth material selection: selecting a semiconductor material with specific photoelectric properties according to target application requirements; s2, wafer growth: realizing epitaxial growth of the wafer in an ultrahigh vacuum environment; s3, data acquisition and high-quality wafer screening: acquiring growth process data in real time, and detecting wafer quality through X-ray diffraction and an atomic force microscope; s4, data processing: performing signal processing and dimension reduction operation on the acquired data, and removing noise and abnormal values; s5, model training; and S6, optimizing and controlling. The system can automatically analyze and adjust key parameters in the growth process through the artificial intelligence technology, accurate control is achieved, and dependence on experience of operators is reduced. The system automatically processes and optimizes growth parameters, and even if the experience of an operator is insufficient, high-quality material growth can be achieved through the system.

Description

Technical Field

[0001] The invention relates to the field of infrared detection, and in particular to a molecular beam epitaxial growth process control method based on artificial intelligence. Background Art

[0002] Molecular Beam Epitaxy (MBE) is a technology that grows high-quality single-crystal thin films in an ultra-high vacuum environment. It is widely used in semiconductor devices, infrared detectors, quantum devices and other fields. The core of MBE technology is to precisely control the molecular beams of different materials to deposit on the substrate surface and build the crystal structure with atomic-level precision. Due to its ability to precisely control the composition and thickness of the material, MBE has become the preferred technology for preparing superlattice structures and low-dimensional semiconductor materials.

[0003] However, traditional MBE technology mainly relies on the operator's experience and manual control of the equipment. During the growth process, the operator needs to monitor and adjust the chamber temperature, molecular beam intensity, substrate rotation speed and other parameters in real time to ensure the stability of the growth process and the quality of the material. However, many key operations in the MBE growth process rely on the operator's experience and judgment. Even a small operational error may lead to a decline in material quality or growth failure. Summary of the invention

[0004] In view of the shortcomings of the prior art, the present invention provides a molecular beam epitaxial growth process control method based on artificial intelligence to solve the problem that many key operations in the MBE growth process rely on the operator's experience and judgment, and even minor operational errors may lead to material quality degradation or growth failure.

[0005] To achieve the above objectives, the present invention is implemented by the following technical solutions: A molecular beam epitaxial growth process control method based on artificial intelligence, comprising the following steps:

[0006] S1: Growth material selection

[0007] According to the target application requirements, semiconductor materials with specific optoelectronic properties are selected, and the substrate surface is clean and at an appropriate temperature through multi-step cleaning and preheating treatment;

[0008] S2: Wafer Growth

[0009] In an ultra-high vacuum environment with a pressure no higher than 1×10^-10Torr, epitaxial growth of wafers is achieved by precisely controlling the injection angle and temperature of the molecular beam.

[0010] S3: Data collection and high-quality wafer screening

[0011] Collect growth process data in real time, detect wafer quality through X-ray diffraction and atomic force microscopy, and select wafers that meet specific structural standards;

[0012] S4: Data Processing

[0013] Perform signal processing and dimensionality reduction operations on the collected data to remove noise and outliers, and extract key features through principal component analysis;

[0014] S5: Model training

[0015] A multi-layer neural network model is used to train the model using pre-processed data to predict and optimize key parameters in the growth process;

[0016] S6: Optimization and Control

[0017] Based on the model's prediction results, the growth conditions are adjusted in real time, and the cavity temperature and molecular beam intensity are dynamically adjusted through a closed-loop control system.

[0018] Preferably, the growth material selection includes the following specific steps:

[0019] S101: Based on the optoelectronic performance requirements of the target application, select a semiconductor material with a light absorption coefficient of 10^4 to 10^6 cm^-1 in the target wavelength range, a bandgap of 0.1 to 0.5 eV, and a carrier mobility in the range of 1000 to 5000 cm^2 / V·s;

[0020] S102: Perform chemical purity test on the selected material to ensure that the impurity concentration of the material is lower than 10^(-6)ppm to meet the material purity requirements of the molecular beam epitaxial growth process;

[0021] S103: The substrate is cleaned in multiple steps by ultrasonic cleaning, deionized water rinsing, anhydrous ethanol degreasing and drying to ensure that there is no organic residue and oxide layer on the surface;

[0022] S104: The substrate is preheated in a preheating furnace controlled at 350°C ±5°C for at least 2 hours to remove residual gas molecules and moisture on the surface of the substrate, and the surface flatness of the substrate is detected by optical interferometry.

[0023] Preferably, the wafer growth comprises the following specific steps:

[0024] S201: In an ultra-high vacuum environment with a pressure not higher than 1×10^-10Torr, a molecular beam is formed by adjusting the temperature of the beam source furnace to 970°C to 1150°C, and the injection angle of the beam is controlled at 90°±0.5°;

[0025] S202: By adjusting the opening and closing angle and the switching frequency of the mechanical baffle, the injection time and angle of the molecular beam are controlled so that the molecular beam is incident vertically onto the substrate surface at a stable speed, and a multi-layer shielding method is used to reduce the scattering of the molecular beam in the vacuum cavity;

[0026] S203: Use a quadrupole mass spectrometer to monitor the gas composition in the molecular beam in real time, and use a reflection high-energy electron diffraction instrument to analyze the lattice structure changes on the substrate surface in real time, and adjust the beam source temperature, molecular beam intensity and substrate rotation speed during the growth process.

[0027] Preferably, the data collection and high-quality wafer screening includes the following specific steps:

[0028] S301: Deploy temperature sensors, pressure sensors, optical sensors and mass flow controllers to form a multi-sensor data acquisition network to collect multi-dimensional data of chamber temperature, molecular beam intensity, beam type, substrate temperature, rotation speed and chamber pressure during the growth process in real time, and record them synchronously in the central data storage system;

[0029] S302: Use X-ray diffraction equipment to perform nondestructive testing on each wafer, obtain the diffraction pattern of the crystal structure, and analyze the interlayer stress distribution and atomic arrangement order in the superlattice structure by X-ray scattering method;

[0030] S303: Use an atomic force microscope to perform high-precision scanning on the wafer surface to obtain surface roughness, microscopic particle distribution and grain boundary position morphology information, and ensure the consistency of surface morphology through multiple scans;

[0031] S304: Based on the diffraction peak intensity, peak position, half-peak width and diffraction angle difference of the XRD diffraction pattern, as well as the surface morphology data of the AFM scan, a multi-dimensional quality assessment is performed on each wafer to screen out high-quality wafers that meet the design requirements, and the corresponding growth process data and quality assessment results are associated and annotated in the central database.

[0032] Preferably, the data processing includes the following specific steps:

[0033] S401: using a signal processing algorithm based on Fourier transform and wavelet analysis to perform frequency domain analysis on the collected growth process data to remove random noise and periodic interference signals;

[0034] S402: Perform outlier detection on the preliminarily processed data, and use a dual algorithm based on statistical analysis and machine learning to identify and remove extreme values ​​and outliers;

[0035] S403: Using principal component analysis and linear discriminant analysis methods, the processed data is subjected to dimensionality reduction and feature extraction to simplify the input dimension of model training, and the data is standardized to a zero mean and unit variance form to improve the convergence speed and prediction accuracy of the model;

[0036] S404: Divide the dataset after dimension reduction into a training set and a test set, and annotate metadata including the source, acquisition time, equipment parameters and environmental conditions of each dataset in the database.

[0037] Preferably, the model training includes the following steps:

[0038] S501: According to the time series characteristics and nonlinear relationship of the growth process data, a neural network model with time-dependent characteristics is selected, including a long short-term memory network or a gated recurrent unit, and a convolutional neural network is combined to extract the spatial characteristics of the multi-dimensional data;

[0039] S502: using a distributed computing framework, distributing the training data set to multiple computing nodes, and training the neural network model in parallel through a gradient descent algorithm and a back propagation algorithm to speed up the model training;

[0040] S503: Use cross-validation to evaluate the prediction performance of the model, adjust the model's hyperparameters on the validation set, including learning rate, regularization parameter, number of layers and number of nodes, and optimize the model's structure to prevent overfitting;

[0041] S504: Use the Bayesian optimization method to further tune the model hyperparameters, and find the best hyperparameter combination by combining grid search and random search.

[0042] Preferably, the optimization and control include the following specific steps:

[0043] S601: During the growth process, based on the real-time prediction results of the model, the growth parameters including the chamber temperature, molecular beam intensity, beam switching time and substrate rotation speed are dynamically adjusted through a closed-loop control system, and the settings of these parameters are automatically optimized according to the prediction of the growth conditions by the model;

[0044] S602: When the system detects growth parameters that significantly deviate from the model prediction results, the safety mechanism is triggered to automatically suspend the growth process, and the cause of the abnormality is analyzed through the intelligent diagnosis module, and possible fault points and adjustment suggestions are provided;

[0045] S603: Evaluate the final crystal quality of each growth process, compare and analyze the evaluation results with the real-time data and model prediction results during the growth process, and optimize the feedback mechanism and model update strategy of the closed-loop control system;

[0046] S604: The control parameters and corresponding crystal quality data after each optimization are stored in a central database as reference data for subsequent model training and growth process improvement, and historical data in the database are reanalyzed regularly to discover potential patterns and trends.

[0047] Preferably, the growth parameter adjustment in step S2 further comprises:

[0048] The substrate temperature is adjusted in real time through a closed-loop PID control algorithm, and the lattice constant change calculation at different temperatures is used to adjust the substrate temperature in real time to optimize the interlayer stress distribution during the growth process;

[0049] The components of the molecular beam are analyzed online by using multi-spectral mass spectrometry, and the stability and uniformity of the beam intensity are maintained by adjusting the heating power and cooling rate of the beam source furnace during the molecular beam intensity adjustment process.

[0050] The high-quality wafer screening criteria in the S3 step also include:

[0051] Perform high-precision fitting analysis on the multi-level diffraction peaks in the X-ray diffraction pattern to determine the tiny defects and stress concentration areas in the crystal structure, and further evaluate the growth quality of the superlattice material by combining the interlayer thickness and the diffraction peak position of the periodic structure;

[0052] Based on the atomic force microscope scanning, a scanning electron microscope is used to perform secondary surface analysis on the wafer to obtain the morphological information and composition analysis results of the surface particles, and to compare and analyze the AFM and SEM data.

[0053] Preferably, the neural network model in step S5 is a multi-layer deep neural network, and comprises multiple hidden layers, each hidden layer comprises 100 to 1000 neuron nodes, and is used to process multi-dimensional nonlinear data relationships. The input layer of the model accepts data features processed by PCA dimensionality reduction, and the output layer is used to predict key parameters in the growth process.

[0054] The hyperparameter optimization in the model training also includes:

[0055] Compare and analyze the structures of different types of neural networks, select the network structure that best suits the current data characteristics and prediction needs, and adjust parameters including learning rate, regularization coefficient, weight initialization method, and activation function type according to the characteristics of different networks;

[0056] During the model training process, through multi-step pre-training and transfer learning methods, the existing growth process data and training models of similar materials are used to initialize and accelerate the training of the new model, thereby reducing the training time.

[0057] Preferably, the closed-loop control system in step S6 further includes:

[0058] A multi-objective optimization module based on genetic algorithm and particle swarm optimization algorithm, which can automatically adjust the weights of multiple growth parameters during the growth process and find the optimal solution among multiple growth conditions according to different materials and target crystal characteristics;

[0059] In the anomaly detection of the growth process, Bayesian reasoning and Markov chain Monte Carlo methods are used to make probabilistic inferences on possible anomaly causes, and a variety of repair strategies are proposed based on the diagnosis results for operators to choose from.

[0060] The present invention provides a molecular beam epitaxial growth process control method based on artificial intelligence, which has the following beneficial effects:

[0061] 1. The present invention uses artificial intelligence technology to enable the system to automatically analyze and adjust key parameters in the growth process, achieve precise control, and reduce dependence on operator experience. The system automatically processes and optimizes growth parameters, and even if the operator is inexperienced, high-quality material growth can be achieved through the system.

[0062] 2. The system of the present invention monitors the growth process in real time through a variety of sensors and quickly adjusts parameters to ensure the stability of growth conditions. In addition, the system can automatically adjust growth parameters according to changes in different materials and growth environments, reducing experimental requirements and having strong adaptability.

[0063] 3. The system of the present invention improves the growth rate and material quality, reduces the defective rate, and improves production efficiency by optimizing key parameters. And by optimizing control and reducing defective products, the system reduces production costs, while reducing energy consumption and equipment wear. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 This is a typical X-ray spectrum of the InAs / GaSb type II superlattice material of the present invention;

[0065] Figure 2 This is an AFM scan of a typical InAs / GaSb type II superlattice material of the present invention;

[0066] Figure 3 Schematic diagram of a typical neural network structure of the present invention. DETAILED DESCRIPTION

[0067] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings 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 creative work are within the scope of protection of the present invention.

[0068] Please refer to the attached Figure 1 - Attachment Figure 3 The embodiment of the present invention provides a molecular beam epitaxial growth process control method based on artificial intelligence, comprising the following steps:

[0069] S1: Growth material selection

[0070] S101: Material selection:

[0071] Based on the optoelectronic performance requirements of the target application, the target wavelength range is first determined, such as the near-infrared region (0.75 to 1.4 μm) or the mid-infrared region (3 to 5 μm).

[0072] Analyze the light absorption coefficient of different semiconductor materials (such as GaSb, InAs, InSb, etc.) within this wavelength range, which should range from 10^4 to 10^6cm^-1.

[0073] The bandgap widths of these materials were further calculated to ensure that they were in the range of 0.1 to 0.5 eV to match the performance requirements of optoelectronic devices.

[0074] Compared with the carrier mobility, materials with a mobility of 1000 to 5000 cm^2 / V·s are preferred to ensure high photoelectric conversion efficiency.

[0075] S102: Purity test:

[0076] The selected semiconductor materials are chemically analyzed using high-precision instruments such as ICP-MS (Inductively Coupled Plasma Mass Spectrometry) to ensure that the impurity concentration of the material does not exceed 10^(-6)ppm.

[0077] During the inspection process, special attention is paid to impurity elements that may affect the growth quality, such as oxygen, carbon, and metallic impurities.

[0078] S103: Substrate cleaning:

[0079] The substrates were first cleaned using ultrasonic waves for 10 min with a frequency set to 40 kHz to remove particles and contaminants attached to the surface.

[0080] Deionized water rinsing and anhydrous ethanol degreasing treatment were performed in sequence, each step lasting 5 minutes to remove organic matter and other impurities.

[0081] Finally, place the substrate in a clean air flow drying cabinet to dry to avoid secondary contamination.

[0082] S104: Preheating treatment:

[0083] Place the cleaned substrate into a preheating furnace, set the temperature to 350°C ± 5°C, and preheat for no less than 2 hours.

[0084] An optical interferometer is used to monitor the substrate surface in real time to ensure that its surface flatness is within a quarter of the wavelength (i.e., the fringes generated by optical interference do not exceed a deviation of 1 / 4 wavelength) to ensure the uniformity of the growth layer.

[0085] S2: Wafer Growth

[0086] S201: Growth preparation in ultra-high vacuum environment:

[0087] Start the MBE system and ensure that the pressure in the vacuum chamber drops below 1×10^-10Torr.

[0088] Start the heating system of the beam source furnace, gradually increase the temperature to 970℃ to 1150℃, and select the appropriate evaporation temperature according to different materials.

[0089] Adjust the molecular beam injection angle to 90°±0.5° to ensure that the beam is perpendicular to the substrate surface.

[0090] S202: Molecular beam control:

[0091] By adjusting the opening and closing angle (0° to 90°) and the switching frequency (1 to 10 Hz) of the mechanical baffle, the injection time of the molecular beam is controlled to ensure the stability of the beam.

[0092] Multi-layer shielding technology is used, and metal shielding plates and graphite partitions are installed to reduce the scattering of the beam in the vacuum cavity and avoid interference with other devices.

[0093] Real-time monitoring of the position and opening and closing status of the mechanical baffle ensures that the beam flows at a constant rate (0.1 to ) is incident on the substrate surface.

[0094] S203: Real-time monitoring and adjustment:

[0095] A quadrupole mass spectrometer (QMS) is used to analyze the gas composition of the evaporated molecular beam, detect and adjust the concentration ratio of each element, and ensure the accuracy of the stoichiometric ratio of the material.

[0096] The reflection high-energy electron diffraction (RHEED) system is used to monitor the diffraction pattern of the substrate surface lattice and analyze the crystal structure and quality of the grown layer in real time.

[0097] According to the monitoring results, the beam source temperature (within ±5℃), molecular beam intensity ( range) and substrate rotation speed (1 to 10 rpm) to optimize the growth process.

[0098] S3: Data collection and high-quality wafer screening

[0099] S301: Construction of multi-sensor data acquisition network:

[0100] Deploy temperature sensors (range 200℃ to 800℃, accuracy ±0.5℃), pressure sensors (range 1×10^-10Torr to 1×10^-2Torr, accuracy ±1%), optical sensors (wavelength detection range 200nm to 2000nm, accuracy ±1nm) and mass flow controllers (flow range 0.01 to 10sccm, accuracy ±1%).

[0101] The sensor data is transmitted to the central data storage system via a high-speed data acquisition card (sampling rate 1000Hz), ensuring real-time and synchronous recording of all key parameters in the growth process.

[0102] S302: X-ray diffraction detection:

[0103] Using X-ray diffractometer (XRD), select CuKα radiation (wavelength ), perform a full spectrum scan with a 2θ scanning range of 10° to 90°, and detect the diffraction pattern of the chip.

[0104] High-precision fitting of the half-width, intensity and peak position of the diffraction peak is performed to analyze the interlayer stress and atomic arrangement of the superlattice structure to ensure that the layer thickness error is less than

[0105] S303: AFM scanning analysis:

[0106] An atomic force microscope (AFM) is used to perform high-precision scans of the wafer surface with a resolution of 1nm to analyze the surface roughness, particle distribution, and grain boundary position, ensuring that the repeatability error of the scanned area does not exceed ±0.5nm.

[0107] Each wafer was scanned multiple times, and the results were averaged to ensure data consistency and compared with XRD results to verify the uniformity of surface morphology.

[0108] S304: Multi-dimensional quality assessment and annotation:

[0109] Based on the XRD and AFM test results, multi-dimensional quality assessment was carried out, with special attention paid to the diffraction peak intensity (≥95%), peak position deviation (≤0.01°) and surface particle diameter (≤5nm).

[0110] Wafers that meet the requirements are marked, and their growth process data, quality assessment results, and detection map data are associated and stored in a central database to facilitate subsequent analysis and optimization.

[0111] S4: Data Processing

[0112] S401: Signal Processing:

[0113] The collected growth process data were converted into the frequency domain using a frequency domain analysis method based on Fourier transform to remove high-frequency noise (≥50 Hz) and low-frequency drift (≤0.01 Hz).

[0114] Wavelet analysis method is used to decompose the periodic interference signals in the data to ensure the purity and stability of the signal.

[0115] S402: Outlier Detection:

[0116] The 3σ principle based on statistical analysis is used to detect extreme values ​​and outliers in the data, and the isolation forest algorithm based on machine learning is used to further eliminate outliers to ensure the integrity and consistency of the data.

[0117] S403: Dimensionality reduction and feature extraction:

[0118] The principal component analysis (PCA) method is used to reduce the dimensionality of high-dimensional data, extract the key features covered by the first 90% cumulative variance, and reduce the data dimension to 2 to 5 principal components.

[0119] Linear discriminant analysis (LDA) was used to further optimize feature extraction and ensure the simplicity and representativeness of the input data for model training.

[0120] S404: Dataset construction and annotation:

[0121] The processed data set is divided into a training set (accounting for 80%) and a test set (accounting for 20%), and metadata such as the source (such as equipment model, operating parameters), collection time, and environmental conditions (such as room temperature and humidity) of each data set are annotated in the database to ensure data traceability and accuracy of analysis.

[0122] S5: Model training

[0123] S501: Neural network model selection:

[0124] According to the time series characteristics and nonlinear relationship of the growth process data, the long short-term memory network (LSTM) or the gated recurrent unit (GRU) is selected, and the number of hidden layers is set to 2 to 3, with 100 to 500 neurons in each layer.

[0125] The convolutional neural network (CNN) is combined to extract spatial features to enhance the feature capture ability of the model, and the convolution kernel size is set to 3×3 or 5×5.

[0126] S502: Distributed training:

[0127] Use a distributed computing framework (such as TensorFlow or PyTorch), distribute the training dataset to multiple computing nodes, use GPU acceleration, set the batch size to 32 to 128, and the number of training iterations to 100 to 1000.

[0128] Through the stochastic gradient descent (SGD) algorithm and the back propagation algorithm, the neural network model is trained in parallel to accelerate the convergence of the model.

[0129] S503: Model verification and optimization:

[0130] The cross-validation method (K = 5) was used to evaluate the predictive performance of the model on the validation set, and the learning rate (0.001 to 0.01), regularization parameter (0.01 to 0.1), number of layers (2 to 3 layers), and number of nodes (100 to 500) were adjusted.

[0131] Bayesian optimization and grid search are combined to further tune hyperparameters to improve the generalization ability and prediction accuracy of the model.

[0132] S504: Hyperparameter Tuning:

[0133] Use the Bayesian optimization method to perform a global search on the model's hyperparameters. By combining grid search with random search, find the optimal hyperparameter combination to ensure that the model performs best on the test set.

[0134] S6: Optimization and Control

[0135] S601: Real-time adjustment of closed-loop control system:

[0136] During the growth process, the cavity temperature (±1°C) and molecular beam intensity are dynamically adjusted through a closed-loop control system based on the real-time prediction results of the model. Beam switch time (±0.1 sec) and substrate rotation speed (±0.1 rpm).

[0137] Using PID control algorithm, the substrate temperature is adjusted in real time to ensure the uniformity of lattice constant and minimize interlayer stress.

[0138] S602: Abnormal detection and processing:

[0139] When the system detects that the growth parameters deviate from the model prediction results by more than 5%, the safety mechanism is triggered to automatically suspend growth and start the intelligent diagnosis module to analyze the cause of the abnormality through Bayesian reasoning and Markov chain Monte Carlo method.

[0140] Provides a variety of repair strategies, such as adjusting beam source temperature, adjusting substrate rotation speed, recalibrating mechanical shutters, etc., for operators to choose and execute.

[0141] S603: Quality Assessment and Feedback Optimization:

[0142] The final crystal quality of each growth process was evaluated, including the intensity and position of XRD diffraction peaks, surface roughness and particle distribution by AFM scanning, etc.

[0143] The evaluation results are compared with the real-time data and model prediction results during the growth process, the feedback mechanism of the closed-loop control system is optimized, and the model update strategy is adjusted to ensure that the next round of growth process is more stable and efficient.

[0144] S604: Data storage and reanalysis:

[0145] The control parameters and crystal quality data after each optimization are stored in a central database, and the relevant metadata are recorded to ensure data integrity and traceability.

[0146] Historical data in the database are reanalyzed regularly to identify potential areas for improvement through pattern recognition and trend analysis, so as to continuously optimize the growth process and control system.

[0147] Implementation process description

[0148] Step 1: Growth Material Selection

[0149] Select appropriate semiconductor materials, here GaSb is selected as the epitaxial substrate, InAl / GaSb / AlSb material is selected as the absorption region, and AlAsSb / GaSb material is selected as the barrier region. These materials show good sensitivity in the infrared spectrum range.

[0150] Step 2: Wafer Growth

[0151] A thin film of the selected material is periodically grown on a substrate using molecular beam epitaxy.

[0152] Step 3: Data collection and high-quality wafer screening

[0153] When using the data acquisition module, data such as cavity temperature, molecular beam intensity and type during material growth are collected. After the wafer growth is completed, the sample is characterized by high-resolution X-ray diffraction (XRD), which will not cause damage while checking the quality of the superlattice crystal. A good X-ray image should have high-order diffraction, a small peak position difference between the superlattice zero-order peak and the substrate peak, and a small half-peak width of the diffraction image. In addition, atomic force microscopy (AFM) can also be used to measure surface morphology. If there are no protrusions or depressions on the surface under long-term stable atomic force scanning, the surface quality is considered to be good. Figure 1 The typical X-ray spectrum of InAs / GaSb type II superlattice materials is shown; Figure 2 The typical AFM scan of InAs / GaSbⅡ type superlattice material is shown; after XRD and AFM detection, high-quality wafers are screened out, the X-ray images and surface morphology characteristics of high-quality wafers are recorded, and the corresponding data in the growth process are found for one-to-one correspondence. In this process, a large number of wafer growth experiments are required to obtain a large number of data in the growth process corresponding to high-quality wafers.

[0154] Step 4: Data Processing

[0155] The data of the growth process corresponding to the screened high-quality wafers are preprocessed, including the removal and standardization of noise and outliers, to ensure the quality and consistency of the data.

[0156] Step 5: Model training

[0157] The preprocessed data set is divided into a training set and a test set, and the training set data is trained using a neural network. A predictive model that can relate growth process parameters to film properties and optimize growth conditions to obtain the desired results is developed and validated using the test set. Figure 3 A typical schematic diagram of a neural network structure is given.

[0158] Step 6: Optimization and Control

[0159] According to the results predicted by the model, the growth conditions, such as cavity temperature, molecular beam intensity and other parameters, are intelligently adjusted to achieve the expected crystal growth effect.

[0160] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A molecular beam epitaxial growth process control method based on artificial intelligence, characterized in that: The following steps are involved: S1: Growth material selection According to the target application requirements, semiconductor materials with specific optoelectronic properties are selected, and the substrate surface is clean and at an appropriate temperature through multi-step cleaning and preheating treatment; S2: Wafer Growth In an ultra-high vacuum environment with a pressure no higher than 1×10^-10Torr, epitaxial growth of wafers is achieved by precisely controlling the injection angle and temperature of the molecular beam. S3: Data collection and high-quality wafer screening Collect growth process data in real time, test wafer quality through X-ray diffraction and atomic force microscopy, and select wafers that meet specific structural standards; S4: Data Processing Perform signal processing and dimensionality reduction operations on the collected data to remove noise and outliers, and extract key features through principal component analysis; S5: Model training A multi-layer neural network model is used to train the model using pre-processed data to predict and optimize key parameters in the growth process; S6: Optimization and Control Based on the model's prediction results, the growth conditions are adjusted in real time, and the cavity temperature and molecular beam intensity are dynamically adjusted through a closed-loop control system.

2. The method for controlling molecular beam epitaxial growth process based on artificial intelligence according to claim 1, characterized in that: The growth material selection comprises the following specific steps: S101: Based on the optoelectronic performance requirements of the target application, select a semiconductor material with a light absorption coefficient of 10^4 to 10^6 cm^-1 in the target wavelength range, a bandgap of 0.1 to 0.5 eV, and a carrier mobility in the range of 1000 to 5000 cm^2 / V·s; S102: Perform chemical purity test on the selected materials to ensure that the impurity concentration of the materials is lower than 10^(-6)ppm to meet the material purity requirements of the molecular beam epitaxial growth process; S103: The substrate is cleaned in multiple steps by ultrasonic cleaning, deionized water rinsing, anhydrous ethanol degreasing and drying to ensure that there is no organic residue and oxide layer on the surface; S104: The substrate is preheated in a preheating furnace controlled at 350°C ±5°C for at least 2 hours to remove residual gas molecules and moisture on the surface of the substrate, and the surface flatness of the substrate is detected by optical interferometry.

3. The method for controlling molecular beam epitaxial growth process based on artificial intelligence according to claim 1, characterized in that: The wafer growth comprises the following specific steps: S201: In an ultra-high vacuum environment with a pressure not higher than 1×10^-10Torr, a molecular beam is formed by adjusting the temperature of the beam source furnace to 970°C to 1150°C, and the injection angle of the beam is controlled at 90°±0.5°; S202: By adjusting the opening and closing angle and the switching frequency of the mechanical baffle, the injection time and angle of the molecular beam are controlled so that the molecular beam is incident vertically onto the substrate surface at a stable speed, and a multi-layer shielding method is used to reduce the scattering of the molecular beam in the vacuum cavity; S203: Use a quadrupole mass spectrometer to monitor the gas composition in the molecular beam in real time, and use a reflection high-energy electron diffraction instrument to analyze the lattice structure changes on the substrate surface in real time, and adjust the beam source temperature, molecular beam intensity and substrate rotation speed during the growth process.

4. The method for controlling molecular beam epitaxial growth process based on artificial intelligence according to claim 1, characterized in that: The data collection and high-quality wafer screening include the following specific steps: S301: Deploy temperature sensors, pressure sensors, optical sensors and mass flow controllers to form a multi-sensor data acquisition network to collect multi-dimensional data of chamber temperature, molecular beam intensity, beam type, substrate temperature, rotation speed and chamber pressure during the growth process in real time, and record them synchronously in the central data storage system; S302: Use X-ray diffraction equipment to perform nondestructive testing on each wafer, obtain the diffraction pattern of the crystal structure, and analyze the interlayer stress distribution and atomic arrangement order in the superlattice structure by X-ray scattering method; S303: Use an atomic force microscope to perform high-precision scanning on the wafer surface to obtain surface roughness, microscopic particle distribution and grain boundary position morphology information, and ensure the consistency of surface morphology through multiple scans; S304: Based on the diffraction peak intensity, peak position, half-peak width and diffraction angle difference of the XRD diffraction pattern, as well as the surface morphology data of the AFM scan, a multi-dimensional quality assessment is performed on each wafer to screen out high-quality wafers that meet the design requirements, and the corresponding growth process data and quality assessment results are associated and annotated in the central database.

5. The method for controlling molecular beam epitaxy growth process based on artificial intelligence according to claim 1, characterized in that: The data processing includes the following specific steps: S401: using a signal processing algorithm based on Fourier transform and wavelet analysis to perform frequency domain analysis on the collected growth process data to remove random noise and periodic interference signals; S402: Perform outlier detection on the preliminarily processed data, and use a dual algorithm based on statistical analysis and machine learning to identify and remove extreme values ​​and outliers; S403: Using principal component analysis and linear discriminant analysis methods, the processed data is subjected to dimensionality reduction and feature extraction to simplify the input dimension of model training, and the data is standardized to a zero mean and unit variance form to improve the convergence speed and prediction accuracy of the model; S404: Divide the dataset after dimension reduction into a training set and a test set, and annotate metadata including the source, acquisition time, equipment parameters and environmental conditions of each dataset in the database.

6. The method for controlling molecular beam epitaxial growth process based on artificial intelligence according to claim 1, characterized in that: The model training includes the following steps: S501: According to the time series characteristics and nonlinear relationship of the growth process data, a neural network model with time-dependent characteristics is selected, including a long short-term memory network or a gated recurrent unit, and a convolutional neural network is combined to extract the spatial characteristics of the multi-dimensional data; S502: using a distributed computing framework, distributing the training data set to multiple computing nodes, and training the neural network model in parallel through a gradient descent algorithm and a back propagation algorithm to speed up the model training; S503: Use cross-validation to evaluate the prediction performance of the model, adjust the model's hyperparameters on the validation set, including learning rate, regularization parameter, number of layers and number of nodes, and optimize the model's structure to prevent overfitting; S504: Use the Bayesian optimization method to further tune the model hyperparameters, and find the best hyperparameter combination by combining grid search and random search.

7. The method for controlling molecular beam epitaxial growth process based on artificial intelligence according to claim 1, characterized in that: The optimization and control include the following specific steps: S601: During the growth process, based on the real-time prediction results of the model, the growth parameters including the chamber temperature, molecular beam intensity, beam switching time and substrate rotation speed are dynamically adjusted through a closed-loop control system, and the settings of these parameters are automatically optimized according to the prediction of the growth conditions by the model; S602: When the system detects growth parameters that significantly deviate from the model prediction results, the safety mechanism is triggered to automatically suspend the growth process, and the cause of the abnormality is analyzed through the intelligent diagnosis module, and possible fault points and adjustment suggestions are provided; S603: Evaluate the final crystal quality of each growth process, compare and analyze the evaluation results with the real-time data and model prediction results during the growth process, and optimize the feedback mechanism and model update strategy of the closed-loop control system; S604: The control parameters and corresponding crystal quality data after each optimization are stored in a central database as reference data for subsequent model training and growth process improvement, and historical data in the database are reanalyzed regularly to discover potential patterns and trends.

8. The method for controlling molecular beam epitaxial growth process based on artificial intelligence according to claim 1, characterized in that: The growth parameter adjustment in step S2 further includes: The substrate temperature is adjusted in real time through a closed-loop PID control algorithm, and the lattice constant change calculation at different temperatures is used to adjust the substrate temperature in real time to optimize the interlayer stress distribution during the growth process; The components of the molecular beam are analyzed online by using multi-spectral mass spectrometry, and the stability and uniformity of the beam intensity are maintained by adjusting the heating power and cooling rate of the beam source furnace during the molecular beam intensity adjustment process. The high-quality wafer screening criteria in the S3 step also include: Perform high-precision fitting analysis on the multi-level diffraction peaks in the X-ray diffraction pattern to determine the tiny defects and stress concentration areas in the crystal structure, and further evaluate the growth quality of the superlattice material by combining the interlayer thickness and the diffraction peak position of the periodic structure; Based on the atomic force microscope scanning, a scanning electron microscope is used to perform secondary surface analysis on the wafer to obtain the morphological information and composition analysis results of the surface particles, and to compare and analyze the AFM and SEM data.

9. The method for controlling molecular beam epitaxial growth process based on artificial intelligence according to claim 1, characterized in that: The neural network model in step S5 is a multi-layer deep neural network, and includes multiple hidden layers, each hidden layer includes 100 to 1000 neuron nodes, and is used to process multi-dimensional nonlinear data relationships. The input layer of the model accepts data features that have been processed by PCA dimensionality reduction, and the output layer is used to predict key parameters in the growth process; The hyperparameter optimization in the model training also includes: Compare and analyze the structures of different types of neural networks, select the network structure that best suits the current data characteristics and prediction needs, and adjust parameters including learning rate, regularization coefficient, weight initialization method, and activation function type according to the characteristics of different networks; During the model training process, through multi-step pre-training and transfer learning methods, the existing growth process data and training models of similar materials are used to initialize and accelerate the training of the new model, thereby reducing the training time.

10. The method for controlling molecular beam epitaxial growth process based on artificial intelligence according to claim 1, characterized in that: The closed-loop control system in step S6 further includes: A multi-objective optimization module based on genetic algorithm and particle swarm optimization algorithm, which can automatically adjust the weights of multiple growth parameters during the growth process and find the optimal solution among multiple growth conditions according to different materials and target crystal characteristics; In the anomaly detection of the growth process, Bayesian reasoning and Markov chain Monte Carlo methods are used to make probabilistic inferences on possible anomaly causes, and a variety of repair strategies are proposed based on the diagnosis results for operators to choose from.

Images