Semiconductor material multispectral optical characterization system and method based on deep learning

Abstract

The invention discloses a semiconductor material multispectral optical characterization system and method based on deep learning. In a light path system of the system, a light source irradiates a semiconductor material sample to excite optical response; the light path beam splitter receives light provided by the light source, reflects the light to the surface of the semiconductor material sample, and transmits the light reflected by the semiconductor material sample to the spectrograph; the spectrograph modulates and processes the received optical signals and then transmits the optical signals to the detector; and the detector converts a received optical signal into an electric signal. A data processing module in the computer system acquires multispectral data of the semiconductor material sample through a multispectral imaging technology according to the electric signal; and carrying out feature extraction and analysis on the multispectral data by the deep learning model, and outputting a multispectral optical characterization result. Automatic analysis and feature extraction of spectral data are realized by using an integrated optical measurement technology and an artificial intelligence algorithm, and the measurement accuracy and the measurement efficiency of semiconductor material characteristics are improved.

Description

Technical Field

[0001] This application relates to the technical field of optical characterization of semiconductor materials, and particularly to a multi-spectral optical characterization system and method for semiconductor materials based on deep learning. Background Art

[0002] With the wide application of semiconductor materials in the fields of electronics, optoelectronics and new energy, the accurate measurement of the physical and chemical properties of semiconductor materials has become crucial. At present, traditional spectral analysis methods mainly rely on experienced researchers to interpret data, which means that the analysis results of traditional spectral analysis methods are greatly affected by human factors and are less efficient when facing large-scale data. Therefore, the system and method for realizing automatic analysis and feature extraction of spectral data to improve the measurement accuracy and efficiency of semiconductor material properties are hot research directions in this field. Summary of the Invention

[0003] Based on the above problems, this application provides a multi-spectral optical characterization system and method for semiconductor materials based on deep learning. The purpose is to realize automatic analysis and feature extraction of spectral data through the optical measurement technology and artificial intelligence algorithm integrated in the system, realize high-precision optical characterization of semiconductor materials, and improve the measurement accuracy and efficiency of semiconductor material properties.

[0004] The embodiments of this application disclose the following technical solutions:

[0005] The first aspect of this application provides a multi-spectral optical characterization system for semiconductor materials based on deep learning. The system includes: an optical path system and a computer system; the optical path system includes: a light source, an optical path beam splitter, a sample stage, a spectrometer and a detector; the computer system is integrated with a data processing module and a deep learning model, and the output end of the data processing module is connected to the input end of the deep learning model; the detector is electrically connected to the computer system; the light source and the sample stage are located on the first side of the reflection surface of the optical path beam splitter; the spectrometer is located on the second side of the reflection surface;

[0006] The sample stage is used to carry the semiconductor material sample to be measured;

[0007] The light source is used to provide illumination to stimulate the optical response of the semiconductor material sample;

[0008] The optical path beam splitter is used to receive the light provided by the light source and reflect the light to the surface of the semiconductor material, and is also used to transmit the light reflected by the semiconductor material sample to the spectrometer;

[0009] The spectrometer is used to modulate and process the received optical signal and then transmit it to the detector;

[0010] The detector is used to convert the optical signal received by itself into an electrical signal, and then transmit the electrical signal to the data processing module;

[0011] The data processing module is used to obtain the multi-spectral data of the semiconductor material sample through multi-spectral imaging technology according to the electrical signal, and input the multi-spectral data into the deep learning model;

[0012] The deep learning model is used to extract and analyze features from the multi-spectral data, and output the multi-spectral optical characterization result of the semiconductor material sample; the multi-spectral optical characterization result includes at least one of a defect detection result, a composition analysis result, or a structure characterization result.

[0013] In an alternative implementation, the deep learning model is trained based on a labeled semiconductor material spectral image dataset, and the semiconductor material spectral image dataset includes spectral images corresponding to multiple different wavelengths; the annotation information of the spectral images includes at least one of the following: defect detection annotation, composition analysis annotation, or structure characterization annotation.

[0014] In an alternative implementation, the deep learning model is trained based on a labeled first spectral image dataset by means of transfer learning on the basis of a target model that has been preliminarily trained; the target model is trained based on a labeled second spectral image dataset;

[0015] Wherein, the second spectral image dataset includes spectral images of one or more semiconductor materials at multiple different wavelengths;

[0016] The first spectral image dataset also includes spectral images of one or more semiconductor materials at multiple different wavelengths; the type of at least one semiconductor material involved in the first spectral image dataset is not covered by the type of semiconductor materials involved in the second spectral image dataset;

[0017] The annotation information of the spectral images in the first spectral image dataset and the second spectral image dataset includes at least one of the following: defect detection annotation, composition analysis annotation, or structure characterization annotation.

[0018] In an alternative implementation, the deep learning model includes: a shared feature extraction module and a multi-task branch module; the output end of the shared feature extraction module is connected to the input end of the multi-task branch module; the multi-task branch includes a defect detection branch, a composition analysis branch, and a structure characterization branch;

[0019] The shared feature extraction module is used to extract features from the multi-spectral data input to the deep learning model;

[0020] The defect detection branch is used to analyze the features output by the shared feature extraction module by means of upsampling and convolutional blocks, and output a defect mask;

[0021] The component analysis branch is used to learn the features output by the shared feature extraction module by means of a regression algorithm, and output the regional component content distribution;

[0022] The structure characterization branch is used to analyze the features output by the shared feature extraction module, and output the classification result of the lattice orientation or stress distribution, or output the regression result of the lattice orientation or stress distribution.

[0023] In an alternative implementation, the shared feature extraction module includes introducing multi-scale convolutional kernels in low-level feature processing for adapting to different spectral feature sizes; the shared feature extraction module also introduces a channel attention mechanism to strengthen the features of key bands.

[0024] In an alternative implementation, during the training process of the deep learning model, the loss weights of each task in the loss function are dynamically adjusted based on the importance of different tasks.

[0025] In an alternative implementation, the output end of the deep learning model is connected to an input end of the data processing module; the computer system further includes a visualization interface;

[0026] The data processing module is used to sort, statistically analyze, and quantitatively analyze the multi-spectral optical characterization results, generate a visualization chart of the semiconductor material sample; standardize the visualization chart and related analysis data into a multi-spectral optical characterization report; the visualization chart includes one or more of a defect distribution map, a component heat map, and a strain analysis map; the content of the multi-spectral optical characterization report covers one or more of detection methods, sample information, detailed analysis results, or comprehensive conclusions;

[0027] The visualization interface is used to display the information in the multi-spectral optical characterization report.

[0028] In an alternative implementation, the multi-spectral optical characterization system for semiconductor materials based on deep learning further includes: a chopper; the chopper is located between the light source and the optical path beam splitter, and the chopper is also located on the first side of the reflecting surface;

[0029] The chopper is used to perform pulse modulation on the light beam output by the light source to convert a continuous optical signal into a pulsed optical signal.

[0030] In an alternative implementation, the multi-spectral optical characterization system for semiconductor materials based on deep learning further includes: a lens assembly; the lens assembly is located between the light source and the chopper;

[0031] The lens assembly is used to modulate the light beam output by the light source into parallel light.

[0032] In an alternative implementation, the data processing module is further configured to perform preprocessing operations on the multispectral data after obtaining the multispectral data; the preprocessing operations include at least one of the following:

[0033] Denoising processing, normalization processing, image background correction, or size adjustment.

[0034] In an alternative implementation, the light source is a xenon lamp, and the xenon lamp supports providing light in a wavelength range of 200 nm to 2500 nm for exciting the optical responses of different types of semiconductor materials.

[0035] In an alternative implementation, the computer system includes a control module, the sample stage is driven by a stepper motor with a positioning accuracy of 1 μm, and the stepper motor is connected to the control module;

[0036] The control module is configured to send a driving signal to the stepper motor to move the sample stage in any one or more of the x-direction, y-direction, or z-direction.

[0037] The second aspect of the present application provides a method for multispectral optical characterization of semiconductor materials based on deep learning, and this method can be applied to the multispectral optical characterization system of semiconductor materials based on deep learning described in any implementation manner of the first aspect. This method includes:

[0038] Turn on the light source;

[0039] Reflect the light provided by the light source to the surface of the semiconductor material sample to be measured on the sample stage through an optical path beam splitter;

[0040] Transmit the light reflected by the semiconductor material sample through the optical path beam splitter and transmit it to the spectrometer;

[0041] The spectrometer modulates and processes the received light and then transmits it to the detector;

[0042] The detector converts the received optical signal into an electrical signal and then transmits the electrical signal to the data processing module;

[0043] The data processing module obtains the multispectral data of the semiconductor material sample through multispectral imaging technology according to the electrical signal and inputs the multispectral data into the deep learning model;

[0044] Feature extraction and analysis are performed on the multispectral data through the deep learning model, and the multispectral optical characterization results of the semiconductor material sample are output; the multispectral optical characterization results include at least one of defect detection results, composition analysis results, or structural characterization results.

[0045] Compared with the prior art, the present application has the following beneficial effects:

[0046] The present application discloses a multispectral optical characterization system and method for semiconductor materials based on deep learning. In the optical path system of the system, a light source irradiates a semiconductor material sample to excite an optical response; an optical path beam splitter receives the light provided by the light source and reflects the light to the surface of the semiconductor material sample, and transmits the light reflected by the semiconductor material sample to a spectrometer; the spectrometer modulates and processes the received optical signal and then transmits it to a detector; the detector converts the optical signal received by itself into an electrical signal. In the computer system, a data processing module obtains the multispectral data of the semiconductor material sample through multispectral imaging technology; a deep learning model performs feature extraction and analysis on the multispectral data and outputs the multispectral optical characterization results. By using the integrated optical measurement technology and artificial intelligence algorithm, automatic analysis and feature extraction of spectral data are realized, and the measurement accuracy and measurement efficiency of semiconductor material characteristics are improved. Description of the Drawings

[0047] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0048] Figure 1 It is a schematic structural diagram of a multispectral optical characterization system for semiconductor materials based on deep learning provided by an embodiment of the present application;

[0049] Figure 2 It is a schematic structural diagram of a deep learning model provided by an embodiment of the present application;

[0050] Figure 3 It is a schematic structural diagram of another multispectral optical characterization system for semiconductor materials based on deep learning provided by an embodiment of the present application;

[0051] Figure 4 It is a schematic structural diagram of yet another multispectral optical characterization system for semiconductor materials based on deep learning provided by an embodiment of the present application;

[0052] Figure 5Schematic structural diagram of yet another multi - spectral optical characterization system for semiconductor materials based on deep learning provided by the embodiments of the present application;

[0053] Figure 6 Flowchart of a multi - spectral optical characterization method for semiconductor materials based on deep learning provided by the embodiments of the present application. Detailed implementation manners

[0054] In many fields such as scientific research experiments, industrial inspections, and semiconductor manufacturing, there are relatively high requirements for the intelligent level, automation level, and accuracy of the optical characterization of semiconductor materials. The current traditional approach relies on experienced researchers to interpret spectral data. Interpreting spectral data manually is inevitably subject to excessive subjectivity and insufficient objectivity, and the richness of experience is likely to directly affect the accuracy of interpretation. In addition, when a large amount of data needs to be interpreted for the optical characterization of semiconductor materials, the efficiency of manual processing is low, resulting in the material measurement efficiency being difficult to meet the measurement requirements.

[0055] Regarding the above problems, the inventors have provided a system that integrates optical measurement technology and artificial intelligence algorithms through research, that is, a multi - spectral optical characterization system for semiconductor materials based on deep learning. In addition, the present application further proposes a multi - spectral optical characterization method for semiconductor materials based on deep learning. The system includes an optical path system and a computer system. The optical path system mainly uses a light source to emit light to stimulate the optical response of a semiconductor material sample to be measured, realizes the deflection of the optical path and the transmission of light through an optical path beam splitter, and collects the light reflected from the surface of the semiconductor material sample into a spectrometer. The spectrometer modulates and processes the light, and finally a detector performs the conversion of photoelectric signals. The computer system integrates a data processing module and a deep learning model. Among them, the data processing module obtains multi - spectral data of the semiconductor material sample through multi - spectral imaging technology based on the electrical signal and inputs it into the deep learning model. The deep learning model then uses artificial intelligence technology to automatically extract features and analyze the multi - spectral data, and outputs multi - spectral optical characterization results such as defect detection results, composition analysis results, and structural characterization results. In this way, through the multi - spectral optical characterization system for semiconductor materials based on deep learning, automatic analysis and feature extraction of multi - spectral data can be performed, and objective and intelligent analysis can achieve high - precision optical characterization of semiconductor materials, improving the measurement accuracy and measurement efficiency of semiconductor material characteristics.

[0056] To enable those skilled in the art to better understand the solution of this application, the technical solutions in the embodiments of this application will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of this application. Obviously, the described embodiments are only a part of the embodiments of this application, rather than all the embodiments. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without making creative efforts belong to the scope of protection of this application.

[0057] See Figure 1 , which is a schematic structural diagram of a multi-spectral optical characterization system for semiconductor materials based on deep learning provided by an embodiment of this application. As Figure 1 shown, the multi-spectral optical characterization system for semiconductor materials based on deep learning includes an optical path system 01 and a computer system 02.

[0058] As Figure 1 shown, the optical path system 01 includes: a light source 11, an optical path beam splitter 12, a sample stage 13, a spectrometer 14, and a detector 15. The computer system 02 is integrated with a data processing module 21 and a deep learning model 22. In Figure 1 it, the red lines represent the optical paths, and the green lines represent the transmission of electrical data. The output end of the data processing module 21 is connected to the input end of the deep learning model 22, and the data processing module 21 can call the deep learning model 22 to perform deep learning operation and analysis on the pre-processed data at the front end.

[0059] In addition, in a possible implementation manner, the output end of the deep learning model 22 can also be connected to an input end of the data processing module 21, so that the data processing module 21 can also further process the output result of the deep learning model 22. In the following text, the further processing method of the output result of the deep learning model 22 by the data processing module 21 will be introduced in detail.

[0060] In the entire multi-spectral optical characterization system for semiconductor materials based on deep learning, the optical path system 01 and the computer system 02 are electrically connected. Among them, the detector 15, as the output component of the electrical signal of the optical path system 01, is specifically electrically connected to the computer system 02. As Figure 1 shown, in the computer system 02, the data processing module 21 first receives the electrical signal transmitted from the front end, so the detector 15 is electrically connected to the data processing module 21 in this computer system 02.

[0061] In the multi-spectral optical characterization system for semiconductor materials based on deep learning, the optical path beam splitter 12 has the function of reflecting and transmitting light beams. The optical path beam splitter 12 can also be called a beam splitter, an optical splitter, an optical beam splitter, etc. Although Figure 1The optical path beam splitter 12 shown in the figure is in the form of a thin sheet. However, in practical applications, an element in the form of a prism with a larger thickness can also be selected. The optical path beam splitter 12 has a reflective surface. For the convenience of introducing the positions of the devices in the optical path system 01, with the reflective surface of the optical path beam splitter 12 as a reference, one side of the reflective surface is called the first side, and the other side is called the second side.

[0062] In the multi-spectral optical characterization system of semiconductor materials based on deep learning, the light source 11 and the sample stage 13 are located on the first side of the reflective surface of the optical path beam splitter 12; the spectrometer 14 is located on the second side of the reflective surface of the optical path beam splitter 12.

[0063] The sample stage 13 is used to hold the semiconductor material sample M to be measured. In this application, the sample stage 13 can be a high-precision three-dimensional movable sample stage, which can perform precise positioning and movement in the x direction, y direction, and z direction. Specifically, the sample stage 13 can be driven by a stepper motor, and the positioning accuracy reaches 1 μm. This design of the sample stage 13 not only supports single-point measurement of the semiconductor material sample M to be measured, but also can perform large-area scanning on it. Therefore, this sample stage 13 can be applied to the testing of semiconductor materials with different shapes and sizes. It ensures the accuracy and flexibility of the measurement.

[0064] As Figure 1 shown by the red line representing the optical path, the light source 11 is used to provide light illumination to stimulate the optical response of the semiconductor material sample M. In this process, the optical path beam splitter 12 is used to receive the light provided by the light source 11 and reflect the light to the surface of the semiconductor material sample through the reflective surface. In this application, the light source 11 can be a xenon lamp, and the xenon lamp supports providing light in the wavelength range of 200 nm to 2500 nm. Such a wide-band light illumination can be used to stimulate the optical responses of different types of semiconductor materials. Thus, the multi-spectral optical characterization system of semiconductor materials based on deep learning provided by the embodiments of this application has strong applicability to the measurement requirements of various different types of semiconductor materials, and there is no need to change the light source 11 in the optical path system multiple times for different semiconductor material samples to be measured.

[0065] The optical response excited from the semiconductor material sample M is transmitted to the spectrometer through an optical signal. Specifically, the optical path beam splitter 12 is also used to transmit the light reflected by the semiconductor material sample M to the spectrometer 14 in a transmissive manner. The spectrometer 14 is used to analyze, modulate, and process the received optical signal and then transmit it to the detector 15. After modulating the broadband light, it is convenient to achieve high-precision spectral measurement. For example, the received reflected light can be spectroscopically processed to obtain the spectral information of the semiconductor material sample M to be measured at different wavelengths, thereby realizing spectral analysis. In practical applications, the spectrometer can be set through different programs to measure various different types of spectra. For example, Raman spectra and photoluminescence spectroscopy (abbreviated as PL spectra). The optical signal of the spectrometer measured is transmitted to the detector 15 in a form for subsequent analysis and use.

[0066] In this application, the spectrometer can measure various different types of spectra simultaneously or successively. These different types of spectra may differ in wavelength bands or physical mechanisms. Thus, each spectrum can reflect different properties of the semiconductor material sample to be measured. For example: Raman spectra reveal lattice vibrations, molecular structures, etc.; photoluminescence spectra reflect electronic structures, band information, etc.

[0067] The detector 15 is used to convert the optical signal received by itself into an electrical signal and then transmit the electrical signal to the data processing module 21. The detector can adopt a high-sensitivity detector type, such as a CCD, an InGaAs detector, etc.

[0068] In this application, the computer system 02 is mainly responsible for the processing and analysis of electrical signals. Specifically, the data processing module 21 is used to obtain the multi-spectral data of the semiconductor material sample through multi-spectral imaging technology based on the electrical signal and input the multi-spectral data into the deep learning model 22. Traditional single spectra (such as a single Raman spectrum or a single photoluminescence spectrum) often can only capture information on one aspect of the material, easily leading to incomplete feature extraction and affecting the analysis accuracy. Therefore, this system uses a spectrometer to measure various types of spectra such as Raman spectra and photoluminescence spectra. The purpose is to perform electrical signal imaging on it using multi-spectral imaging technology in the computer system 02 to form multi-spectral data. Here, the multi-spectral data can be understood as spectral images of various different types of spectra. In this application, comprehensively using the spectral information of different spectral images of the same semiconductor material sample to be measured can improve the comprehensiveness and accuracy of the optical characterization of semiconductor materials.

[0069] In addition, in an alternative implementation, the data processing module 21 can also be used to perform preprocessing operations on the multispectral data after obtaining the multispectral data; the preprocessing operations include at least one of the following: denoising processing, normalization processing, image background correction, or size adjustment. The purpose of the data processing module 21 to perform preprocessing on the multispectral data is to improve the data quality of the multispectral data and reduce measurement errors. For the technical implementation with preprocessing, the multispectral data transmitted by the data processing module 21 to the deep learning model 22 can specifically be the preprocessed multispectral data.

[0070] The deep learning model 22 is used to extract and analyze features from the multispectral data and output the multispectral optical characterization results of the semiconductor material sample; the multispectral optical characterization results include at least one of defect detection results, composition analysis results, or structural characterization results. In this application, the deep learning model 22 is pre-trained using a labeled semiconductor material spectral image dataset, and the semiconductor material spectral image dataset includes spectral images corresponding to multiple different wavelengths; the annotation information of the spectral images includes at least one of the following: defect detection annotation, composition analysis annotation, or structural characterization annotation. It can be understood that if it is expected that the deep learning model 22 can output a certain type of characterization result in the future, then during pre-training, spectral images with corresponding annotations also need to be used as training data. For example, if it is expected that the deep learning model 22 can output defect detection results and composition analysis results for a certain semiconductor material sample in the future, then when pre-training this model, spectral images corresponding to multiple different wavelengths with defect detection annotations and composition analysis annotations also need to be used.

[0071] The defect detection annotation can be to mark the defect areas existing in the material in the image, such as microcracks, voids, non-uniform deposition, impurities, etc. The composition analysis annotation can be to mark the material composition or chemical component information corresponding to each sample area, such as the percentage of silicon content, the type and concentration of doping elements, etc. The specific annotation information may vary depending on the type of semiconductor material, and the specific content of the annotation information is not strictly limited here. The structural characterization annotation can include angle information of lattice orientation, lattice structure type, stress state type, etc. For example, the lattice orientation angle is marked as a specific value, and the stress distribution is marked as "tensile stress area", "compressive stress area", etc. By learning these structural information, the model can output corresponding structural characterization results when inputting the multispectral image of a new sample.

[0072] In this application, the essential purpose of training the model is to enable the model to learn the feature information related to the semiconductor material characteristics of spectral images with different wavelengths, and then establish a mapping relationship between the spectral image features and material properties in the model, which is convenient to output accurate multispectral optical characterization results using the learned mapping relationship during the subsequent use of the model.

[0073] The above is a multi - spectral optical characterization system for semiconductor materials based on deep learning provided by the embodiments of the present application. In the optical path system of the system, a light source irradiates a semiconductor material sample to excite an optical response; an optical path beam splitter receives the light provided by the light source, reflects the light onto the surface of the semiconductor material sample, and transmits the light reflected by the semiconductor material sample to a spectrometer; the spectrometer modulates and processes the received optical signal and then transmits it to a detector; the detector converts the optical signal received by itself into an electrical signal. In the computer system, a data processing module obtains multi - spectral data of the semiconductor material sample through multi - spectral imaging technology based on the electrical signal; a deep learning model extracts and analyzes features from the multi - spectral data and outputs a multi - spectral optical characterization result. By using integrated high - precision optical detection / measurement technologies, intelligent data analysis algorithms, and automated testing technologies, automatic analysis of spectral data and feature extraction are realized, achieving efficient and accurate characterization of semiconductor materials. The measurement accuracy and measurement efficiency of semiconductor material characteristics are improved. This system can not only detect the optical properties of materials, but also perform defect detection, composition analysis, and structural feature recognition, greatly enhancing the intelligence and automation level of semiconductor material characterization and being applicable to multiple fields such as scientific research experiments, industrial inspections, and semiconductor manufacturing. The present application overcomes the problems of traditional methods that usually rely on expert experience for data analysis, such as large subjective errors, slow analysis speed, and poor repeatability.

[0074] In the traditional deep - learning model training method, a large number of labeled samples are usually required to enable the model to obtain good performance. However, in the field of semiconductor material characterization, especially when dealing with new or rare semiconductor materials, high - quality labeled data is often scarce and expensive. Using traditional training methods often means longer data collection time and higher cost. The existence of this problem also hinders the high - accuracy optical characterization of new or rare semiconductor materials. Therefore, the present application introduces transfer learning to effectively solve the modeling difficulties caused by data scarcity. The multi - spectral optical characterization system for semiconductor materials based on deep learning provided by the technical solution of the present application trains a deep - learning model using the idea of transfer learning, transferring the model knowledge trained on existing tasks to new tasks and avoiding training the model from scratch.

[0075] Specifically, in the multi - spectral optical characterization system for semiconductor materials based on deep learning provided by the technical solution of the present application, first, a large - scale existing semiconductor material database (such as common silicon - based and gallium nitride - based materials) is used for preliminary training to enable the model to master the basic spectral feature recognition ability. Then, when encountering new materials (such as perovskite materials and two - dimensional materials), only a small amount of new data is needed to fine - tune the model, which can quickly adapt to new tasks, significantly reducing the data demand and training time. The following is illustrated with embodiments.

[0076] As introduced above, in an optional implementation, the deep learning model 22 in the computer system 02 in the embodiments of the present application can be obtained by training, through transfer learning, on the basis of a target model that has completed preliminary training, using a first spectral image dataset with annotations; the target model is trained using a second spectral image dataset with annotations. The target model can be understood as a model obtained after preliminary training using the second spectral image dataset with annotations.

[0077] Among them, the second spectral image dataset includes spectral images at multiple different wavelengths of one or more semiconductor materials; the first spectral image dataset also includes spectral images at multiple different wavelengths of one or more semiconductor materials; the types of at least one semiconductor material involved in the first spectral image dataset are not covered by the types of semiconductor materials involved in the second spectral image dataset. As an example, the second spectral dataset involves silicon-based and gallium nitride-based materials; the first spectral dataset involves new materials such as perovskite materials and two-dimensional materials. The role of the first spectral image dataset with annotations is to fine-tune the target model, reducing the model training cost and the model training cycle. The problem of insufficient data volume of new materials is overcome through transfer learning.

[0078] It should be noted that in the present application, the annotation information of the spectral images in the first spectral image dataset and the second spectral image dataset includes at least one of the following: defect detection annotation, composition analysis annotation, or structural characterization annotation. The role of these annotation information is to improve the accuracy of the model's optical characterization of semiconductor materials, and the accuracy of the annotation information affects the analysis performance of the model for the input multi-spectral data.

[0079] For ease of understanding, the following combines Figure 2 to introduce an example structure of the deep learning model. Figure 2 This is a schematic structural diagram of a deep learning model provided by an embodiment of the present application. In the deep learning model, the overall framework is based on a convolutional neural network CNN and an attention mechanism. As Figure 2 shown, in an optional implementation, the deep learning model includes: a shared feature extraction module and a multi-task branch module. The output end of the shared feature extraction module is connected to the input end of the multi-task branch module.

[0080] The multi-task branch includes a defect detection branch, a composition analysis branch, and a structural characterization branch. The defect detection branch, the composition analysis branch, and the structural characterization branch respectively correspond to the defect detection result, the composition analysis result, and the structural characterization result finally output by the multi-spectral optical characterization system of semiconductor materials based on deep learning.

[0081] In a deep learning model, a shared feature extraction module is used to extract features from the multi-spectral data input to the deep learning model. In an optional implementation, the shared feature extraction module includes introducing multi-scale convolutional kernels in the low-level feature processing to adapt to different spectral feature sizes. For example, the shared feature extraction module includes multiple convolutional layers, where the convolutional layer closer to the input is the low-level feature extraction layer, and the convolutional layer farther from the input is the high-level feature extraction layer. For multiple low-level feature extraction layers, multi-scale convolutional kernels are respectively set. The diversity of the sizes of the convolutional kernels meets the feature of the size diversity of the spectral features in the multi-spectral image. If the convolutional kernel is of a single size, it is very likely that the effective extraction of spectral features of various sizes cannot be achieved. Now, the design of the multi-scale convolutional kernels in the shared feature extraction module effectively ensures the accuracy and comprehensiveness of the extraction of spectral features of various sizes. Furthermore, it is convenient for the multi-task branch module to achieve high-precision optical characterization of the semiconductor material sample to be measured by analyzing accurate and comprehensive spectral features.

[0082] In addition, the shared feature extraction module can also introduce a channel attention mechanism to strengthen the features of the key bands. Here, the key bands refer to the bands that are particularly important for optical characterization. The application of the channel attention mechanism improves the feature extraction ability and expression ability of the shared feature extraction module for the features of the key bands, strengthens the important information in the spectral data, and is also beneficial for the characterization system to achieve high-precision optical characterization.

[0083] The features extracted by the shared feature extraction module will be input into the above three parallel branches, namely the defect detection branch, the composition analysis branch, and the structure characterization branch, for these three branches to analyze and process. In the multi-task branch module, the defect detection branch is used to analyze the features output by the shared feature extraction module by using upsampling and convolutional blocks, and output a defect mask. In the multi-task branch module, the composition analysis branch is used to learn the features output by the shared feature extraction module by using a regression algorithm, and output the regional composition content distribution. Regression refers to a machine learning task whose goal is to predict continuous values, rather than the discrete labels predicted in a classification task. In the multi-task branch module, the structure characterization branch is used to analyze the features output by the shared feature extraction module, and output the classification results of the lattice orientation or stress distribution, or output the regression results of the lattice orientation or stress distribution. The structure characterization branch has the characteristic of being compatible with both classification and regression output forms in structure, and can include parallel classification sub-modules and regression sub-modules to meet the needs of multi-task learning. The structure of the structure characterization branch pays more attention to the fine perception of local spatial structure and physical distribution features compared with other branches.

[0084] It can be seen that in the deep learning model, the defect detection branch, the component analysis branch, and the structure characterization branch are each responsible for corresponding detection, analysis, and characterization tasks. The setting of multiple branches enables the model to have more complete detection and optical characterization capabilities, so as to output richer and more comprehensive results. Nowadays, in the fields of semiconductor manufacturing, scientific research experiments, industrial inspections, etc., the demand for optical characterization may vary, and the characterization requirements for different types of semiconductor materials may also be diverse. The branch structures of the deep learning model in the embodiments of the present application can provide sufficient and powerful support for diverse characterization requirements and diverse semiconductor materials to be measured, and serve fields such as semiconductor manufacturing, scientific research experiments, and industrial inspections with complete functions. In the field of semiconductor manufacturing, the characterization system proposed in the embodiments of the present application can be used to detect the composition and structure of wafers and thin film materials, optimize the production process, and improve the product yield; in scientific research experiments, the characterization system proposed in the embodiments of the present application can be used to explore the optical properties of new semiconductor materials and promote the development of materials science; in industrial inspections, the characterization system proposed in the embodiments of the present application can be used for quality control of semiconductor devices to achieve efficient and high-precision automated inspections.

[0085] In an alternative implementation, during the training process of the deep learning model, the loss weights of each task in the loss function are dynamically adjusted based on the importance of different tasks. Specifically, in the loss function, the above three types of tasks are respectively configured with loss weights. As the training iterates, the performance of some tasks may be significantly improved. Therefore, if the loss values of the model are still calculated using the initially set loss weights of each task, it may be difficult to quickly improve the performance of other tasks. For this reason, the loss weights of each task in the loss function can be dynamically adjusted. Dynamically adjusting the loss weights of tasks according to the training effect is beneficial to quickly improving the performance of task branches and achieving balanced optimization of the performance of multiple task branches.

[0086] Figure 3 FIG. is a schematic structural diagram of another multi-spectral optical characterization system for semiconductor materials based on deep learning provided by the embodiments of the present application. In Figure 3 the example, a reflector 16 is further provided between the optical path beam splitter 12 and the spectrometer 14 to realize optical path folding conversion and direction change. In addition, different from Figure 1 the shown structure, in Figure 3 the example, the computer system further includes a display 23, and the display is used to present a visual interface. Figure 3 In the example, the output end of the deep learning model 22 is connected to an input end of the data processing module 21; in the computer system, the data processing module 21 is connected to the display 23, and a visual interface can be displayed through the display 23, and the data processed and analyzed by the characterization system provided by the present application is presented by the visual interface.

[0087] In Figure 3 the example, the data processing module 21 is used to sort, statistically analyze, and quantitatively analyze the multi-spectral optical characterization results, and generate a visualization chart of the semiconductor material sample. The visualization chart includes one or more of a defect distribution map, a composition heat map, and a strain analysis map. The data processing module 21 also standardizes the visualization chart and related analysis data into a multi-spectral optical characterization report. The content of the multi-spectral optical characterization report covers one or more types of information such as the detection method, sample information, detailed analysis results, or comprehensive conclusions. It can be understood that the multi-spectral optical characterization report presents diverse information about this material analysis and characterization in a standardized format. Since the data processing module 21 is connected to the above-mentioned display 23, the content such as the chart in the multi-spectral optical report can be rendered and presented in the visualization interface 231. The visualization interface 231 is used to display the information in the multi-spectral optical characterization report. The visualization interface 231 adopted by the system supports interactive query and analysis, enabling users to intuitively view the spectral curve and compare the data of different samples, further improving the convenience and practicality of the test.

[0088] If the light output by the light source 11 is a continuous light signal, in order to reduce the influence caused by noise and improve the signal-to-noise ratio, a chopper can also be added in the multi-spectral optical characterization system of semiconductor materials based on deep learning. Figure 4 FIG. shows a schematic structural diagram of another multi-spectral optical characterization system of semiconductor materials based on deep learning provided by an embodiment of the present application. In Figure 4 it shows the implementation manner of setting the chopper 17 in the optical path system 01. As Figure 4 shown, the chopper 17 is located between the light source 11 and the optical path beam splitter 12, and the chopper 17 is also located on the first side of the reflection surface of the optical path beam splitter 12. The chopper is used to pulse-modulate the light beam output by the light source to convert the continuous light signal into a pulsed light signal, and uses the lock-in amplification technology to improve the signal-to-noise ratio and enhance the detection ability of weak light signals. This helps to improve the optical characterization accuracy of the entire system.

[0089] In addition, in an optional implementation manner, the multi-spectral optical characterization system of semiconductor materials based on deep learning may further include: a lens assembly. Figure 5 FIG. shows the structural implementation effect of including a lens assembly in the system. As Figure 5 shown, in the optical path system 01, the lens assembly 18 is located between the light source 11 and the chopper 17. The lens assembly 18 is used to modulate the light beam output by the light source into a parallel light beam. By modulating the light beam through the lens assembly 18, the uniformity of the illumination of the semiconductor material sample M to be measured is greatly improved, and thus the stability of the measurement is improved.

[0090] In an alternative implementation, the computer system further includes a control module. The sample stage is driven by a stepper motor with a positioning accuracy of 1 μm, and the stepper motor is connected to the control module. The control module is configured to send a driving signal to the stepper motor to move the sample stage in any one or more of the x-direction, y-direction, or z-direction. That is, within the system, three-dimensional high-precision movement of the semiconductor material sample to be measured can be realized, and the feedback displacement can be received in real time, supporting multi-point measurement of the semiconductor material sample to be measured.

[0091] In addition, the control module can also be connected to a chopper to control the operation of the chopper. Moreover, the control module can also be connected to a spectrometer to control the operation of the spectrometer.

[0092] In summary, the present application provides a multi-spectral optical characterization system for semiconductor materials based on deep learning, which combines a high-precision optical detection system, an intelligent data analysis algorithm, and an automated testing technology to achieve efficient and accurate characterization of semiconductor materials. This system can not only detect the optical properties of materials, but also perform defect detection, composition analysis, and structural feature recognition, greatly improving the intelligence and automation level of semiconductor material characterization. It is applicable to multiple fields such as scientific research experiments, industrial inspections, and semiconductor manufacturing. It can meet the characterization requirements of different types of semiconductor materials, including but not limited to materials such as silicon-based, gallium arsenide, and gallium nitride, and has broad application value in fields such as scientific research, industrial inspection, and semiconductor manufacturing. In addition, the multi-spectral optical characterization system for semiconductor materials based on deep learning can also adopt GPU accelerated computing to greatly improve the data processing speed and meet the needs of large-scale data analysis.

[0093] Based on the system introduced in the foregoing embodiments, correspondingly, the present application also provides a multi-spectral optical characterization method for semiconductor materials based on deep learning. This method can be applied to the multi-spectral optical characterization system for semiconductor materials based on deep learning introduced in the above embodiments. Figure 6 The flowchart of this method is as follows. As Figure 6 shown, this method includes:

[0094] S61: Turn on the light source.

[0095] In the technical solution of the present application, the light source is the light provider of the entire optical path system, and turning on the light source is a prerequisite for the optical path system to start working. Only after the light source is turned on can optical measurement and optical characterization be realized.

[0096] S62: Reflect the light provided by the light source to the surface of the semiconductor material sample to be measured on the sample stage through an optical beam splitter.

[0097] As shown in the relevant drawings of the foregoing embodiments, the optical path splitter in the optical path system provides two functions. One of the functions is, as described in step S62, to use its reflection function to reflect the light beam provided by the light source onto the surface of the semiconductor material sample to be measured, so as to stimulate the optical response of the semiconductor material sample to be measured.

[0098] S63. Transmit the light reflected by the semiconductor material sample through the optical path splitter and transmit it to the spectrometer.

[0099] Step S63 represents the second function of the optical path splitter: the transmission function for the light beam. In step S63, the transmission of the light beam reflected by the semiconductor material sample to the spectrometer is completed, making it possible for the spectrometer to perform spectral measurement in combination with the collected light.

[0100] S64. After the spectrometer modulates and processes the received light, it transmits it to the detector.

[0101] In step S64, the spectrometer mainly performs its own functions. The functions of the spectrometer can be set by itself or controlled by a program to complete the measurement of specific types of spectra and the measurement of spectra in specific wavelength bands.

[0102] S65. The detector converts the optical signal received by itself into an electrical signal and then transmits the electrical signal to the data processing module.

[0103] The detector mainly completes the conversion of optical and electrical signals. The deep learning model in the backend computer system can process electrical data. Therefore, it is necessary to pre-process the optical signal into an electrical signal in step S65 to facilitate the operation, analysis, and processing of the backend computer system.

[0104] S66. The data processing module obtains the multi-spectral data of the semiconductor material sample according to the electrical signal through multi-spectral imaging technology and inputs the multi-spectral data into the deep learning model.

[0105] The data processing module can use imaging technology to process the collected electrical signal into an electrical digital image, that is, spectral data. Since a variety of spectral data have been measured by the spectrometer in the early stage, the multi-spectral data here can also be regarded as multiple spectral images processed by the data processing module. The data processing module can further pre-process the multi-spectral data, such as denoising, normalization, image background correction, or size adjustment, to improve the data quality of the multi-spectral data.

[0106] S67. The deep learning model extracts and analyzes the features of the multi-spectral data and outputs the multi-spectral optical characterization results of the semiconductor material sample.

[0107] The multi-spectral optical characterization results include at least one of defect detection results, composition analysis results, or structural characterization results. Regarding the structure and training method of the deep learning model, they have been introduced in detail in the above system embodiments and will not be elaborated here. Relevant details can be referred to the descriptions in the above system embodiments.

[0108] The multi-spectral optical characterization method for semiconductor materials based on deep learning provided by the embodiments of this application utilizes the optical measurement technology and artificial intelligence algorithms integrated in the system to achieve automatic analysis and feature extraction of spectral data, improving the measurement accuracy and efficiency of semiconductor material characteristics.

[0109] As described above, it is only a specific implementation manner of this application, but the protection scope of this application is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by this application should be covered by the protection scope of this application. Therefore, the protection scope of this application should be subject to the protection scope of the claims.

Claims

1. A multi-spectral optical characterization system for semiconductor materials based on deep learning, characterized in that, The system includes: an optical path system and a computer system; the optical path system includes: a light source, an optical path beam splitter, a sample stage, a spectrometer, and a detector; the computer system is integrated with a data processing module and a deep learning model, and the output end of the data processing module is connected to the input end of the deep learning model; the detector is electrically connected to the computer system; the light source and the sample stage are located on the first side of the reflecting surface of the optical path beam splitter; the spectrometer is located on the second side of the reflecting surface; The sample stage is used to carry a semiconductor material sample to be measured; The light source is used to provide illumination to stimulate the optical response of the semiconductor material sample; The optical path beam splitter is used to receive the light provided by the light source and reflect the light to the surface of the semiconductor material, and is also used to transmit the light reflected by the semiconductor material sample to the spectrometer; The spectrometer is used to modulate and process the received optical signal and then transmit it to the detector; The detector is used to convert the optical signal received by itself into an electrical signal and then transmit the electrical signal to the data processing module; The data processing module is used to obtain the multi-spectral data of the semiconductor material sample through multi-spectral imaging technology according to the electrical signal and input the multi-spectral data into the deep learning model; The deep learning model is used to extract and analyze the features of the multi-spectral data and output the multi-spectral optical characterization result of the semiconductor material sample; the multi-spectral optical characterization result includes at least one of a defect detection result, a composition analysis result, or a structure characterization result.

2. The system according to claim 1, wherein The deep learning model is trained based on a labeled semiconductor material spectral image data set, and the semiconductor material spectral image data set includes spectral images corresponding to multiple different wavelengths; the annotation information of the spectral images includes at least one of the following: defect detection annotation, composition analysis annotation, or structure characterization annotation.

3. The system according to claim 1, wherein The deep learning model is trained based on a labeled first spectral image data set by means of transfer learning on the basis of a target model that has been preliminarily trained; the target model is trained based on a labeled second spectral image data set; Among them, the second spectral image data set includes spectral images of one or more semiconductor materials at multiple different wavelengths; The first spectral image data set also includes spectral images of one or more semiconductor materials at multiple different wavelengths; the type of at least one semiconductor material involved in the first spectral image data set is not covered by the type of semiconductor materials involved in the second spectral image data set; The annotation information of the spectral images in the first spectral image data set and the second spectral image data set includes at least one of the following: defect detection annotation, composition analysis annotation, or structure characterization annotation.

4. The system according to any one of claims 1 to 3, characterized in that, The deep learning model includes: a shared feature extraction module and a multi-task branch module; the output end of the shared feature extraction module is connected to the input end of the multi-task branch module; the multi-task branch includes a defect detection branch, a composition analysis branch, and a structure characterization branch; The shared feature extraction module is used to extract features from the multi-spectral data input into the deep learning model; The defect detection branch is used to analyze the features output by the shared feature extraction module by using upsampling and convolutional blocks, and output a defect mask; The component analysis branch is used to learn the features output by the shared feature extraction module by using a regression algorithm, and output the distribution of regional component contents; The structural characterization branch is used to analyze the features output by the shared feature extraction module, and output a classification result of lattice orientation or stress distribution, or output a regression result of lattice orientation or stress distribution.

5. The system according to claim 4, wherein The shared feature extraction module includes introducing multi-scale convolutional kernels in low-level feature processing to adapt to different spectral feature sizes; the shared feature extraction module also introduces a channel attention mechanism to strengthen the features of key bands.

6. The system according to claim 4, wherein During the training process of the deep learning model, the loss weights of each task in the loss function are dynamically adjusted based on the importance of different tasks.

7. The system according to any one of claims 1-3, characterized in that, The output end of the deep learning model is connected to an input end of the data processing module; the computer system also includes a visualization interface; The data processing module is used to sort, statistically analyze, and quantitatively analyze the multi-spectral optical characterization results, and generate a visualization chart of the semiconductor material sample; Standardize the visualization chart and related analysis data into a multi-spectral optical characterization report; the visualization chart includes one or more of a defect distribution chart, a component heat map, and a strain analysis chart; The content of the multi-spectral optical characterization report covers one or more of detection methods, sample information, detailed analysis results, or comprehensive conclusions; The visualization interface is used to display the information in the multi-spectral optical characterization report.

8. The system according to any one of claims 1 to 3, characterized in that, It also includes: A chopper; The chopper is located between the light source and the optical path beam splitter, and the chopper is also located on the first side of the reflecting surface; The chopper is used to pulse-modulate the light beam output by the light source to convert a continuous optical signal into a pulsed optical signal.

9. The system according to claim 8, wherein It also includes: A lens assembly; the lens assembly is located between the light source and the chopper; The lens assembly is used to modulate the light beam output by the light source into a parallel light beam.

10. The system according to any one of claims 1 to 3, characterized in that, The data processing module is also used to perform preprocessing operations on the multi-spectral data after obtaining the multi-spectral data; the preprocessing operations include at least one of the following: Denoising processing, normalization processing, image background correction, or size adjustment.

11. The system according to any one of claims 1 to 3, characterized in that, The light source is a xenon lamp, and the xenon lamp supports providing light in the wavelength range of 200 nm to 2500 nm to be used to excite the optical responses of different types of semiconductor materials.

12. The system according to any one of claims 1-3, characterized in that, The computer system includes a control module, the sample stage is driven by a stepping motor with a positioning accuracy of 1 μm, and the stepping motor is connected to the control module; The control module is used to send a driving signal to the stepping motor to move the sample stage in any one or more of the x direction, y direction, or z direction.

13. A multi-spectral optical characterization method for semiconductor materials based on deep learning, characterized in that Applied to the deep learning-based multi-spectral optical characterization system for semiconductor materials according to any one of claims 1-12, the method includes: Turn on the light source; The light provided by the light source is reflected to the surface of the semiconductor material sample to be measured on the sample stage through an optical path beam splitter; The light reflected by the semiconductor material sample is transmitted through the optical path beam splitter and transmitted to a spectrometer; The light received by the spectrometer is modulated and processed and then transmitted to a detector; The detector converts the optical signal received by itself into an electrical signal, and then transmits the electrical signal to a data processing module; The data processing module obtains the hyperspectral data of the semiconductor material sample through hyperspectral imaging technology based on the electrical signal, and inputs the hyperspectral data into a deep learning model; The deep learning model performs feature extraction and analysis on the hyperspectral data, and outputs the hyperspectral optical characterization result of the semiconductor material sample; the hyperspectral optical characterization result includes at least one of a defect detection result, a composition analysis result or a structure characterization result.

Images