Method, system, device and storage medium for adjusting semiconductor process parameters

Abstract

The application provides a semiconductor process parameter adjustment method, system, device and storage medium, relates to the technical field of semiconductor detection, and is based on an original reflectance spectrum and a process parameter of a current process cycle, acquires a reference threshold of a search parameter from a preset function library, reduces the dependence on a fixed threshold, and improves the detection stability under complex spectrum conditions. Based on historical reflectance spectra and corresponding process parameters, a preset prediction model is used to generate a characteristic wavelength range of the current detection cycle, so that the search range is dynamically updated with process drift, and the positioning efficiency and anti-interference ability are improved. Based on the reference threshold and the characteristic wavelength range, a feature search is performed on the original reflectance spectrum to output a fusion characteristic wavelength, higher-precision key wavelength extraction is realized, and reliable basis is provided for process decision. The fusion characteristic wavelength is fed back to a main control system and used for parameter adjustment of the next process cycle only when the effectiveness judgment meets the preset condition, and reliable closed-loop control is formed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of semiconductor detection, and in particular to a semiconductor process parameter adjustment method, system, device and storage medium. BACKGROUND

[0002] Vertical-Cavity Surface-Emitting Laser (VCSEL) is widely used in three-dimensional sensing, data communication and optical interconnection due to its low threshold current, circular light spot, easy array integration and other advantages. The key optical performances of VCSEL devices, such as emission wavelength, threshold characteristics and temperature stability, are mainly determined by the reflection spectrum characteristics of the Distributed Bragg Reflector (DBR) and the cavity resonance mode of the Fabry-Perot (FP) resonator. Among them, the stop band center position and width of the DBR determine the reflection boundary and gain coupling condition of the cavity mode, and the center wavelength of the cavity resonance peak (valley) of the cavity resonance mode directly corresponds to the resonance working point of the device. Therefore, during the epitaxial growth of VCSEL, by monitoring the reflection spectrum of the semiconductor growth surface in real time and accurately extracting the two key characteristic parameters of the stop band center wavelength and the cavity resonance wavelength, it is the core technology to control the thickness of the epitaxial layer, predict the performance of the final device, and realize the closed-loop control of the process.

[0003] In the traditional technology, an in-situ monitoring system based on spectrum method is generally used to monitor the epitaxial growth process online, for example, a broadband light source is used to irradiate the growth surface and the reflection spectrum is collected, and the growth state is inferred by analyzing the multi-layer thin film interference and reflection characteristics. However, in the high-temperature epitaxial growth scene of VCSEL, the optical constants such as the refractive index and the band gap of semiconductor materials have significant temperature dependence, which leads to a large redshift between the characteristic wavelength measured in-situ and the final performance of the device at room temperature; in addition, the reflection spectrum in the actual growth process is affected by factors such as surface roughness, gradual change of material composition, and multi-layer interference superposition, and the stop band edge and cavity resonance peak (valley) may not be obvious, or there may be multiple secondary peaks interference; the above situations will seriously affect the detection accuracy, and further affect the prediction and closed-loop control of the epitaxial growth process. SUMMARY

[0004] The purpose of the present application is to provide a semiconductor process parameter adjustment method, system, device and storage medium to overcome the defects of the traditional technology that affects the detection accuracy under high temperature and complex reflection spectrum conditions, and further affects the prediction and closed-loop control of the epitaxial growth process.

[0005] In a first aspect, the present application provides a semiconductor process parameter adjustment method, comprising:

[0006] acquire an original reflectance spectrum and process parameters of a current process cycle, wherein the original reflectance spectrum is generated based on a reflected light signal generated in a semiconductor epitaxial growth process;

[0007] at each detection cycle of the current process cycle, analyze the original reflectance spectrum frame by frame, and according to the analysis result and the process parameters, query a reference threshold of a search parameter from a preset function library;

[0008] based on the historical reflectance spectrum and the corresponding process parameters, generate a characteristic wavelength range of the current detection cycle by using a preset prediction model;

[0009] based on the characteristic wavelength range, perform feature search on the original reflectance spectrum under the constraint of the reference threshold to obtain a fused characteristic wavelength;

[0010] perform effectiveness judgment according to the original reflectance spectrum and the fused characteristic wavelength, and when the judgment result meets a preset condition, send the fused characteristic wavelength to a main control system of an epitaxial growth device, so that the main control system adjusts process parameters of a next process cycle based on the fused characteristic wavelength.

[0011] In one of the embodiments, the method further comprises:

[0012] at each detection cycle of the current process cycle, perform Gaussian pyramid decomposition on the original reflectance spectrum frame by frame to generate a plurality of spectrum sets of different scales;

[0013] based on the characteristic wavelength range, perform feature search on the spectrum sets of different scales under the constraint of the reference threshold to obtain a fused characteristic wavelength.

[0014] In one of the embodiments, the Gaussian pyramid decomposition on the original reflectance spectrum frame by frame to generate a plurality of spectrum sets of different scales comprises:

[0015] perform Gaussian kernel smoothing processing on an initial spectrum, wherein the initial spectrum is a current frame of the original reflectance spectrum;

[0016] perform down-sampling processing on the initial spectrum after the smoothing processing to generate a low-resolution spectrum of a first layer based on a preset down-sampling factor;

[0017] based on the low-resolution spectrum of the first layer, repeatedly perform the smoothing processing and the down-sampling processing to iteratively generate low-resolution spectra of a second layer and above;

[0018] the initial spectrum and the low-resolution spectra constitute a plurality of spectrum sets of different scales.

[0019] In one of the embodiments, the process parameters of the current process cycle include at least one of a growth layer number, a material component, a set temperature, and a set growth rate.

[0020] The original reflectance spectrum is analyzed frame by frame in each detection cycle of the current process cycle, and a reference threshold of the search parameter is obtained from a preset function library according to an analysis result and a process parameter, including:

[0021] In each detection cycle, the original reflectance spectrum is analyzed frame by frame to obtain a spectral feature parameter corresponding to each frame;

[0022] Based on a statistical result of each spectral feature parameter in the entire detection cycle, an index parameter of the original reflectance spectrum in a preset wavelength band is determined, wherein the index parameter includes a signal-to-noise ratio estimate value and a smoothness index;

[0023] Based on the index parameter and the process parameter, a table lookup and an interpolation calculation are performed in a preset function library to obtain a reference threshold of the search parameter.

[0024] In one of the embodiments, based on the historical reflectance spectrum and the corresponding process parameter, a preset prediction model is used to predict a feature wavelength range of the current detection cycle, including:

[0025] A preset number of historical reflectance spectra and corresponding process parameters are obtained, and a feature wavelength sequence is extracted from each historical reflectance spectrum to form a historical feature wavelength input sequence;

[0026] The historical feature wavelength input sequence and the corresponding process parameter are input into the preset prediction model to output a reference range of the current detection cycle;

[0027] Based on the reference range, a preset wavelength list is dynamically updated to generate a feature wavelength range for the current detection cycle.

[0028] In one of the embodiments, based on the feature wavelength range, a feature search is performed on different scale spectral sets under the constraint of the reference threshold to obtain a fusion feature wavelength, including:

[0029] In a low-scale low-resolution spectrum, each candidate wavelength in the feature wavelength range is taken as an initial center, and a candidate peak search and screening are performed in a preset search window according to the reference threshold to determine a low-scale estimate value of a stop band center;

[0030] Determine a prior search region based on the low-scale estimation value, perform stopband edge detection in the prior search region to determine a high-scale estimation value of the stopband center, search for a target trough in the prior search region, and calculate a cavity resonance characteristic wavelength of the target trough;

[0031] Weighted fusion of the low-scale estimation value and the high-scale estimation value to obtain a stopband center characteristic wavelength; and the cavity resonance characteristic wavelength and the stopband center characteristic wavelength constitute the fusion characteristic wavelength.

[0032] In one of the embodiments, the searching for a target trough in the prior search region and calculating a cavity resonance characteristic wavelength of the target trough include:

[0033] In the prior search region, trough searching is performed on the high-scale high-resolution spectrum to obtain at least one candidate trough.

[0034] For each candidate trough, a preset spectral line model is fitted by using a nonlinear least square fitting algorithm in an interval centered at the candidate trough and having a preset width to obtain a trough center wavelength, a spectral line depth, and a fitting goodness of the candidate trough; wherein the preset spectral line model is a parameterized model for representing the corresponding relationship between the wavelength and the reflectivity of the candidate spectral line.

[0035] The candidate troughs are screened based on the spectral line depth and the fitting goodness, and the trough center wavelength corresponding to the screened candidate trough is taken as the cavity resonance characteristic wavelength.

[0036] In one of the embodiments, the effectiveness judgment according to the original reflectance spectrum and the fusion characteristic wavelength includes:

[0037] Based on the original reflectance spectrum and the fusion characteristic wavelength, a two-dimensional feature vector is constructed.

[0038] The two-dimensional feature vector is input into a preset deep learning model to output a confidence score and a classification label; wherein the confidence score is in a preset numerical interval, and the classification label is used to represent the effectiveness state of the fusion characteristic wavelength.

[0039] Based on the confidence score and / or the classification label, a judgment result and a processing strategy of the fusion characteristic wavelength in a current detection period are determined.

[0040] In one of the embodiments, the preset numerical interval is between 0 and 1; and the classification label includes normal, suspicious, and failure.

[0041] determine a judgment result and a processing strategy of the fusion characteristic wavelength in a current detection period based on the confidence score and / or the classification label, including:

[0042] in a case where the confidence score is greater than or equal to a first threshold value and the classification label is normal, it is considered that a preset condition is met, and the fusion characteristic wavelength is output to a main control system of the epitaxial growth device;

[0043] in a case where the confidence score is less than or equal to a second threshold value or the classification label is suspicious, it is considered that the preset condition is not met, a prompt information is generated, and the prompt information and the fusion characteristic wavelength are output to the main control system of the epitaxial growth device;

[0044] in a case where the confidence score is less than the second threshold value or the classification label is failure, it is considered that the preset condition is not met, an alarm signal is generated, and the alarm signal is output to the main control system of the epitaxial growth device; wherein the second threshold value is less than the first threshold value.

[0045] In one of the embodiments, the method further includes:

[0046] after each detection period, detection data generated in the current detection period is acquired;

[0047] the detection data is added with a time sequence identifier, and after compression processing, the detection data added with the time sequence identifier is stored in a historical database;

[0048] new detection data is periodically extracted from the historical database as an incremental training sample, and the prediction model and the deep learning model are incrementally trained and updated based on the incremental training sample.

[0049] In a second aspect, the application provides a semiconductor process parameter adjustment system, the system includes:

[0050] an acquisition module, configured to acquire an original reflection spectrum and a process parameter in a current process period; wherein the original reflection spectrum is generated based on a reflection light signal generated in a semiconductor epitaxial growth process;

[0051] an adaptive threshold generation module, configured to analyze the original reflection spectrum frame by frame in each detection period of the current process period, and obtain a reference threshold value of a search parameter from a preset function library according to an analysis result and the process parameter;

[0052] an intelligent optimization module, configured to generate a characteristic wavelength range of the current detection period based on a historical reflection spectrum and a corresponding process parameter by using a preset prediction model;

[0053] The search module is used to perform a feature search on the original reflection spectrum based on the feature wavelength range and under the constraint of the reference threshold to obtain the fused feature wavelength;

[0054] The verification module is used to determine the validity based on the original reflectance spectrum and the fused characteristic wavelength;

[0055] The output module is used to send the fusion characteristic wavelength to the main control system of the epitaxial growth equipment when the judgment result meets the preset conditions, so that the main control system can adjust the process parameters of the next process cycle based on the fusion characteristic wavelength.

[0056] Thirdly, this application also provides a computer device. The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method steps of the first aspect.

[0057] Fourthly, this application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program thereon, which, when executed by a processor, implements the method steps of the first aspect.

[0058] The methods, systems, equipment, and storage media for adjusting the aforementioned semiconductor process parameters have at least the following advantages:

[0059] This application obtains the original reflectance spectrum and process parameters of the current process cycle, analyzes the original reflectance spectrum frame by frame and combines it with the process parameters, and obtains the reference threshold for the search parameters from a preset function library, reducing dependence on fixed thresholds and improving detection stability under complex spectral conditions. Furthermore, based on historical reflectance spectra and their corresponding process parameters, a preset prediction model is used to generate the characteristic wavelength range of the current detection cycle, allowing the search range to be dynamically updated with process drift, thereby improving feature localization efficiency and enhancing anti-interference capabilities. Furthermore, under the constraint of the reference threshold, feature search is performed on the original reflectance spectrum using the characteristic wavelength range, and the fused characteristic wavelength is output, achieving higher accuracy and repeatability in key wavelength extraction, providing a reliable basis for process decisions. Furthermore, the fused characteristic wavelength is fed back to the main control system only when the validity judgment meets preset conditions and used for parameter tuning in the next process cycle, forming a reliable closed-loop control to improve the controllability, wavelength consistency, and yield of the epitaxial process. Attached Figure Description

[0060] Figure 1 This is a flowchart illustrating a method for adjusting semiconductor process parameters in one embodiment;

[0061] Figure 2 This is a flowchart illustrating the step of querying and obtaining a reference threshold in one embodiment;

[0062] Figure 3 a flowchart of a step of obtaining a characteristic wavelength range of a current detection period in an embodiment;

[0063] Figure 4 a flowchart of a step of obtaining a fusion characteristic wavelength in an embodiment;

[0064] Figure 5 a flowchart of a step of validity determination in an embodiment;

[0065] Figure 6 a structural block diagram of a semiconductor process parameter adjustment system in an embodiment;

[0066] Figure 7 an internal structural diagram of a computer device in an embodiment. DETAILED DESCRIPTION

[0067] The present application is described and explained with additional specificity and detail through the use of the accompanying drawings in which:

[0068] Some example embodiments of the present application are described for illustrative purposes, and it is to be understood that the present application can be carried out by other ways and means without departing from the spirit and substance of the present application.

[0069] Reference will now be made to the drawings, wherein: Figure 1 In one example embodiment, the present application provides a semiconductor process parameter adjustment method, which specifically includes the following steps:

[0070] In step 102, an original reflectance spectrum and a process parameter of a current process period are obtained, wherein the original reflectance spectrum is generated based on a reflected light signal generated in a semiconductor epitaxial growth process.

[0071] Specifically, the process period refers to a process execution unit that can be independently controlled and adjusted in the semiconductor epitaxial growth process, and usually corresponds to a step segment (Step) in an epitaxial recipe. In the same process period, the epitaxial growth equipment performs growth control according to the preset process setting value.

[0072] The process parameter refers to a process setting quantity or a state quantity of the process period controlled by the main control system, including at least one of a growth layer number, a material component, a set temperature, and a set growth rate.

[0073] The semiconductor reaction cavity is provided with a bearing disc made of high-purity graphite, and a plurality of trays for placing semiconductors are arranged on the bearing disc. The bearing disc is supported by a rotating shaft driven by a motor and can be continuously rotated during the epitaxial growth of the semiconductor. The top of the semiconductor reaction cavity is provided with an incident optical window and an outgoing optical window. A halogen tungsten lamp or an ultrashort continuous spectrum light source is used as a broadband light source. The output probe light is irradiated to the surface of the semiconductor through the incident optical window at a preset incident angle, and the reflected light is coupled to a spectrometer by an optical fiber through the outgoing optical window. The spectrometer continuously collects the reflection spectrum during the growth process at a preset sampling frequency, and performs dark field subtraction and reference spectrum normalization processing on the reflection spectrum, so as to obtain the original reflection spectrum of the current process period.

[0074] Step 104, in each detection period of the current process period, the original reflection spectrum is analyzed frame by frame, and the reference threshold of the search parameter is obtained from the preset function library according to the analysis result and the process parameter.

[0075] Specifically, a process period usually lasts for several seconds to several minutes. In order to realize in-situ and real-time monitoring, the present application further divides the process period into a plurality of continuous detection periods, wherein each detection period corresponds to a sequence of spectral frames in the event window, that is, a plurality of continuous original reflection spectra. The original reflection spectrum is analyzed frame by frame, and the characteristic is extracted and summarized to obtain the index parameter used for evaluation in the current detection period.

[0076] The preset function library refers to a set of mapping relationships or function tables constructed and stored in advance, which is used to associate the process parameters and spectral index parameters to the threshold parameters suitable for the current working condition.

[0077] The preset function library is constructed based on the historical data in the historical database, wherein the historical data includes historical reflection spectrum data, corresponding process parameters, historical analysis results and quality annotation information. Exemplarily, the way of constructing the function library includes: classifying the above-mentioned historical data according to the process parameters, for example, dividing according to the growth layer number, material composition, set temperature and set growth rate to obtain a plurality of categories. In each category, further stratified statistics is carried out according to the index parameter to obtain a plurality of samples. For each group of samples, the threshold range allowed by the search parameter meeting the data requirement is counted, and thus a set of mapping relationships of the reference threshold of the process parameter, the index parameter and the search parameter is obtained.

[0078] In this way, during execution, the reference threshold suitable for the current working condition can be obtained by looking up the preset function library and optionally performing interpolation calculation according to the analysis result obtained by frame-by-frame analysis and the current process parameter in each detection period.

[0079] At step 106, a preset prediction model is used to generate a characteristic wavelength range of the current detection period based on the historical reflectance spectrum and its corresponding process parameters.

[0080] Specifically, the historical reflectance spectrum refers to reflectance spectrum data collected and generated from the semiconductor growth surface reflectance signal during the past operation of the epitaxial growth device, which corresponds to the actual measurement results of one or more previous process periods. The historical reflectance spectrum is usually recorded in the form of wavelength and its corresponding reflectivity, and is stored in association with the process parameters at the time of collection.

[0081] The prediction model is used to make prior prediction of the possible key spectral feature positions in the current detection period, thereby providing a search starting point and a search range for subsequent feature search. Exemplarily, the prediction model can use a machine learning model or a statistical model, such as a time series prediction network of long short-term memory (LSTM) or a regression model based on historical statistics.

[0082] The historical reflectance spectrum and its corresponding process parameters are input into the above-mentioned prediction model, and the prediction model outputs a characteristic wavelength prediction result of the current detection period. The prediction result can be a reference value, a prediction interval or a probability distribution of the characteristic wavelength. Based on the prediction result, a characteristic wavelength range for feature search is further constructed.

[0083] At step 108, based on the characteristic wavelength range, a feature search is performed on the original reflectance spectrum under the constraint of a reference threshold to obtain a fused characteristic wavelength.

[0084] Specifically, each candidate wavelength in the characteristic wavelength range is used as a search starting point or a search center, and the original reflectance spectrum is scanned and candidate features are extracted within a preset search window to obtain possible candidate feature positions, wherein the candidate feature positions include stopband centers, stopband edges, cavity resonance peaks or cavity resonance valleys. The candidate features are fused to obtain the fused characteristic wavelength.

[0085] During the search process, the reference threshold is used as a constraint condition to effectively screen the above-mentioned candidate features, so as to eliminate invalid candidate results.

[0086] At step 110, an effectiveness judgment is made according to the original reflectance spectrum and the fused characteristic wavelength, and when the judgment result meets a preset condition, the fused characteristic wavelength is sent to the main control system of the epitaxial growth device, so that the main control system adjusts the process parameters of the next process period based on the fused characteristic wavelength.

[0087] Specifically, to improve the anti-noise and anti-transient disturbance ability of the closed-loop parameter adjustment, the application takes the results of multiple detection periods in the same process period as the basis for adjustment, and statistically summarizes the fusion characteristic wavelength and the effectiveness judgment result of multiple detection periods to judge the consistency of the detection results. If the fusion characteristic wavelength is within the preset range and the fluctuation amplitude of the fusion characteristic wavelength is within the preset range, it is considered that the characteristic feedback of the process period is reliable, and the confirmed fusion characteristic wavelength is sent to the main control system of the epitaxial growth equipment. The main control system takes the fusion characteristic wavelength as the process feedback quantity, compares it with the target wavelength, and adjusts the process parameters of the next process period accordingly.

[0088] The purpose of effectiveness judgment is to perform reliability screening and gating control on the fusion characteristic wavelength obtained in the current detection period, that is, to determine whether the fusion characteristic wavelength is in an effective state by combining the spectral feature of the original reflectance spectrum and the consistency of the results of each detection period, so as to avoid unreliable characteristic wavelengths from being sent to the main control system of the epitaxial growth equipment and triggering false parameter adjustment.

[0089] The above semiconductor process parameter adjustment method acquires the original reflectance spectrum and process parameters of the current process period, analyzes the original reflectance spectrum frame by frame and combines the process parameters, acquires the reference threshold of the search parameter from the preset function library, reduces the dependence on fixed threshold, and improves the detection stability under complex spectral conditions. Further, based on the historical reflectance spectrum and its corresponding process parameters, a preset prediction model is used to generate a characteristic wavelength range of the current detection period, so that the search range is dynamically updated with process drift, thereby improving the characteristic positioning efficiency and enhancing the anti-interference ability. Further, under the constraint of the reference threshold, the characteristic wavelength range is used to perform feature search on the original reflectance spectrum and output the fusion characteristic wavelength, realizing higher precision and higher repeatability of key wavelength extraction, and providing a reliable basis for process decision. Further, the fusion characteristic wavelength is fed back to the main control system and used for parameter adjustment of the next process period only when the effectiveness judgment meets the preset conditions, forming a reliable closed-loop control to improve the controllability, wavelength consistency and yield of the epitaxial process.

[0090] Optionally, the above semiconductor process parameter adjustment method further comprises:

[0091] In each detection period of the current process period, the original reflectance spectrum is decomposed frame by frame using a Gaussian pyramid to generate a plurality of spectral sets of different scales;

[0092] Based on the characteristic wavelength range, the feature search is performed on the spectral sets of different scales under the constraint of the reference threshold to obtain the fusion characteristic wavelength.

[0093] Specifically, the original reflectance spectrum is further decomposed into a plurality of scale spectrum sets, and due to different sensitivities of scale spectrum to noise and features, low scale spectrum can suppress high frequency noise and more stably reflect global feature profile; high scale spectrum retains complete detail information and is used for accurate positioning. Based on the fusion feature wavelength searched on the above-mentioned spectrum sets of different scales, high-precision and stable detection results can be obtained even in a complex spectral background.

[0094] Optionally, the original reflectance spectrum is decomposed into a plurality of different scale spectrum sets by Gaussian pyramid decomposition frame by frame, including:

[0095] The initial spectrum is subjected to Gaussian kernel smoothing, wherein the initial spectrum is a current frame of the original reflectance spectrum; the initial spectrum after the smoothing is subjected to down-sampling processing to generate a first layer of low resolution spectrum based on a preset down-sampling factor; the first layer of low resolution spectrum is repeatedly subjected to smoothing and down-sampling processing to iteratively generate second layer and above low resolution spectrum; the initial spectrum and each low resolution spectrum constitute a plurality of different scale spectrum sets.

[0096] Specifically, the current frame of the original reflectance spectrum collected at the current time is denoted as an initial spectrum S_0. First, S_0 is subjected to Gaussian kernel smoothing to suppress random noise and local peak interference; then the smoothed spectrum is subjected to down-sampling processing, for example, down-sampling according to a preset down-sampling factor, and a first layer of low resolution spectrum S_1 is obtained. On this basis, S_1 is subjected to the same Gaussian kernel smoothing and down-sampling processing to iteratively generate second layer and above low resolution spectrum, thereby constituting a multi-scale spectrum set {S_0, S_1, S_2,...} in a pyramid structure. In this embodiment, the down-sampling factor is set to 2.

[0097] By using the above scheme, the subsequent feature search process can be cooperatively performed on the spectrum of different scales, for example, the approximate region of the stop band center is quickly located on the low resolution scale, and the feature wavelength position is further determined in detail on the high resolution scale, thereby greatly improving the stability and efficiency of detection.

[0098] Please refer to Figure 2 Optionally, in each detection period of the current process cycle, the original reflectance spectrum is analyzed frame by frame, and according to the analysis result and the process parameter, a reference threshold of the search parameter is obtained from a preset function library, including:

[0099] In step 202, the original reflectance spectrum is analyzed frame by frame in each detection period to obtain the spectral feature parameters corresponding to each frame.

[0100] At step 204, based on the statistical results of the spectral characteristic parameters in the entire detection period, an index parameter of the original reflection spectrum in the preset waveband is determined, wherein the index parameter includes a signal-to-noise ratio estimation value and a smoothness index.

[0101] At step 206, based on the index parameter and the process parameter, table lookup and interpolation calculation are performed in the preset function library to obtain a reference threshold of the search parameter.

[0102] Specifically, in each detection period, a plurality of frames of continuous original reflection spectra are acquired according to a preset sampling frequency, and each frame is analyzed to obtain spectral characteristic parameters for characterizing the spectral quality and waveform stability. Illustratively, analyzing the original reflection spectrum includes: performing waveband interception and wavelength alignment on each frame of spectrum in a preset waveband, and performing background normalization processing, and then extracting the spectral characteristic parameters, including intensity characteristic parameters, noise characteristic parameters and smoothness characteristic parameters. The preset waveband refers to a key waveband near the stop band, which is 850-910 nm in this embodiment.

[0103] Further, the spectral characteristic parameters corresponding to each frame in the detection period are statistically aggregated, and abnormal frames are removed, and then a signal-to-noise ratio estimation value is determined according to the aggregated intensity characteristic parameters and noise characteristic parameters, and a smoothness index is determined according to the aggregated smoothness characteristic parameters, to form the index parameter.

[0104] Based on the index parameter and the process parameter, a query is performed in the preset function library, if the query falls between the discrete nodes of the function table, then the threshold values corresponding to the adjacent nodes are interpolated to obtain a continuous reference threshold matching the current working condition. Illustratively, let the original reflection spectrum at the current time t be S_raw(λ, t), and let the current process parameter be State(t). The signal-to-noise ratio estimation value SNR_est and the smoothness index SM_est of S_raw(λ, t) in the key waveband are calculated, and then table lookup and interpolation calculation are performed in the preset function library in combination with State(t) to output a dynamic parameter set of the reference threshold Params_adapt(t) = {P_adapt, F_adapt, D_adapt}. Wherein, P_adapt is an adaptive peak drop threshold; F_adapt is an adaptive peak filtering threshold, and D_adapt is an adaptive valley rise threshold. According to a preset rule, the value of the reference threshold is dynamically configured according to the index parameter and the process parameter, for example, when it is determined that the signal-to-noise ratio is low, a more conservative threshold value will be automatically adopted, the value of P_adapt is increased from 0.8 to 0.9, so as to reduce the false judgment caused by noise interference.

[0105] With the above embodiment, the frame-level spectral feature parameters are obtained by frame-by-frame analysis and statistically summarized in the detection period, so that the single-frame noise and transient interference can be suppressed without increasing the hardware complexity, and the stability of the spectral quality evaluation is improved. Further, based on the signal-to-noise ratio estimate value and the smoothness index, and in combination with the current process parameters, a reference threshold value is obtained from a preset function library, so that the search threshold value can be adaptively adjusted according to the working condition and the spectral quality, and the optimal parameters are dynamically output, thereby ensuring the scientificity of the threshold setting under different signal qualities and overcoming the inherent defect of performance degradation of the fixed parameters when the signal-to-noise ratio fluctuates.

[0106] Referring to Figure 3 Optionally, based on the historical reflection spectrum and the corresponding process parameters, a preset prediction model is used to predict the characteristic wavelength range of the current detection period, including:

[0107] In step 302, a preset number of historical reflection spectra of process periods and corresponding process parameters are obtained, and a sequence of characteristic wavelengths is extracted from each historical reflection spectrum to form a historical characteristic wavelength input sequence.

[0108] In step 304, the historical characteristic wavelength input sequence and the corresponding process parameters are input into a preset prediction model to output a reference range of the current detection period.

[0109] In step 306, based on the reference range, a preset wavelength list is dynamically updated to generate a characteristic wavelength range for the current detection period.

[0110] Specifically, a preset number of historical reflection spectra and process parameters of process periods are obtained from a historical database, and the historical reflection spectrum of each process period is analyzed to extract characteristic wavelengths related to the device structure, and then the historical characteristic wavelength sequence [λ_bragg(t-N),..., λ_bragg(t-1)] is formed in chronological order.

[0111] Further, the historical characteristic wavelength input sequence and the corresponding process parameters are input into a preset prediction model, which learns the evolution law of the historical characteristic wavelength with the process state change, and outputs the possible interval λ_predict(t) of the characteristic wavelength of the current detection period.

[0112] Further, in each detection period, a preset wavelength list is used as an initial static list, and the initial static list is updated based on the predicted reference range λ_base(t). For example, λ_base(t) and the initial static list are fused to generate an optimized dynamic guess wavelength list G_optimized(t) by weighted average, and the list is used as the characteristic wavelength range, so that the starting point of subsequent feature search can be updated in real time with the process drift.

[0113] With the above scheme, the prediction model can actively predict and correct the search starting point of the characteristic wavelength based on historical data, so that the application can actively track and adapt to the wavelength changes caused by process drift, limit the search to the vicinity of the predicted range, significantly reduce the search space, and reduce the computational overhead caused by full-band blind search.

[0114] Please refer to Figure 4 Optionally, based on the characteristic wavelength range, a feature search is performed on different scale spectrum sets under the constraint of a reference threshold to obtain a fused characteristic wavelength, including:

[0115] Step 402: On a low-resolution spectrum of a low scale, each candidate wavelength in the characteristic wavelength range is taken as an initial center, a candidate peak search and screening are performed in a preset search window according to the reference threshold, to determine a low-scale estimated value of the stopband center.

[0116] Step 404: On a high-resolution spectrum of a high scale, a prior search region is determined based on the low-scale estimated value, a stopband edge detection is performed in the prior search region to determine a high-scale estimated value of the stopband center, a target trough is searched in the prior search region, and a cavity resonance characteristic wavelength of the target trough is calculated.

[0117] Step 406: The low-scale estimated value and the high-scale estimated value are weighted and fused to obtain a stopband center characteristic wavelength; the cavity resonance characteristic wavelength and the stopband center characteristic wavelength constitute the fused characteristic wavelength.

[0118] Specifically, on a low-scale spectrum, the reference threshold is used to quickly lock the approximate region of the stopband center λ_band_low. Taking S_0 as a high-scale spectrum and S_1 as a low-scale spectrum as examples, on the S_1 spectrum, each wavelength g_i in G_optimized(t) is taken as an initial center, a highest reflectivity point is searched in a range of ±25nm using P_adapt and F_adapt threshold rules. If multiple highest points are obtained, invalid peaks are removed based on a preset rule, and then an optimal target peak is selected from the remaining candidate peaks as a low-scale estimated value λ_band_low of the stopband center under a low scale.

[0119] Optionally, a target trough is searched in the prior search region, and a cavity resonance characteristic wavelength of the target trough is calculated, including:

[0120] In the prior search region, a trough search is performed on the high-scale high-resolution spectrum to obtain at least one candidate trough; for each candidate trough, a preset spectral line model is fitted in an interval centered on the candidate trough and having a preset width using a nonlinear least squares fitting algorithm to obtain a trough center wavelength, a spectral line depth, and a fitting goodness of the candidate trough; wherein the preset spectral line model is a parameterized model for representing a corresponding relationship between the candidate spectral line and wavelength and reflectivity; the candidate troughs are screened based on the spectral line depth and the fitting goodness, and the trough center wavelength corresponding to the screened candidate trough is taken as the cavity resonance characteristic wavelength.

[0121] Specifically, on the high-scale spectrum, the above region is used as prior knowledge to accurately locate the stop band edge and the cavity resonance trough using a more refined algorithm. Exemplarily, on the S_0 spectrum, the prior region (λ_band_low ± Δλ) provided by λ_band_low is focused on, a more refined derivative-based accurate zero-crossing detection algorithm is used to perform accurate stop band edge detection to obtain a high-scale estimate of the stop band center λ_band_high. Further, a cavity resonance trough search is also performed in the prior search region to obtain multiple candidate troughs. For each candidate trough, a Gaussian function nonlinear least squares fitting is performed using the Levenberg-Marquardt algorithm in an interval centered on it and having a width of W_fit = N × FWHM_initial (N can be dynamically adjusted according to the spectrum quality) to accurately solve the trough center wavelength λ_dip_high, the depth Depth, and the fitting goodness R². Then, the target trough is selected based on the depth Depth and the fitting goodness R².

[0122] Further, a weighted average algorithm is used to fuse the detection results of each scale. For the stop band center, a weighted average is used for fusion: λ_band_final= w_low×λ_band_low+w_high×λ_band_high, where w_low and w_high are the weights of the low-scale estimate and the high-scale estimate, respectively, and are dynamically assigned according to the fitting goodness or signal quality of the results of each scale.

[0123] By using the above scheme, a multi-scale feature search scheme is used to suppress high-frequency noise in a low-resolution scale, stabilize the capture of global feature profiles, accurately locate details in a high-resolution scale, and integrate the advantages of each scale through a weighted fusion algorithm, so that high-precision and stable detection results can be obtained even in a complex spectral background.

[0124] Please refer to Figure 5 Optionally, the effectiveness is judged according to the original reflectance spectrum and the fused feature wavelength, including:

[0125] Step 502, constructing a two-dimensional feature vector based on the original reflectance spectrum and the fusion feature wavelength.

[0126] Step 504, inputting the two-dimensional feature vector into a preset deep learning model to output a confidence score and a classification label; wherein the confidence score is within a preset numerical interval, and the classification label is used to represent the effectiveness state of the fusion feature wavelength.

[0127] Step 506, determining the judgment result and processing strategy of the fusion feature wavelength in the current detection period based on the confidence score and / or the classification label.

[0128] Specifically, a two-dimensional feature vector is composed based on the original reflectance spectrum S_raw and the fusion feature wavelength (λ_band_final, λ_dip_final). Exemplarily, the deep learning model adopts a CNN model, which is pre-trained based on historical data. The two-dimensional feature vector is input into the CNN model to output a confidence score C_score and / or a classification label.

[0129] Optionally, the preset numerical interval is between 0-1; the classification label includes normal, suspicious and failure; based on the confidence score and / or the classification label, the judgment result and processing strategy of the fusion feature wavelength in the current detection period are determined, including:

[0130] In the case that the confidence score is greater than or equal to the first threshold value and the classification label is normal, it is considered that the preset condition is met, and the fusion feature wavelength is output to the main control system of the epitaxial growth device.

[0131] In the case that the confidence score is less than or equal to the second threshold value or the classification label is suspicious, it is considered that the preset condition is not met, a prompt information is generated, and the prompt information and the fusion feature wavelength are output to the main control system of the epitaxial growth device.

[0132] In the case that the confidence score is less than the second threshold value or the classification label is failure, it is considered that the preset condition is not met, an alarm signal is generated, and the alarm signal is output to the main control system of the epitaxial growth device; wherein the second threshold value is less than the first threshold value.

[0133] Exemplarily, the first threshold value in the embodiment is 0.8, and the second threshold value is 0.5.

[0134] If C_score ≥ 0.8 and the classification label is "normal", the fusion feature wavelength (λ_band_final, λ_dip_final) calculated this time is taken as a high-reliability result and immediately sent to the epitaxial growth control system.

[0135] If 0.5≤ C_score<0.8 or the classification label is "suspicious", the fusion feature wavelength is outputted while a prompt information is sent.

[0136] If C_score<0.5 or the classification label is "invalid", the result output is suspended, and an audible and light alarm is triggered. Further, according to the configuration, the last valid value can be used or a safe waiting mode can be entered.

[0137] With the above scheme, an objective confidence score is outputted each time the feature wavelength is detected, so that the downstream process control system can adopt differentiated strategies according to the confidence score, for example, high confidence results are used for closed-loop control, and low confidence results trigger alarms, thereby avoiding the risk of directly driving process adjustment by false detection data.

[0138] Further, the trained deep learning model is used to jointly analyze the original spectrum and the judgment result, identify problems that are difficult to determine by traditional algorithms such as spectral anomalies and fitting distortion, and output quantitative confidence evaluation, thereby providing a reliable judgment basis for process safety and greatly improving the reliability of automatic process control based on the judgment result.

[0139] Optionally, the above semiconductor process parameter adjustment method further comprises:

[0140] After each detection cycle ends, detection data generated in the current detection cycle is acquired; the detection data is added with a time sequence identifier, and after compression processing, the detection data added with the time sequence identifier is stored in a historical database; newly added detection data is periodically extracted from the historical database as an incremental training sample, and the prediction model and the deep learning model are incrementally trained and updated based on the incremental training sample.

[0141] With the above scheme, the detection data from the original spectrum, dynamic parameters, intermediate results of each scale to the final verification conclusion in each detection cycle are synchronously stored in the historical database. These detection data provide data asset value that cannot be matched by traditional single output systems for subsequent deep analysis of the influence of process fluctuations on spectral features, tracing the root cause of abnormal batches, and further optimizing the epitaxial growth model. At the same time, the above detection data are also used for continuous self-optimization of the prediction model and the deep learning model, thereby further improving the detection accuracy of the present application.

[0142] It should be understood that although the steps in the flowcharts involved in the embodiments described above are shown in sequence according to the arrows, the steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, the execution of the steps is not strictly limited in sequence, and the steps can be executed in other sequences. Moreover, at least some of the steps in the flowcharts involved in the embodiments described above can include multiple steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution sequence of the steps or stages is not necessarily sequential, but can be executed alternately or alternately with at least part of other steps or steps or stages in other steps.

[0143] Based on the same inventive concept, the embodiments of the present application also provide a semiconductor process parameter adjustment system, which is suitable for the semiconductor process parameter adjustment method described above. The implementation scheme for solving the problem provided by the system is similar to the implementation scheme described in the above method. Therefore, the specific limitations in one or more device embodiments provided below can refer to the limitations of the method described above, and will not be repeated here.

[0144] Please refer to Figure 6 In one embodiment, the present application also provides a semiconductor process parameter adjustment system, comprising: an acquisition module, an adaptive threshold generation module, an intelligent optimization module, a search module, a verification module and an output module.

[0145] The acquisition module is configured to acquire an original reflectance spectrum and a process parameter of a current process cycle, wherein the original reflectance spectrum is generated based on a reflected light signal generated during a semiconductor epitaxial growth process.

[0146] The adaptive threshold generation module is configured to analyze the original reflectance spectrum frame by frame in each detection cycle of the current process cycle, and query a reference threshold of a search parameter from a preset function library according to an analysis result and the process parameter.

[0147] The intelligent optimization module is configured to generate a characteristic wavelength range of a current detection cycle based on a historical reflectance spectrum and a corresponding process parameter thereof using a preset prediction model.

[0148] The search module is configured to perform feature search on the original reflectance spectrum under the constraint of the reference threshold based on the characteristic wavelength range, to obtain a fused characteristic wavelength.

[0149] The verification module is configured to perform effectiveness judgment according to the original reflectance spectrum and the fused characteristic wavelength.

[0150] The output module is configured to send the fused characteristic wavelength to a main control system of the epitaxial growth device when the determination result meets the preset condition, so that the main control system adjusts process parameters of a next process cycle based on the fused characteristic wavelength.

[0151] Optionally, the semiconductor process parameter adjustment system further includes a spectrum decomposition module.

[0152] The spectrum decomposition module is configured to perform Gaussian pyramid decomposition on the original reflection spectrum frame by frame to generate a plurality of spectrum sets of different scales in each detection period of the current process cycle.

[0153] The search module is further configured to perform feature search on the spectrum sets of different scales under the constraint of the reference threshold based on the characteristic wavelength range to obtain the fused characteristic wavelength.

[0154] Optionally, the search module performs Gaussian pyramid decomposition on the original reflection spectrum frame by frame to generate a plurality of spectrum sets of different scales, including: performing Gaussian kernel smoothing processing on an initial spectrum, wherein the initial spectrum is a current frame of the original reflection spectrum; performing down-sampling processing on the initial spectrum after the smoothing processing to generate a low-resolution spectrum of a first layer based on a preset down-sampling factor; repeatedly performing the smoothing processing and the down-sampling processing based on the low-resolution spectrum of the first layer to iteratively generate low-resolution spectrums of a second layer and above; and the initial spectrum and the low-resolution spectrums constitute the plurality of spectrum sets of different scales.

[0155] Optionally, the adaptive threshold generation module analyzes the original reflection spectrum frame by frame in each detection period of the current process cycle, and obtains the reference threshold of the search parameter from a preset function library according to an analysis result and a process parameter, including: analyzing the original reflection spectrum frame by frame to obtain spectrum feature parameters corresponding to each frame in each detection period; determining an index parameter of the original reflection spectrum in a preset wavelength band based on a statistical result of the spectrum feature parameters in the entire detection period, wherein the index parameter includes a signal-to-noise ratio estimate value and a smoothness index; and performing table lookup and interpolation calculation in the preset function library based on the index parameter and the process parameter to obtain the reference threshold of the search parameter.

[0156] Optionally, the intelligent optimization module predicts a characteristic wavelength range of a current detection period based on historical reflection spectrums and corresponding process parameters thereof using a preset prediction model, including: obtaining historical reflection spectrums of a preset number of process cycles and corresponding process parameters thereof, and extracting a feature wavelength sequence from each historical reflection spectrum to form a historical feature wavelength input sequence; inputting the historical feature wavelength input sequence and the corresponding process parameters into the preset prediction model to output a reference range of the current detection period; and based on the reference range, dynamically updating a preset wavelength list to generate the characteristic wavelength range for the current detection cycle.

[0157] Optionally, the search module performs feature search on the spectrum set of different scales under the constraint of the reference threshold based on the feature wavelength range, to obtain the fusion feature wavelength, including: performing candidate peak search and screening in a preset search window according to the reference threshold, with each candidate wavelength in the feature wavelength range as an initial center, on a low-resolution spectrum of a low scale, to determine a low-scale estimation value of the stopband center; determining a prior search region based on the low-scale estimation value, performing stopband edge detection in the prior search region to determine a high-scale estimation value of the stopband center, and searching for a target trough in the prior search region to calculate a cavity resonance feature wavelength of the target trough, on a high-resolution spectrum of a high scale; performing weighted fusion on the low-scale estimation value and the high-scale estimation value to obtain a stopband center feature wavelength; and the cavity resonance feature wavelength and the stopband center feature wavelength constitute the fusion feature wavelength.

[0158] Optionally, the search module searches for a target trough in the prior search region to calculate a cavity resonance feature wavelength of the target trough, including: performing trough search on a high-resolution spectrum of a high scale in the prior search region to obtain at least one candidate trough; for each candidate trough, performing fitting on a preset spectral line model in an interval with a preset width centered on the candidate trough using a nonlinear least squares fitting algorithm to obtain a trough center wavelength, a spectral line depth, and a fitting goodness of the candidate trough; wherein the preset spectral line model is a parameterized model for representing a corresponding relationship between a candidate spectral line and wavelength and reflectivity; screening the candidate troughs based on the spectral line depth and the fitting goodness, and taking a trough center wavelength corresponding to a screened candidate trough as the cavity resonance feature wavelength.

[0159] Optionally, the verification module performs effectiveness judgment according to the original reflectance spectrum and the fusion feature wavelength, including: constructing a two-dimensional feature vector based on the original reflectance spectrum and the fusion feature wavelength; inputting the two-dimensional feature vector into a preset deep learning model to output a confidence score and a classification label; wherein the confidence score is in a preset numerical interval, and the classification label is used to represent an effectiveness state of the fusion feature wavelength; and determining a judgment result of the fusion feature wavelength in a current detection period based on the confidence score and / or the classification label.

[0160] Optionally, the preset numerical interval is between 0 and 1; the classification label includes normal, suspicious and failure cases; and the verification module determines the judgment result and the processing strategy of the fusion feature wavelength in the current detection period based on the confidence score and / or the classification label, including: in the case that the confidence score is greater than or equal to the first threshold value and the classification label is normal, it is considered that the preset condition is met, and the fusion feature wavelength is output to the main control system of the epitaxial growth equipment; in the case that the confidence score is less than or equal to the second threshold value or the classification label is suspicious, it is considered that the preset condition is not met, a prompt information is generated, and the prompt information and the fusion feature wavelength are output to the main control system of the epitaxial growth equipment; in the case that the confidence score is less than the second threshold value or the classification label is failure, it is considered that the preset condition is not met, an alarm signal is generated, and the alarm signal is output to the main control system of the epitaxial growth equipment; wherein the second threshold value is less than the first threshold value.

[0161] Optionally, the semiconductor process parameter adjustment system further comprises an updating module.

[0162] The updating module is configured to, at the end of each detection period, acquire detection data generated in the current detection period; add a time sequence identifier to the detection data, and store the detection data after compression processing to a historical database; periodically extract newly added detection data from the historical database as an incremental training sample, and perform incremental training on the prediction model and the deep learning model based on the incremental training sample and update the prediction model and the deep learning model.

[0163] The semiconductor process parameter adjustment system acquires the original reflection spectrum and the process parameter of the current process period, analyzes the original reflection spectrum frame by frame and combines the process parameter, acquires the reference threshold value of the search parameter from a preset function library, reduces the dependence on the fixed threshold value, and improves the detection stability under complex spectrum conditions. Further, based on the historical reflection spectrum and the corresponding process parameter, a preset prediction model is used to generate a feature wavelength range of the current detection period, so that the search range is dynamically updated with process drift, thereby improving the feature positioning efficiency and enhancing the anti-interference ability. Further, under the constraint of the reference threshold value, the feature wavelength range is used to perform feature search on the original reflection spectrum and output the fusion feature wavelength, realizing higher precision and higher repeatability of key wavelength extraction, and providing a reliable basis for process decision. Further, the fusion feature wavelength is fed back to the main control system and used for parameter adjustment of the next process period only when the effectiveness judgment meets the preset condition, forming a reliable closed-loop control to improve the controllability, wavelength consistency and yield of the epitaxial process.

[0164] The modules in the semiconductor process parameter adjustment system can be implemented by software, hardware, or a combination thereof. The modules can be embedded in a processor in a computer device or independent of the processor, or stored in a memory in the computer device to be invoked by the processor to perform the operations corresponding to the modules.

[0165] In an embodiment, a computer device, which can be a terminal, has an internal structure as shown in Figure 7 The computer device includes a processor, a memory, an input / output interface, a communication interface, a display unit, and an input device. The processor, the memory, and the input / output interface are connected through a system bus, and the communication interface, the display unit, and the input device are connected to the system bus through the input / output interface. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for running the operating system and the computer program in the non-volatile storage medium. The input / output interface of the computer device is configured to exchange information between the processor and external devices. The communication interface of the computer device is configured to perform wired or wireless communication with external terminals. The wireless communication can be achieved by WIFI, mobile cellular network, NFC (Near Field Communication), or other technologies. The computer program is executed by the processor to implement the semiconductor process parameter adjustment method. The display unit of the computer device is configured to form a visually visible screen, which can be a display screen, a projection device, or a virtual reality imaging device. The display screen can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or a key, trackball, or touchpad arranged on the housing of the computer device, or an external keyboard, touchpad, or mouse, etc.

[0166] Those skilled in the art can understand that Figure 7 The structure shown in the above embodiment is only a block diagram of part of the structure related to the scheme of the present application, and does not limit the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0167] In an embodiment, a computer device includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method steps in the semiconductor process parameter adjustment method.

[0168] In an implementable embodiment, a computer readable storage medium is provided, and a computer program is stored on the computer readable storage medium, and the computer program is executed by a processor to implement the method steps in the method for adjusting semiconductor process parameters.

[0169] In an implementable embodiment, a computer program product is provided, and the computer program product comprises a computer program, and the computer program is executed by a processor to implement the method steps in the method for adjusting semiconductor process parameters.

[0170] The technical features of the above embodiments can be combined in any manner. To make the description concise, not all possible combinations of the technical features in the above embodiments are described, but it should be considered that any combination of the technical features is within the scope of the present disclosure as long as the combination does not result in contradictions.

[0171] The above embodiments only express several implementation manners of the present application, and the description is relatively specific and detailed, but it should not be considered as a limitation on the patent scope of the present application. It should be pointed out that, for those skilled in the art, some modifications and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. A method for adjusting semiconductor process parameters, characterized in that, The method comprises: acquiring an original reflectance spectrum and process parameters of a current process cycle; wherein the original reflectance spectrum is generated based on a reflected light signal generated in a semiconductor epitaxial growth process; in each detection cycle of the current process cycle, frame-by-frame analyzing the original reflectance spectrum, and according to the analysis result and the process parameters, querying a reference threshold of a search parameter from a preset function library; based on historical reflectance spectra and their corresponding process parameters, generating a characteristic wavelength range of the current detection cycle using a preset prediction model; based on the characteristic wavelength range, performing feature search on the original reflectance spectrum under the constraint of the reference threshold to obtain a fused characteristic wavelength; performing effectiveness judgment according to the original reflectance spectrum and the fused characteristic wavelength, and when the judgment result meets a preset condition, sending the fused characteristic wavelength to a main control system of an epitaxial growth device, so that the main control system adjusts process parameters of a next process cycle based on the fused characteristic wavelength.

2. The method of claim 1, wherein, The method further comprises: in each detection cycle of the current process cycle, performing Gaussian pyramid decomposition on the original reflectance spectrum frame by frame to generate a plurality of spectrum sets of different scales; based on the characteristic wavelength range, performing feature search on the spectrum sets of different scales under the constraint of the reference threshold to obtain a fused characteristic wavelength.

3. The method of claim 2, wherein, The Gaussian pyramid decomposition on the original reflectance spectrum frame by frame to generate a plurality of spectrum sets of different scales comprises: performing Gaussian kernel smoothing on an initial spectrum, wherein the initial spectrum is a current frame of the original reflectance spectrum; performing downsampling on the initial spectrum after smoothing to generate a low-resolution spectrum of a first layer based on a preset downsampling factor; based on the low-resolution spectrum of the first layer, repeatedly performing smoothing and downsampling to iteratively generate low-resolution spectra of a second layer and above; the initial spectrum and the low-resolution spectra constitute a plurality of spectrum sets of different scales.

4. The method according to claim 1 or 2, characterized in that, The process parameters of the current process cycle include at least one of a growth layer number, a material composition, a set temperature, and a set growth rate; The frame-by-frame analysis of the original reflectance spectrum in each detection cycle of the current process cycle, and the querying of the reference threshold of the search parameter from the preset function library according to the analysis result and the process parameters, comprise: in each detection cycle, frame-by-frame analyzing the original reflectance spectrum to obtain spectrum feature parameters corresponding to each frame; based on the statistical result of the spectrum feature parameters in the entire detection cycle, determining an index parameter of the original reflectance spectrum in a preset waveband, wherein the index parameter includes a signal-to-noise ratio estimate value and a smoothness index; based on the index parameter and the process parameters, performing table lookup and interpolation calculation in the preset function library to obtain the reference threshold of the search parameter.

5. The method according to claim 1 or 2, characterized in that, The prediction of the characteristic wavelength range of the current detection cycle based on the historical reflectance spectra and their corresponding process parameters using the preset prediction model comprises: acquiring historical reflectance spectra of a preset number of process cycles and their corresponding process parameters, and extracting a feature wavelength sequence from each of the historical reflectance spectra to form a historical feature wavelength input sequence; inputting the historical characteristic wavelength input sequence and corresponding process parameters into the preset prediction model to output a reference range of a current detection period; based on the reference range, dynamically updating a preset wavelength list to generate a characteristic wavelength range for the current detection period.

6. The method of claim 3, wherein, based on the characteristic wavelength range, performing characteristic search on different scale spectrum sets under the constraint of the reference threshold to obtain a fusion characteristic wavelength, including: on a low-resolution spectrum of a low scale, taking each candidate wavelength in the characteristic wavelength range as an initial center, performing candidate peak search and screening within a preset search window according to the reference threshold to determine a low-scale estimation value of a stopband center; on a high-resolution spectrum of a high scale, determining a prior search region based on the low-scale estimation value, performing stopband edge detection in the prior search region to determine a high-scale estimation value of the stopband center, and searching in the prior search region to obtain a target trough to calculate a cavity resonance characteristic wavelength of the target trough; weighting the low-scale estimation value and the high-scale estimation value to obtain a stopband center characteristic wavelength; the cavity resonance characteristic wavelength and the stopband center characteristic wavelength constitute the fusion characteristic wavelength.

7. The method of claim 6, wherein, the searching in the prior search region to obtain a target trough to calculate a cavity resonance characteristic wavelength of the target trough, including: in the prior search region, performing trough search on the high-resolution spectrum of the high scale to obtain at least one candidate trough; for each candidate trough, in an interval with a preset width centered on the candidate trough, a nonlinear least squares fitting algorithm is used to fit a preset spectral line model to obtain a trough center wavelength, a spectral line depth, and a fitting goodness of the candidate trough; wherein the preset spectral line model is a parameterized model for representing the corresponding relationship between the candidate spectral line and the wavelength and reflectivity; based on the spectral line depth and the fitting goodness, the candidate trough is screened, and the trough center wavelength corresponding to the screened candidate trough is taken as the cavity resonance characteristic wavelength.

8. The method of claim 1 or 2, wherein, the effectiveness judgment according to the original reflectance spectrum and the fusion characteristic wavelength, including: based on the original reflectance spectrum and the fusion characteristic wavelength, a two-dimensional feature vector is constructed; inputting the two-dimensional feature vector into a preset deep learning model to output a confidence score and a classification label; wherein the confidence score is in a preset numerical interval, and the classification label is used to represent the effectiveness state of the fusion characteristic wavelength; based on the confidence score and / or the classification label, determining the judgment result and the processing strategy of the fusion characteristic wavelength in the current detection period.

9. The method of claim 8, wherein, the preset numerical interval is between 0 and 1; the classification label includes normal, suspicious, and failure; the determination of the judgment result and the processing strategy of the fusion characteristic wavelength in the current detection period based on the confidence score and / or the classification label, including: in the case that the confidence score is greater than or equal to a first threshold and the classification label is normal, it is considered that the preset condition is met, and the fusion characteristic wavelength is output to the main control system of the epitaxial growth device; In a case that the confidence score is less than or equal to a second threshold value or the classification label is suspicious, it is considered that the preset condition is not met, a prompt information is generated, and the prompt information and the fusion feature wavelength are output to a main control system of the epitaxial growth device; In a case that the confidence score is less than a second threshold value or the classification label is failure, it is considered that the preset condition is not met, an alarm signal is generated, and the alarm signal is output to a main control system of the epitaxial growth device; wherein the second threshold value is less than the first threshold value.

10. The method of claim 8, wherein, The method further comprises: After each detection period ends, detection data generated in a current detection period is acquired; The detection data is added with a time sequence identifier, and after compression processing of the detection data added with the time sequence identifier, the detection data is stored to a historical database; Periodically, newly added detection data is extracted from the historical database as an incremental training sample, and the prediction model and the deep learning model are incrementally trained and updated based on the incremental training sample.

11. A semiconductor process parameter adjustment system, characterized in that, The system comprises: An acquisition module is configured to acquire an original reflectance spectrum and process parameters of a current process period; wherein the original reflectance spectrum is generated based on a reflectance signal generated in a semiconductor epitaxial growth process; An adaptive threshold generation module is configured to analyze the original reflectance spectrum frame by frame in each detection period of the current process period, and obtain a reference threshold value of a search parameter from a preset function library according to an analysis result and the process parameters; An intelligent optimization module is configured to generate a feature wavelength range of a current detection period based on a historical reflectance spectrum and corresponding process parameters by using a preset prediction model; A search module is configured to perform feature search on the original reflectance spectrum under the constraint of the reference threshold value based on the feature wavelength range, and obtain a fusion feature wavelength; A verification module is configured to perform effectiveness judgment according to the original reflectance spectrum and the fusion feature wavelength; An output module is configured to send the fusion feature wavelength to a main control system of an epitaxial growth device when a judgment result meets a preset condition, so that the main control system adjusts process parameters of a next process period based on the fusion feature wavelength.

12. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that, The processor executes the computer program to implement the steps of the method of any one of claims 1-10.

13. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method of any one of claims 1-10.

Images