Detection system and compensation method for semiconductor surface structure morphology and computer readable medium

Abstract

The invention discloses a semiconductor surface structure morphology detection system, a compensation method and a computer readable medium, and the method comprises the steps: providing feature information corresponding to a frequency spectrum signal for a neural network-like model; and providing the characteristic information corresponding to the spectrum signal, the detection height value and the actually measured actual height value for the other neural network-like model, so that the trained combined neural network-like model can generate a compensation value for a to-be-corrected height value corresponding to a to-be-corrected spectrum signal with variability. The compensation value is used for providing necessary compensation for height information, so that the accuracy of detection of the surface structure morphology of the semiconductor is improved, and the reliability of a detection system is also improved.

Description

Technical Field

[0001] The present invention relates to the detection of the surface structure topography of a semiconductor. More specifically, the present invention relates to a compensation method for improving the interpretation ability of the surface structure topography, a detection system using the compensation method, and a computer-readable medium.

Background Art

[0002] In the manufacturing process of a semiconductor, many processes and corresponding detections are experienced. The surface structures of semiconductors all have their corresponding forming positions and shapes, which need to be effectively monitored during the manufacturing process to ensure the yield. In addition to some process defects that may cause defects in the surface structure topography of the semiconductor, during the detection process, contact detection (such as the point measurement process using a probe) may also affect the surface structure topography of the semiconductor. For example, contact detection using a probe may form needle marks on the contact pads of the semiconductor structure, that is, the probe causes depressions and protrusions on the surface topography of the contacted pads. The size and depth of the needle mark area will affect the quality of wire bonding in subsequent packaging processes. Therefore, whether the defects occur during the manufacturing process or the detection process, these defects need to be effectively detected to prevent the defective grains from flowing into the backend process.

[0003] Traditionally, the surface structure topography of a semiconductor can be detected by an optical measurement device using optical interference technology. For example, when the surface to be measured has undulations in topography, after the measurement light reflected from the surface to be measured is mixed with the reference light reflected from the reference surface, due to the optical path difference, the optical measurement device obtains an interference signal of the reflected light spectrum. By converting the interference signal through a mathematical model (such as Fourier transform), a waveform signal presenting the distribution of power spectral density in the frequency domain can be obtained. This waveform signal is called a spectral signal. Among them, the peaks in the spectral signal are the basis for determining the optical path difference. That is, based on the obtained optical path difference, the height condition of the detection point can be known. After the detection of each detection point in the target detection area is completed, the surface structure topography of the semiconductor in this target detection area can be known.

[0004] However, when the detected semiconductor structure has a material with a high reflectivity or the semiconductor structure has a multi-layer film structure, and / or when the detection point is located at the edge of the semiconductor structure, the waveform signal of the power spectral density distribution is prone to variability, such as skewness, double wave packets, multiple wave packets, etc. Many algorithms are used to analyze the characteristics of the spectral signal, but usually only for a single type of characteristic, which cannot provide effective benefits in the use of the overall detection system. Therefore, these possible variant characteristics will cause the detection system to misjudge the peaks in the spectral signal, resulting in a high error rate in the detection result of the semiconductor surface structure topography and the reliability cannot be improved.

Summary of the Invention

[0005] In some embodiments disclosed in the present invention, a detection system for semiconductor surface structure topography and its compensation method provide a learning and training method to solve the problem that the peaks in the spectral signal may be misjudged.

[0006] In some embodiments disclosed in the present invention, a detection system for semiconductor surface structure topography and its compensation method improve the accuracy of determining the semiconductor surface structure topography and also increase the reliability of the detection system.

[0007] According to some embodiments, the present invention provides a compensation method for a detection system for semiconductor surface structure topography, which is used to compensate a to-be-corrected height value corresponding to a to-be-corrected spectral signal with variability. The compensation method includes: providing a plurality of eigenvalue of the to-be-corrected spectral signal to a first type of neural network model executed by an arithmetic processing device to obtain a first set of feature information, and providing the to-be-corrected spectral signal to a second type of neural network model executed by the arithmetic processing device to obtain a second set of feature information; and fusing the first set of feature information representing waveform features and the second set of feature information representing compensation degree information to generate a waveform category information and a compensation information, and the compensation information is used for the detection system to correct the to-be-corrected height value to a corrected height value. Wherein, the first type of neural network model is trained based on the eigenvalues corresponding to a plurality of abnormal signals and a plurality of normal signals in one-to-one correspondence. The second type of neural network model is trained based on the eigenvalues corresponding to abnormal signals and abnormal height information and based on the eigenvalues corresponding to normal signals and normal height information.

[0008] According to some embodiments of the present invention, the first type of neural network model can be trained in the following manner: providing a plurality of abnormal eigenvalues corresponding to abnormal signals that can present a variable spectral state and a plurality of normal eigenvalues corresponding to normal signals that do not have a variable spectral state as input vectors to the first type of neural network model.

[0009] According to some embodiments of the present invention, the above-mentioned second type of neural network model can be trained in the following manner: Each abnormal signal, the corresponding abnormal eigenvalue, and an abnormal height difference value, and each normal signal, the corresponding normal eigenvalue, and a normal height difference value are provided as input vectors to the second type of neural network model. The abnormal height difference value and the normal height difference value are the differences between a detected height value corresponding to each signal and an actual measured true height value. Among them, the abnormal height information is the abnormal height difference value, and the normal height information is the normal height difference value.

[0010] According to some embodiments of the present invention, the second processing step may further include a compensation control program. The compensation control program is based on a threshold value, and only when the correction amplitude of the compensation information corresponding to the spectrum signal to be corrected is greater than the threshold value, the arithmetic processing device corrects the height value to be corrected of the spectrum signal to be corrected.

[0011] According to some embodiments of the present invention, in the above-mentioned manner of training the second type of neural network model, the input vector provided to the second type of neural network model may be an input vector recombined from the spectrum signal to be corrected and an additional spectrum signal corresponding to each of a plurality of additional detection points. Among them, each signal presenting a variable spectrum state corresponds to a detection point, and an additional range around the detection point is defined as a plurality of additional detection points. Further, the additional range may be the range covered by 1 detection point.

[0012] According to some embodiments of the present invention, in the above-mentioned second processing step, a preprocessing program may further be included. The preprocessing program may be arranged to be executed before providing the input vector to the second type of neural network model. The preprocessing program is used to individually and sequentially perform a DC component removal step, a signal enhancement step, and a signal mean shift step on all spectrum signals provided to the second type of neural network model.

[0013] According to some embodiments, the present invention also provides a detection system for the surface structure topography of a semiconductor, including: an optical measurement device and an arithmetic processing device. The optical measurement device is used to scan a target detection area of a to-be-detected object with an illumination light to obtain a plurality of interference signals formed by the reflected light reflected from the to-be-detected object. The arithmetic processing device is used to receive such interference signals and convert them into a plurality of spectrum signals without variability and at least one spectrum signal to be corrected with variability. The arithmetic processing device is also used to execute the compensation method as described above to correct a height value to be corrected corresponding to the at least one spectrum signal to be corrected into a corrected height value, so as to correctly generate the surface structure topography of the target detection area.

[0014] According to some embodiments, the present invention further provides a non-volatile computer-readable storage medium. The computer-readable storage medium stores a computer program. The computer program includes a first type of neural network model trained based on eigenvalues corresponding to a plurality of spectral signals and a second type of neural network model trained based on such spectral signals and corresponding height information. The computer program is used to execute the following method: obtaining a to-be-corrected spectral signal based on the surface structure topography of a to-be-tested object, obtaining a plurality of signal eigenvalues of the to-be-corrected spectral signal and providing them to the first type of neural network model to obtain a first set of feature information, and providing the to-be-corrected spectral signal to the second type of neural network model to obtain a second set of feature information, fusing the first set of feature information and the second set of feature information to obtain a compensation information for the to-be-corrected spectral signal, and the compensation information is used to compensate a to-be-corrected height value corresponding to the to-be-corrected spectral signal.

[0015] According to some embodiments, the present invention further provides a detection system for the surface structure topography of a semiconductor, including: an optical measurement device, a topography measurement device, and an arithmetic processing device. Through the configuration of the topography measurement device, at the user end, by providing a plurality of standard to-be-tested objects (each standard to-be-tested object has corresponding regions on its surface topography where the spectral signal can be normal and abnormal), the neural network model can be trained by itself. The arithmetic processing device can summarize the information related to each spectral signal corresponding to these standard to-be-tested objects to establish and train the neural network model. Among them, the optical measurement device is configured to scan a target detection area with an illumination light and is used to obtain the optical interference state formed by the reflected light from the target detection area. The topography measurement device is configured to obtain the true height value of the target detection area. The arithmetic processing device is coupled to the optical measurement device and the topography measurement device and is configured to execute the following steps: a first training step, a second training step, an analysis step, and a compensation step. The first training step corresponds to the training method of the first type of neural network model described above. The second training step corresponds to the training method of the second type of neural network model described above. Through the analysis step, a compensation information is obtained based on the abnormal features included in the to-be-corrected spectral signal. Through the compensation step, the to-be-corrected height value corresponding to the to-be-corrected spectral signal is corrected to a corrected height value to correctly generate the surface structure topography of the to-be-tested object in the target detection area.

[0016] In this way, the spectral signal used to evaluate the surface structure topography of the semiconductor can be correspondingly compensated in the case of having errors or mistakes, solving the problem that would have been misjudged, and also improving the accuracy of the determination of the surface structure topography of the semiconductor.

Description of the Drawings

[0017] Figure 1 It is a schematic diagram of a detection system for the surface structure topography of a semiconductor according to some embodiments; Figure 2 Schematic diagram of a spectral signal with skewness; Figure 3 Schematic diagram of a spectral signal with double wave packets; Figure 4 Compensation method for a detection system for semiconductor surface structure topography according to some embodiments; Figure 5 Schematic diagram of the use of a first type of neural network model and a second type of neural network model according to some embodiments; Figure 6 Schematic diagram of the training of a first type of neural network model and a second type of neural network model according to some embodiments; and Figure 7 Schematic diagram of a detection system for semiconductor surface structure topography according to another embodiment.

Detailed Description of the Embodiments

[0018] To fully understand the purpose, features, and effects of the present invention, the following specific embodiments are hereby used in conjunction with the accompanying drawings to provide a detailed description of the present invention as follows:

[0019] In this application, the terms "a" or "an" are used to describe elements or features. This is merely for convenience of description and provides a general meaning to the scope of this document. Therefore, unless clearly indicated otherwise, such description should be understood to include one or at least one, and the singular also includes the plural.

[0020] In this application, the terms "comprising", "including", "having", or any other similar terms are not limited to only the elements or features listed herein, but may include other parts that are not explicitly listed but are usually inherent to the elements or features.

[0021] In this application, terms such as "first" or "second" and other similar ordinal terms are used to distinguish or refer to related elements or features that are the same or similar, and do not necessarily imply the order of these elements or features in the process. It should be understood that in certain situations or configurations, the ordinal terms can be used interchangeably without affecting the disclosed or related embodiments of the present invention.

[0022] Optical measurement devices are often used in the detection of surface topography, such as measurement devices based on the reflected light from the surface topography, or topography measurement devices that, although slower, can be used to accurately obtain the surface topography (e.g., scanning electron microscope (SEM), atomic force microscope (AFM), or scanning tunneling microscope (STM), etc.). Among them, these slower topography measurement devices have another drawback in that they may damage the surface of the object to be measured.

[0023] In order to obtain detection results more quickly, measurement devices based on the amount of reflected light from the surface topography are usually employed. For example, an optical measurement device with a reference mirror (such as a white light interferometer) is used, or an optical measurement device without a reference mirror that directly uses the surface of the object to be measured as the reference surface (such as a spectrometer for measuring the depth of TSV holes). The optical interference spectrum signal of the obtained reflected light is converted by an algorithm (such as Fourier transform) to obtain the converted frequency spectrum signal. Since the optical path difference that generates the interference phenomenon of the reflected light is the key factor affecting the phase change in the optical interference spectrum signal, after converting to the frequency domain of the frequency spectrum signal, the coordinate information of the peaks in the waveform can correspond to the information of the optical path difference, and then the height information of the detection point (or the illuminated area) can be obtained. Among them, the spectral signal without interference phenomenon also represents that the height information of this detection point is the preset height information (that is, the surface topography is flat and without undulations). Among them, the method of obtaining depth information through the frequency spectrum signal converted from the optical interference signal is a well-known technology and will not be elaborated here.

[0024] Please refer to Figure 1 , which is a schematic diagram of a detection system for the surface structure topography of a semiconductor according to some embodiments. The optical measurement device 100 is used to provide illumination light 101 to the object to be measured 300 for scanning the target detection area (detecting each point one by one). Through the beam splitting unit 130, a part of the illumination light 101 forms transmitted light 102 towards the reference mirror 140. With the coaxial illumination configuration of the light source unit 120 and the beam splitting unit 130, the reflected light from the object to be measured 300 and the reference light from the reference mirror 140 can be captured as spectral signals by the imaging unit 110, and the imaging unit 110 can capture spectral signals with interference phenomena. The arithmetic processing device 200 is coupled to the light source unit 120 and the imaging unit 110 of the optical measurement device 100 to control the scanning operation, receive the spectral signals, and perform subsequent signal analysis.

[0025] The optical measurement device 100 can be a white light interferometer or other optical measurement devices that can obtain optical interference phenomena, Figure 1 and a white light interferometer is taken as an example. The arithmetic processing device 200 can be a single computer or multiple computers, or a single arithmetic processing module or multiple arithmetic processing modules configured in the overall detection system. The arithmetic processing device 200 is also used to receive the signals provided by the optical measurement device 100.

[0026] Among them, the spectral signals generated based on some states of the surface topography are usually symmetric. When the spectral signals are abnormal and become variant spectral signals, the algorithms generally used to automatically capture the wave peaks to obtain the optical path difference information often lead to misjudgment of the optical path difference information, and then misestimate the surface topography. These variant spectral signals with abnormalities have characteristics in their waveforms, such as asymmetric forms like skewness, double wave packets, multiple wave packets, etc. Please refer to Figure 2 and Figure 3 , Figure 2 which is a schematic diagram of a spectral signal with skewness, Figure 3 and Figure 2 is a schematic diagram of a spectral signal with a double wave packet. In Figure 3 , the wave peak pointed by arrow A is captured by the algorithm. However, after precise measurement (such as using an atomic force microscope), in fact, the wave peak pointed by arrow B is the location of the correct optical path difference information. Similarly, in

[0027] , the wave peak pointed by arrow A is captured by the algorithm. However, after precise measurement (such as using an atomic force microscope), in fact, the wave peak pointed by arrow B is the location of the correct optical path difference information. Therefore, for the mutated waveform, the wave peak captured by the algorithm often does not represent the true height information of the detection point.

[0027] The spectral signal can extract the feature information of the waveform through many feature extraction tools (algorithms) for the signal waveform. For example, using a graphic feature extraction method based on gray values, which extracts the waveform features in the graphic by calculating the maximum, minimum, and average gray values in the graphic. For another example, using an image feature extraction method based on scale space (Scale Derivative), which extracts the waveform features in the graphic by calculating the gradient values of the graphic at different scales.

[0028] The spectral signal can extract the waveform feature information in the signal through these analysis tools. In addition, for the obtained spectral signal, it can be classified based on the concerned forms, and these forms will have corresponding waveform feature information. The aforementioned analysis tools are generally the basis for classifying the spectral signal. For example, if the concerned forms are the center of gravity position, skewness, double wave packet, and multiple wave packet, then there will be corresponding characteristic values respectively. In general use, a single comparison or viewing is performed for the items that need to be concerned to check for possible missing (wrong judgment of surface topography) areas. These analysis tools belong to well-known technologies, and the detailed analysis methods will not be elaborated here.

[0029] In embodiments of the present invention, these eigenvalues are used for planning and / or further processing. The classification items can be increased or decreased according to actual needs, and the eigenvalues of the spectral signals corresponding to each detection point under each category can be further summarized. For example, what are the eigenvalues corresponding to the spectral signals belonging to the skewness category in terms of waveform; and what are the eigenvalues corresponding to the spectral signals belonging to the double wave packet category in terms of waveform; and what are the eigenvalues corresponding to the spectral signals belonging to the multiple wave packet category in terms of waveform.

[0030] As another example, based on the foregoing method, it is possible to distinguish which spectral signals have variability and which do not. For example, one classification method is as follows: A spectral signal is classified as a spectral signal with variability as long as it has any one of the forms of skewness, double wave packet, or multiple wave packet. This can be adjusted according to actual needs (such as matching the characteristics of the object to be measured) to include which forms to be concerned about.

[0031] Next, please refer to Figure 4 and Figure 5 , Figure 4 is a compensation method for a detection system for the surface structure topography of a semiconductor according to some embodiments. Figure 5 is a schematic diagram of the use of a first type of neural network model and a second type of neural network model according to some embodiments. Since incorrect or defective spectral signals easily lead to misinterpretation of the optical path difference information, this optical path difference information is the height information of the detection point (or the illumination area), thereby leading to misinterpretation of the surface structure topography of the semiconductor. Therefore, this compensation method is used to compensate the height information (to-be-corrected height value) corresponding to the spectral signals with variability (to-be-corrected spectral signals) so that the detection system can correctly generate the surface structure topography of the target detection area.

[0032] After obtaining the to-be-corrected spectral signal G1, the to-be-corrected spectral signal G1 and its associated eigenvalues are provided to the corresponding neural network model. The compensation method includes the following steps:

[0033] Step S100: Provide the complex eigenvalue G1s corresponding to the to-be-corrected spectral signal G1 to the first type of neural network model 210, and provide the to-be-corrected spectral signal G1 to the second type of neural network model 220.

[0034] Step S200: Provide the first set of feature information V1 output by the first type of neural network model 210 and the second set of feature information V2 output by the second type of neural network model 220 to the fusion layer 230.

[0035] Step S300: Generate a waveform category information and a compensation information. Among them, the compensation information is, for example, Figure 5 the compensation value Hf1 exemplified in, for use in compensating the to-be-corrected height value.

[0036] Among them, these eigenvalue provided to the first type of neural network model 210 are the eigenvalues of each spectral signal under the selected classification items (for example: when the classification item contains a skewness pattern, the analysis tool (algorithm) for the skewness pattern can output the eigenvalue corresponding to the spectral signal). The more classification items are selected, the more corresponding eigenvalues of the spectral signal will be produced and input into the first type of neural network model 210.

[0037] Among them, the first set of feature information V1 output by the first type of neural network model 210 represents the recognition information of the waveform of the spectral signal G1 to be corrected. This recognition information is associated with the selected classification items and, after being fused with the subsequent second set of feature information V2, serves as the basis for numerical regression to obtain compensation information.

[0038] Among them, the second set of feature information V2 output by the second type of neural network model 220 can be used to judge the compensation degree, that is, the compensation degree information corresponding to the spectral signal G1 to be corrected. Specifically, the trained second type of neural network model 220 is used to correctly interpret the signal and can provide a judgment of the compensation degree information (the magnitude of the compensation value). When the first type of neural network model 210 and the second type of neural network model 220 are used together, based on the fusion of the first set of feature information V1 and the second set of feature information V2, the waveform type of the signal can be referred to in the process of generating compensation information. Subsequently, when performing numerical regression, the compensation information and waveform category information (waveform type) of the spectral signal G1 to be corrected are generated. In this way, it is more capable of judging various signal states and providing more accurate compensation information.

[0039] Among them, the fusion layer 230 performs a concatenate operation on the first set of feature information V1 and the second set of feature information V2 to obtain combined features. As mentioned above, under the concatenation of features, the detection system can obtain more accurate compensation information for the surface topography in terms of height based on the waveform type of the signal.

[0040] Accordingly, since the spectral signal G1 to be corrected also has a corresponding detection height value (i.e., the information obtained traditionally), after obtaining the compensation value Hf1 corresponding to the spectral signal G1 to be corrected, the height difference value (the height value to be corrected) corresponding to the variable spectral signal (the spectral signal G1 to be corrected) can be compensated and converted into a corrected height value. Based on this, after all the spectral signals G1 to be corrected are corrected, the detection system can present the correct surface structure topography of the target detection area.

[0041] Further, in addition to compensating only the spectrum signals with variability (to improve the determination efficiency), the arithmetic processing unit 200 (please refer to Figure 1 ) may execute a compensation control program. Based on a threshold value, when the correction amplitude of the compensation information corresponding to the spectrum signal G1 to be corrected is greater than the threshold value (for example: the absolute value of the compensation value Hf1 is greater than the threshold value), the arithmetic processing unit 200 will perform the compensation operation on the detection height value corresponding to the spectrum signal G1 to be corrected.

[0042] Next, please refer to Figure 6 , which is a training schematic diagram of a first type of neural network model and a second type of neural network model according to some embodiments.

[0043] Regarding the training of the first type of neural network model, information presenting a variable spectrum state and information not presenting a variable spectrum state are provided to the first type of neural network model for training. That is to say, the complex abnormal eigenvalues g1s corresponding to the complex abnormal signals g1 and the complex normal eigenvalues g2s corresponding to the complex normal signals g2 are used as input vectors and respectively provided to the first type of neural network model for the training process. The first type of neural network model may be, for example, a Multilayer Perceptron (MLP) or other neural network models having the same or similar learning ability for the signal waveform pattern.

[0044] Among them, with the use of some selected waveform analysis tools, the waveform pattern of each signal can be correspondingly identified. For example, when focusing on the amount of kurtosis value (such as: high or low), the corresponding waveform analysis tool is an algorithm that can calculate kurtosis, and the corresponding waveform pattern generated may be: flat pattern or steep pattern. Other examples: for the amount of skewness value (such as: high, medium or low) mentioned above, the corresponding waveform analysis tool is an algorithm that can calculate skewness, and the corresponding waveform pattern generated may be: Gaussian normal distribution pattern, left-skewed pattern or right-skewed pattern. The training objective of the first type of neural network model is that the neural network model can identify the waveform pattern based on the eigenvalue of the signal. Taking the Multilayer Perceptron (MLP) as an example of the first type of neural network model, the training can be achieved through the minimization of the loss function, and then the output of the trained Multilayer Perceptron (MLP) is an activation value (activations) or activation vector that can be used to identify the waveform pattern, that is, the first set of feature information.

[0045] Regarding the training of the second type of neural network model, information presenting a variability spectral state and the corresponding abnormal height difference (i.e., a kind of abnormal height information), as well as information without a variability spectral state and the corresponding normal height difference (i.e., a kind of normal height information), are used as input vectors and provided to the second type of neural network model for training. Among them, the height value (regardless of whether it is correct) directly obtained by capturing the peak of the spectral signal for a detection point through an algorithm is the detected height value; similarly, for the same detection point, the height value obtained by a measuring device capable of accurately obtaining the surface topography (such as a scanning electron microscope (SEM), an atomic force microscope (AFM), or a scanning tunneling microscope (STM), etc.) is the true height value, and the height difference value refers to the difference between the detected height value and the true height value. In other words, such abnormal signals g1 and the corresponding abnormal eigenvalue g1s and abnormal height difference value H1d, as well as such normal signals g2 and the corresponding normal eigenvalue g2s and normal height difference value H2d, are used as input vectors and respectively provided to the second type of neural network model for the training process.

[0046] Therefore, in the training of the second type of neural network model, the true height value, the detected height value, and the waveforms of their corresponding spectral signals are used to enable the model to learn and train signal interpretation, so that the resolved feature vector is positively helpful for the subsequent numerical regression result. The model has the ability to output the corresponding compensation degree information for each variability spectral state, and the output vector is an embedding vector based on the waveform to identify the corresponding compensation information, that is, the second set of feature information. Among them, examples of the second type of neural network model can be fully convolutional networks (FCN), residual neural networks (ResNet), or other neural network models with the same or similar learning ability for the form of signal waveforms (for another example: CNN, MCNN, MCDCNN, TWIESN, ENCODER, MLP, INCEPTION, TLENET, programming with HIVE database).

[0047] Among them, the abnormal height difference value H1d belongs to one of the abnormal height information. In addition to using the difference between the detected height value and the true height value as described above, the true height value can also be directly used for the abnormal height information. Also, the normal height difference value H2d belongs to one of the normal height information. In addition to using the difference between the detected height value and the true height value as described above, the true height value can also be directly used for the normal height information (since it is normal height information, the difference here may be zero or only have a tiny difference). Thus, directly using the true height value as described above can also enable the second type of neural network model to learn the height information corresponding to various spectral signals, which can then be used for the subsequent determination of the compensation degree information.

[0048] To elaborate further, during the training process, for a detection position, its spectral signal, waveform feature, detected height value, and true height value are provided for the training of the corresponding neural network model. Among them, since the outputs of the two neural network models are fused, and the waveform feature and the compensation information associated with the detected height value and the true height value are matched to the fused vector (rather than being matched to the vector output by a single neural network model). This enables each fused vector to share all features, that is, based on the weight allocation of each feature during the training process shared by the two neural network models, two neural network models that can obtain more accurate compensation information are trained.

[0049] Furthermore, in some other embodiments, for the training process and the usage process of the second type of neural network model, the information adjacent to each original detection point can be reorganized into a new detection point input information (only the spectral signal is reorganized). Specifically, the information of each detection point and a plurality of additional detection points around it is used as a new input vector. Therefore, the spectral signal of the detection point and the additional spectral signals of each additional detection point are trained within the model of the matching dimension in a data structure arrangement manner. The new input vector further includes the information of the additional detection points around the main detection point. Since the additional detection points are adjacent to the main detection point, they have topographical information highly correlated with the main detection point, which can make the learning and training of the second type of neural network model more abundant. Among them, these additional detection points can be defined as the range of taking an additional detection point around the main detection point as the center within the target detection area. In addition, when the training process uses the reorganized detection point input information, during the process of using the trained second type of neural network model (such as the aforementioned step S100), the spectral signal to be corrected used can also be the reorganized input vector, and the reorganized data structure matches the model of the same dimension that has been trained.

[0050] Further, in some other embodiments, before providing the spectral signal of the components as the input vector to the second neural network model, a preprocessing procedure is performed on the spectral signal. The preprocessing procedure is used to enhance the signal. The preprocessing procedure includes, in sequence: a DC component removal step, a signal enhancement step, and a signal mean shift step.

[0051] Regarding the DC component removal step, it is used to remove the constant term generated over time. Usually, methods such as the first derivative and subtracting the average are used. In addition, this step can also remove some situations that may cause calculation errors (such as the batwing effect).

[0052] Regarding the signal enhancement step, it is used to make most of the negative values positive. Usually, the square method is used for signal enhancement.

[0053] Regarding the signal mean shift step, it is used to further filter out irrelevant information. For example, the Mean shift algorithm can be used to calculate the intensity values of each signal and subtract a shift value together (this shift value is the average value of all signals in terms of intensity, and this shift value also corresponds to the filtering degree to be achieved), so as to separate positive and negative signals (to achieve the filtering effect of only retaining a few main peaks as positive and the others as negative in each spectral signal), and adjust the original ReLU activation function to the LeakyReLU algorithm (for example, set the slope of the LeakyReLU for the negative value part to a very small positive value, such as 0.01), so as to give a certain training influence to the negative values. In this way, the negative value data can still have a certain influence when training the model, so as to effectively improve the performance of the model when processing noisy data.

[0054] Next, please refer to Figure 7 , which is a schematic diagram of a detection system for the surface structure topography of a semiconductor according to another embodiment. The detection system of this embodiment includes: an optical measurement device 100, a topography measurement device 400, and an arithmetic processing device 200.

[0055] The optical measurement device 100 is configured to scan a target detection area (an area on the object to be measured 300) with the illumination light 101 and is used to obtain the optical interference state formed by the reflected light from the target detection area. The topography measurement device 400 is configured to obtain the true height value of the target detection area. The arithmetic processing device 200 is coupled to the optical measurement device 100 and the topography measurement device 400. The arithmetic processing device 200 is configured to perform the following steps: a first training step, a second training step, an analysis step, and a compensation step.

[0056] Figure 7The embodiments incorporate a topography measurement device 400, which enables the user (system operator) to train a neural network model by providing a plurality of standard objects under test (each standard object under test has corresponding regions on its surface topography where the spectral signals exhibit normal and abnormal states). The arithmetic processing device 200 can summarize the information related to each spectral signal corresponding to these standard objects under test to establish and train the neural network model.

[0057] The first training step corresponds to the training method of the aforementioned first type of neural network model. That is to say, it is executed under the condition that a plurality of standard objects under test are individually provided to the optical measurement device. The arithmetic processing device 200 obtains the light interference state from the optical measurement device 100 and converts it into an abnormal signal with a variable spectral state corresponding to each standard object under test, and converts it into a normal signal without a variable spectral state corresponding to each standard object under test. The arithmetic processing device 200 provides such abnormal signals and corresponding plural abnormal eigenvalues and such normal signals and corresponding plural normal eigenvalues as input vectors to the first type of neural network model for training.

[0058] The second training step corresponds to the training method of the aforementioned second type of neural network model. That is to say, it is executed under the condition that a standard object under test is individually provided to the optical measurement device. The arithmetic processing device 200 also causes the topography measurement device 400 to correspondingly obtain a true height value of the surface topography of each standard object under test. The arithmetic processing device 200 obtains the light interference state from the optical measurement device 100 and converts it into an abnormal signal with a variable spectral state corresponding to each standard object under test, and a corresponding abnormal eigenvalue and an abnormal height difference value. And, the arithmetic processing device 200 obtains the light interference state from the optical measurement device 100 and converts it into a normal signal without a variable spectral state corresponding to each standard object under test, and a corresponding normal eigenvalue and a normal height difference value. The arithmetic processing device 200 provides these abnormal signals, normal signals, abnormal eigenvalues, normal eigenvalues, abnormal height difference values and normal height difference values as input vectors to the second type of neural network model for training to obtain compensation information corresponding to each variable spectral state. Wherein the abnormal height difference value and the normal height difference value are the differences between a detected height value corresponding to each signal and the true height value.

[0059] In the analysis step, it is executed under the condition that there is a test object. The arithmetic processing device 200 causes the optical measurement device 100 to obtain a plurality of interference signals formed by the reflected light of the test object in the target detection area. The arithmetic processing device 200 further converts the interference signals into a plurality of frequency spectrum signals, and when there is at least one to-be-corrected frequency spectrum signal with variability in the frequency spectrum signals, provides the to-be-corrected frequency spectrum signal to the first type of neural network model and the second type of neural network model, for obtaining a first set of feature information representing waveform features and a second set of feature information representing compensation degree information. The arithmetic processing device 200 further performs the fusion of the first set of feature information and the second set of feature information, to generate a compensation value based on the abnormal features included in the to-be-corrected frequency spectrum signal.

[0060] In the compensation step, it means that the arithmetic processing device 200, based on the compensation information, corrects a to-be-corrected height value corresponding to the to-be-corrected frequency spectrum signal into a corrected height value, to correctly generate the surface structure topography of the test object in the target detection area.

[0061] During the training process of the neural network, according to the quality of the obtained frequency spectrum signals (such as whether the signals are affected by environmental factors and cause additional effects on the waveform, whether the true height value is correct, whether the feature information of the signals is complete, etc.), the required number of trainings is measured. In addition, when the probability of forming a mean square error approaches zero after training or there is no further progress in a certain number of training cycles (for example, set to 50 times or other times), the training ends. For example but not as a limitation, select 5 positions with pin marks to train these two neural network models, and each neural network model can be trained using about 500,000 pieces of data (where the number of training cycles is 300 times).

[0062] The various functions and operations executed in the form of software can be achieved by storing a computer program in a non-volatile computer-readable storage medium and then executing it. The computer program is stored in the medium, and the computer program includes a plurality of instructions, for causing an electronic device (such as various computer devices, network devices, or other electronic devices, etc.) or a processor to execute the compensation method for the semiconductor surface structure topography described in various embodiments of the present invention.

[0063] In summary, through the learning, training, and use of the combined two neural network models, correct determination can be provided for the frequency spectrum signals and necessary compensation can be provided for the height information, thereby improving the accuracy of the detection of the semiconductor surface structure topography and increasing the reliability of the detection system.

[0064] The present invention discloses preferred embodiments above, but those skilled in the art should understand that the embodiments herein are only used to describe the present invention and should not be interpreted as limiting the scope of the present invention. It should be noted that all changes and substitutions equivalent to the embodiments should be understood to be included in the scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims. [Reference Signs]

[0065] 100 Optical measuring device 101 illumination light 102 Penetrating Light 110 Image acquisition unit 120 Light source unit 130 Spectroscopic Unit 140 Reference mirror 200 Operation processing device 210 The first type of neural network model 220 The second type of neural network model 230 Fusion Layer 300 DUT 400 Profile measurement device g1 abnormal signal g1 s abnormal eigenvalue g2 Normal signal g2s normal eigenvalue G1 Spectrum signal to be corrected G1s eigenvalue Hf1 compensation value H1d Abnormal height difference value H2d Normal height difference S100-S300 Steps V1 The first set of feature information V2 The second set of feature information

Claims

1. A compensation method for a detection system of semiconductor surface structure topography, which is used to compensate a to-be-corrected height value corresponding to a to-be-corrected spectrum signal with variability. The compensation method includes: Providing a plurality of eigenvalue of the to-be-corrected spectrum signal to a first type of neural network model executed by an arithmetic processing device to obtain a first set of feature information, and providing the to-be-corrected spectrum signal to a second type of neural network model executed by the arithmetic processing device to obtain a second set of feature information ； And Fusing the first set of feature information representing waveform features and the second set of feature information representing compensation degree information to generate a waveform category information and a compensation information, and the compensation information is used for the detection system to correct the to-be-corrected height value to a corrected height value; Wherein, the first type of neural network model is trained based on the eigenvalues corresponding to a plurality of abnormal signals and a plurality of normal signals in one-to-one correspondence; and Wherein, the second type of neural network model is trained based on the eigenvalues corresponding to the abnormal signals and abnormal height information and based on the eigenvalues corresponding to the normal signals and normal height information.

2. The compensation method according to claim 1, Wherein, The first type of neural network model is trained in the following manner: providing a plurality of abnormal eigenvalues corresponding to the abnormal signals that can present a variable spectrum state and a plurality of normal eigenvalues corresponding to the normal signals that do not have a variable spectrum state as input vectors to the first type of neural network model.

3. The compensation method according to claim 2, Wherein, The second type of neural network model is trained in the following manner: providing each of the abnormal signals and the corresponding abnormal eigenvalues and an abnormal height difference value, and providing each of the normal signals and the corresponding normal eigenvalues and a normal height difference value as input vectors to the second type of neural network model, and the abnormal height difference value and the normal height difference value are the differences between a detected height value corresponding to each signal and an actual measured true height value.

4. The compensation method according to claim 3, Wherein, It further includes a compensation control step, which is based on a threshold value, and when the correction amplitude of the compensation information corresponding to the to-be-corrected spectrum signal is greater than the threshold value, the arithmetic processing device is enabled to correct the to-be-corrected height value of the to-be-corrected spectrum signal.

5. The compensation method according to claim 3, Wherein, In the training manner of the second type of neural network model, the input vector provided to the second type of neural network model is an input vector recombined by the to-be-corrected spectrum signal and an additional spectrum signal corresponding to each of a plurality of additional detection points. Wherein, the to-be-corrected spectrum signal corresponds to a detection point, and an additional range around the detection point is defined as the additional detection point.

6. The compensation method according to claim 5, Wherein, The additional range is the range covered by 1 detection point.

7. The compensation method according to any one of claims 1 to 6, Wherein, It further includes a preprocessing step, which is executed before providing the input vector to the second type of neural network model. The preprocessing step is used to individually and sequentially perform a DC component removal step, a signal enhancement step, and a signal mean shift step on all the spectral signals provided to the second type of neural network model.

8. The compensation method according to claim 7, wherein, the first type of neural network model is a Multilayer Perceptron (MLP), and the second type of neural network model is a fully convolutional network (FCN) or a residual neural network (ResNet).

9. A detection system for the surface structure topography of a semiconductor, comprising: an optical measurement device configured to scan a target detection area of a test object with an illumination light to obtain a plurality of interference signals formed by the reflected light reflected from the test object; and an arithmetic processing device configured to receive the interference signals and convert them into a plurality of spectral signals without variability and at least one spectral signal to be corrected with variability. The arithmetic processing device is further configured to execute the compensation method according to any one of claims 1 to 8 to correct a to-be-corrected height value corresponding to the at least one spectral signal to be corrected into a corrected height value, so as to correctly generate the surface structure topography of the target detection area.

10. A non-volatile computer-readable storage medium stores a computer program, the computer program includes a first type of neural network model trained based on the eigenvalue corresponding to a plurality of spectral signals and a second type of neural network model trained based on the spectral signals and the corresponding height information. The computer program is used to execute the following method: Based on obtaining a spectral signal to be corrected from the surface structure topography of a test object, obtaining a plurality of signal eigenvalues of the spectral signal to be corrected and providing them to the first type of neural network model to obtain a first set of feature information, and providing the spectral signal to be corrected to the second type of neural network model to obtain a second set of feature information, and fusing the first set of feature information and the second set of feature information to obtain a compensation information of the spectral signal to be corrected, where the compensation information is used to compensate a to-be-corrected height value corresponding to the spectral signal to be corrected.

11. A detection system for the surface structure topography of a semiconductor, comprising: an optical measurement device configured to scan a target detection area with an illumination light and configured to obtain the optical interference state formed by the reflected light from the target detection area; a topography measurement device configured to obtain a true height value of the target detection area; and an arithmetic processing device coupled to the optical measurement device and the topography measurement device, configured to perform the following steps: A first training step is to execute under the condition that a plurality of standard objects to be measured are individually provided to the optical measurement device. The arithmetic processing device obtains the optical interference state from the optical measurement device to convert it into an abnormal signal with a variable spectrum state corresponding to each standard object to be measured, and converts it into a normal signal without a variable spectrum state corresponding to each standard object to be measured. The arithmetic processing device provides a plurality of abnormal eigenvalues corresponding to the abnormal signal and a plurality of normal eigenvalues corresponding to the normal signal as input vectors to a first type of neural network model for training; A second training step is to execute under the condition that the standard object to be measured is individually provided to the optical measurement device. The arithmetic processing device causes the topography measurement device to correspondingly obtain a true height value of the surface topography of each standard object to be measured. The arithmetic processing device obtains the optical interference state from the optical measurement device to convert it into an abnormal signal with a variable spectrum state corresponding to each standard object to be measured, an abnormal eigenvalue, and an abnormal height difference value, and the arithmetic processing device obtains the optical interference state from the optical measurement device to convert it into a normal signal without a variable spectrum state corresponding to each standard object to be measured, a normal eigenvalue, and a normal height difference value. The arithmetic processing device provides the abnormal signal, normal signal, abnormal eigenvalue, normal eigenvalue, abnormal height difference value, and normal height difference value as input vectors to a second type of neural network model for training, where the abnormal height difference value and the normal height difference value are the differences between a detected height value corresponding to each signal and the true height value; An analysis step is to execute under the condition of having a object to be measured. The arithmetic processing device causes the optical measurement device to obtain a plurality of interference signals formed by the reflected light of the object to be measured in the target detection area. The arithmetic processing device converts the interference signals into a plurality of spectrum signals, and when there is at least one spectrum signal to be corrected with variability in the spectrum signals, provides the spectrum signal to be corrected to the first type of neural network model and the second type of neural network model to obtain a first set of feature information with waveform category information and a second set of feature information with compensation degree information, and fuses the first set of feature information and the second set of feature information to generate a compensation information based on the abnormal features contained in the spectrum signal to be corrected; and A compensation step, the arithmetic processing device corrects a to-be-corrected height value corresponding to the spectrum signal to be corrected into a corrected height value based on the compensation information to correctly generate the surface structure topography of the object to be measured in the target detection area.

12. The detection system according to claim 11, wherein, the topography measurement device is a scanning electron microscope (SEM), an atomic force microscope (AFM), or a scanning tunneling microscope (STM).

13. The detection system according to claim 11, wherein, the optical measurement device is a white light interferometer.

14. The detection system according to any one of claims 11 to 13, wherein, the operation processing device is further configured to perform the following steps: in the analysis step, based on a threshold value, when the correction amplitude of the compensation information corresponding to the spectrum signal to be corrected is greater than the threshold value, the operation processing device is caused to perform the compensation step on the height value to be corrected of the spectrum signal to be corrected.

15. The detection system according to any one of claims 11 to 13, wherein, in the second training step, the input vector provided to the second type of neural network model is an input vector recombined from the spectrum signal to be corrected and an additional spectrum signal corresponding to a plurality of additional detection points respectively, wherein the spectrum signal to be corrected corresponds to a detection point, and an additional range around the detection point is defined as the additional detection point.

16. The detection system according to claim 15, wherein, the additional range is the range covered by 1 detection point.

17. The detection system according to any one of claims 11 to 13, wherein, in the second training step and the analysis step, the operation processing device is further configured to perform a preprocessing step, which is executed before providing the input vector to the second type of neural network model, and the preprocessing step is used to individually and sequentially perform a DC component removal step, a signal enhancement step, and a signal mean shift step on all the spectrum signals provided to the second type of neural network model.

18. The detection system according to claim 17, wherein, the first type of neural network model is a Multilayer Perceptron (MLP), and the second type of neural network model is a fully convolutional network (FCN) or a residual neural network (ResNet).

Images