TWI863041B - Method for estimating real-time wafer processing quality and an electronic device - Google Patents

Abstract

Embodiments of the disclosure provide a method for estimating real-time wafer processing quality and an electronic device. The method includes: obtaining a processing state signal generated by a wafer processing equipment during performing a processing operation on a wafer; and obtaining real-time processing quality of the processing operation via inputting the processing state signal into a machine learning model, wherein the machine learning model estimates the real-time processing quality of the processing operation in response to the processing state signal.

Description

Method and electronic device for real-time wafer processing quality estimation

The present invention relates to a technology for managing wafer processing quality, and more particularly to a method and an electronic device for real-time wafer processing quality estimation.

After the silicon wafer is ground, the inspectors will check the various specifications of the silicon wafer, such as the cleanliness, roughness, total thickness variation (TTV), bow and warp of the silicon wafer surface. When checking the roughness of the silicon wafer surface, the inspectors generally use a roughness meter for measurement. For specifications such as TTV, bow and warp, the inspectors use non-contact wafer measurement instruments to ensure that the accuracy or quality of the silicon wafer surface meets the requirements.

Since wafer grinding is a mass production process, but the current inspection manpower cannot afford comprehensive inspection, most of them use sampling to inspect wafer accuracy or quality. However, this method cannot make real-time adjustments in response to abnormal phenomena in the processing process, so it is difficult to ensure that every chip meets the requirements.

In view of this, the present invention provides a method and an electronic device for real-time wafer processing quality estimation, which can be used to solve the above technical problems.

An embodiment of the present invention provides a method for real-time wafer processing quality estimation, which is suitable for an electronic device, comprising: obtaining at least one processing status signal generated by a wafer processing equipment when performing a processing operation on a wafer; and obtaining at least one real-time processing quality of the processing operation by inputting at least one processing status signal into at least one machine learning model, wherein at least one machine learning model estimates at least one real-time processing quality of the processing operation in response to the at least one processing status signal.

The present invention provides an electronic device, including a storage circuit and a processor. The storage circuit stores a program code. The processor is coupled to the storage circuit and accesses the program code to execute: obtaining at least one processing state signal generated by a wafer processing device when performing a processing operation on a wafer; and obtaining at least one real-time processing quality of the processing operation by inputting the at least one processing state signal into at least one machine learning model, wherein the at least one machine learning model estimates the at least one real-time processing quality of the processing operation in response to the at least one processing state signal.

Please refer to FIG. 1, which is a schematic diagram of an electronic device according to an embodiment of the present invention. In different embodiments, the electronic device 100 is, for example, various smart devices and/or computer devices. In some embodiments, the electronic device 100 can also be integrated into a wafer processing device to be used as a processing device/human-machine interface in the wafer processing device, but is not limited thereto.

In FIG1 , the electronic device 100 includes a storage circuit 102 and a processor 104. The storage circuit 102 is, for example, any type of fixed or removable random access memory (RAM), read-only memory (ROM), flash memory, hard disk or other similar devices or a combination of these devices, and can be used to record multiple program codes or modules.

The processor 104 is coupled to the storage circuit 102 and may be a general purpose processor, a special purpose processor, a conventional processor, a digital signal processor, a plurality of microprocessors, one or more microprocessors combined with a digital signal processor core, a controller, a microcontroller, an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), any other type of integrated circuit, a state machine, an Advanced RISC Machine (ARM) based processor, and the like.

In an embodiment of the present invention, the processor 104 can access the modules and program codes recorded in the storage circuit 102 to implement the method of real-time wafer processing quality estimation proposed by the present invention, the details of which are described as follows.

Please refer to FIG. 2 , which is a flowchart of a method for real-time wafer processing quality estimation according to an embodiment of the present invention. The method of this embodiment can be executed by the electronic device 100 of FIG. 1 . The following is a description of the details of each step of FIG. 2 with reference to the components shown in FIG. 1 .

First, in step S210, the processor 104 obtains a processing state signal generated by the wafer processing equipment when performing a processing operation on the wafer. In some embodiments, the processing operation performed by the wafer processing equipment on the wafer includes, for example, grinding the wafer by the wafer processing equipment.

Please refer to FIG. 3, which is a schematic diagram of a wafer processing device according to an embodiment of the present invention. In FIG. 3, the wafer processing device 300 may include a turntable 310 and a grinding wheel 320. In one embodiment, the turntable 310 may be used to carry a wafer 330 to be ground (e.g., having a radius of ), and can rotate Rotation, thereby driving the wafer 330 at a rotation speed The grinding wheel 320 (which has a radius ) may have a main shaft 321 and a grinding part 322, wherein the main shaft 321 may rotate at a speed The grinding portion 322 may be rotated, and abrasive grains for grinding the wafer 330 may be disposed on the grinding portion 322 .

In some embodiments, the grinding wheel 320 may have reference points A, B, and C, and the operator may adjust the positions of the reference points A, B, and C to set the tilt angle of the grinding wheel 320 when grinding the wafer 330. , Speed The feed rate of the grinding wheel 320 and the offset between the turntable 310 and the spindle 321 can also be set by the operator according to the needs. Thus, the wafer processing equipment 300 can perform corresponding grinding processing on the wafer 330 according to the above settings of the operator.

However, in some embodiments, if the operator fails to properly set the above values, after the wafer processing equipment 300 completes the grinding process on the wafer 330, the various specifications of the wafer 330 may not meet the requirements.

To avoid the above situation, the processor 104 can instantly obtain the processing status signal generated when the wafer processing equipment 300 performs a grinding process on the wafer 330 in step S210.

In different embodiments, the processing status signal may include various signals according to the designer's needs. In the first embodiment, a first accelerometer may be installed on the spindle 321. In this case, during the process of the grinding wheel 320 grinding the wafer 330, the spindle 321 will vibrate accordingly, and the first accelerometer will detect the corresponding first vibration signal (hereinafter referred to as V1). Based on this, during the process of the grinding wheel 320 grinding the wafer 330, the processor 104 can obtain the first vibration signal V1 from the first accelerometer as a processing status signal.

In the second embodiment, a second accelerometer may be installed on the main shaft of the turntable 310. In this case, when the grinding wheel 320 grinds the wafer 330, the main shaft of the turntable 310 will vibrate accordingly, and the second accelerometer will detect the corresponding second vibration signal (hereinafter referred to as V2). Therefore, when the grinding wheel 320 grinds the wafer 330, the processor 104 can obtain the second vibration signal V2 from the second accelerometer as a processing status signal.

In the third embodiment, a microphone device may be disposed near the wafer processing equipment 300, and the microphone device may be used to collect the sound generated by the wafer processing equipment 300 during the processing operation on the wafer 330. Therefore, during the process of the grinding wheel 320 grinding the wafer 330, the processor 104 may obtain a sound signal (hereinafter referred to as AU1) from the microphone device as a processing status signal.

In the fourth embodiment, the processor 104 may also take any combination of the first vibration signal V1, the second vibration signal V2 and the sound signal AU1 as the considered processing state signal, but is not limited thereto.

After obtaining the processing state signal under consideration, in step S220, the processor 104 obtains the real-time processing quality of the above-mentioned processing operation by inputting the processing state signal into the machine learning model.

In an embodiment of the present invention, the real-time processing quality includes, for example, at least one of the real-time surface roughness, total thickness variation, warp and curvature of the wafer 330. In some embodiments, the real-time processing quality can be understood as being estimated by a machine learning model based on a processing state signal.

In one embodiment, different real-time processing qualities may correspond to different machine learning models. For example, real-time surface roughness, total thickness variation, warp, and curvature may correspond to a first machine learning model, a second machine learning model, a third machine learning model, and a fourth machine learning model, respectively. In this case, the processor 104 may input the considered processing state signal into the first machine learning model, the second machine learning model, the third machine learning model, and the fourth machine learning model, respectively, so that the first machine learning model, the second machine learning model, the third machine learning model, and the fourth machine learning model provide estimated real-time surface roughness, total thickness variation, warp, and curvature in response to the processing state signal, respectively.

In another embodiment, different real-time processing qualities may also be estimated by the same machine learning model at the same time. For example, real-time surface roughness, total thickness variation, warp and curvature (or any combination thereof) may correspond to a specific machine learning model at the same time. In this case, the processor 104 may input the considered processing state signal into the specific machine learning model so that the specific machine learning model can simultaneously provide estimated real-time surface roughness, total thickness variation, warp and curvature in response to the processing state signal, but is not limited thereto.

In one embodiment, in order to enable each machine learning model to have the above capabilities, during the training process of each machine learning model, the designer can feed specially designed training data into each machine learning model so that each machine learning model can perform corresponding learning. For example, after obtaining a processing state signal that has been labeled as corresponding to a certain surface roughness (represented by R) (for example, measured from the process of grinding a certain wafer), the processor 104 can generate a corresponding feature vector based on it and feed it into the first machine learning model corresponding to the surface roughness. In this way, the first machine learning model can learn the relevant features of the processing state signal related to the above surface roughness (i.e., R) from this feature vector. In this case, when the first machine learning model receives the feature vector corresponding to the above-mentioned processing state signal in the future, the first machine learning model can accordingly estimate the surface roughness of the wafer being polished as R, but it is not limited to this.

For another example, after obtaining a processing state signal that has been labeled as corresponding to a total thickness variation (represented by T) (for example, measured from a grinding process of a wafer), the processor 104 can generate a corresponding feature vector and feed it into the second machine learning model corresponding to the total thickness variation. In this way, the second machine learning model can learn the relevant features of the processing state signal related to the total thickness variation (i.e., T) from this feature vector. In this case, when the second machine learning model receives the feature vector corresponding to the above-mentioned processing state signal in the future, the second machine learning model can accordingly estimate the total thickness variation of the wafer being ground as T, but it is not limited to this.

For another example, after obtaining a processing state signal (e.g., measured during a grinding process of a wafer) labeled as corresponding to a certain surface roughness (e.g., R), total thickness variation (e.g., T), curvature (represented by BW), and curvature (represented by W), the processor 104 can generate a corresponding feature vector and feed it into the corresponding specific machine learning model. In this way, the specific machine learning model can learn the relevant features of the processing state signal related to the above-mentioned surface roughness (i.e., R), total thickness variation (i.e., T), curvature (i.e., BW), and curvature (i.e., W) from the feature vector. In this case, when the specific machine learning model receives the feature vector corresponding to the above-mentioned processing state signal in the future, the specific machine learning model can accordingly estimate the surface roughness of the wafer being polished as R, the total thickness variation as T, the curvature as BW and the curvature as W, but it is not limited to this.

In one embodiment, the pattern/component of the processing state signal obtained by the processor 104 in step S210 must be the same as the training data used to train the machine learning model.

For example, if the training data used when training a machine learning model is composed of a third vibration signal corresponding to the first vibration signal V1 (for example, measured by the first accelerometer when the wafer processing equipment 300 performs a processing operation on a wafer), then the processing state signal obtained by the processor 104 in step S210 must be composed of the first vibration signal V1.

For another example, if the training data used when training a machine learning model is composed of the third vibration signal mentioned above and a fourth vibration signal corresponding to the second vibration signal V2 (for example, measured by the second accelerometer when the wafer processing equipment 300 performs a processing operation on a wafer), then the processing state signal obtained by the processor 104 in step S210 must be composed of the first vibration signal V1 and the second vibration signal V2.

For another example, if the training data used when training a machine learning model is composed of the third vibration signal, the fourth vibration signal and another sound signal corresponding to the sound signal AU1 (for example, measured by a microphone device when the wafer processing equipment 300 performs a processing operation on a wafer), then the processing status signal obtained by the processor 104 in step S210 must be composed of the first vibration signal V1, the second vibration signal V2 and the sound signal AU1.

In different embodiments, the above machine learning models can be implemented based on various existing machine learning algorithms according to the needs of the designer. For the convenience of explanation, it is assumed below that the first, second, third, and fourth machine learning models are all implemented based on the random forest algorithm, but it is not limited to this.

In an embodiment of the present invention, the processor 104 may first select one suitable for training the first machine learning model (corresponding to surface roughness) from a plurality of first hyperparameter combinations based on random search and 5-fold cross-validation. In one embodiment, the top 10 of the first hyperparameter combinations may be illustrated in Table 1 below, and the meaning of each parameter in Table 1 may be illustrated in Table 2. No. Parameter 1 Parameter 2 Parameter 3 Parameter 4 Parameter 5 Parameter 6 0 twenty four 18 77 True 0.091369 1 1 41 12 55 True 0.084644 2 2 121 20 33 True 0.082841 3 3 70 12 33 True 0.080626 4 4 63 14 66 True 0.080136 5 5 124 14 twenty two True 0.077544 6 6 144 8 55 True 0.077255 7 7 137 14 33 True 0.076922 8 8 180 16 44 True 0.073233 9 9 79 4 66 True 0.072590 10 Table 1 Parameters Meaning Parameter 1 (n_estimators) The number of trees in the forest. Default is 100. Parameter 2 (min_samples_split) The minimum number of samples required for internal node subdivision (i.e., how much data is required to subdivide) Parameter 3 (max_depth) Maximum depth of decision tree Parameter 4 (booststrap) The problem of bootstrapping, sampling from one's own sample to estimate the true distribution Parameter 5 (mean_test_score) Average test score Parameter 6 (rank_test_score) Test score ranking Table 2

In the scenario of Table 1, the first hyperparameter combination numbered 0 can be used as the optimal hyperparameter combination for training the first machine learning model, but is not limited thereto.

In one embodiment, the processor 104 may first select one suitable for training the second machine learning model (corresponding to TTV) from multiple second hyperparameter combinations based on random search and 5-fold cross validation. In one embodiment, the top 10 of the second hyperparameter combinations may be shown in Table 3 below, and the meaning of each parameter in Table 3 may refer to the content of Table 2. No. Parameter 1 Parameter 2 Parameter 3 Parameter 4 Parameter 5 Parameter 6 0 20 6 twenty two True 0.046149 1 1 121 2 twenty two True 0.042938 2 2 126 8 88 True 0.036442 3 3 6 10 33 True 0.033491 4 4 27 8 99 True 0.032685 5 5 148 14 twenty two True 0.032246 6 6 61 18 99 True 0.031406 7 7 113 2 66 True 0.031187 8 8 150 16 twenty two True 0.028614 9 9 179 4 33 True 0.027713 10 Table 3

In the scenario of Table 3, the second hyperparameter combination numbered 0 can be used as the optimal hyperparameter combination for training the second machine learning model, but is not limited to this.

In one embodiment, the processor 104 may first select one suitable for training the third machine learning model (corresponding to the curvature) from multiple third hyperparameter combinations based on random search and 5-fold cross validation. In one embodiment, the top 10 of the third hyperparameter combinations may be shown in Table 4 below, and the meaning of each parameter in Table 4 may refer to the content of Table 2. No. Parameter 1 Parameter 2 Parameter 3 Parameter 4 Parameter 5 Parameter 6 0 25 14 88 True 0.025933 1 1 128 4 1 True -0.009232 2 2 101 12 1 True -0.010955 3 3 156 18 1 True -0.014757 4 4 113 18 twenty two True -0.020440 5 5 101 2 1 True -0.020785 6 6 68 14 88 True -0.022272 7 7 176 12 1 True -0.022826 8 8 142 16 66 True -0.022875 9 9 127 18 1 True -0.022938 10 Table 4

In the scenario of Table 4, the third hyperparameter combination numbered 0 can be used as the optimal hyperparameter combination for training the third machine learning model, but is not limited thereto.

In one embodiment, the processor 104 may first select one suitable for training the fourth machine learning model (corresponding to the curvature) from multiple fourth hyperparameter combinations based on random search and 5-fold cross validation. In one embodiment, the top 10 of the fourth hyperparameter combinations may be shown in Table 5 below, and the meaning of each parameter in Table 5 may refer to the content of Table 2. No. Parameter 1 Parameter 2 Parameter 3 Parameter 4 Parameter 5 Parameter 6 0 6 12 55 True -0.304832 1 1 15 20 11 True -0.317027 2 2 11 20 twenty two True -0.417495 3 3 44 2 1 True -0.434407 4 4 28 6 110 True -0.438435 5 5 33 6 33 True -0.486140 6 6 8 18 1 True -0.501083 7 7 133 4 33 True -0.541256 8 8 67 2 44 True -0.552761 9 9 75 6 1 True -0.555465 10 Table 5

In the scenario of Table 5, the fourth hyperparameter combination numbered 0 can be used as the optimal hyperparameter combination for training the fourth machine learning model, for example, but is not limited to this.

2 again, after obtaining the real-time processing quality of the processing operation, the processor 104 may further determine whether the obtained real-time processing quality is abnormal. In one embodiment, in response to determining that the real-time processing quality is abnormal, the processor 104 may provide a warning message (step S230).

For example, if the real-time processing quality obtained is an estimated surface roughness, the processor 104 may compare the estimated surface roughness with the surface roughness threshold. If the estimated surface roughness is higher than the surface roughness threshold, the processor 104 may, for example, determine that the real-time processing quality is abnormal, and then provide a corresponding warning message for the relevant personnel to refer to and take corresponding improvement measures.

For another example, if the real-time processing quality obtained is the estimated TTV and curvature, the processor 104 may compare the estimated TTV and curvature with the TTV threshold and curvature threshold, respectively. If the estimated TTV is higher than the TTV threshold, or the estimated curvature is higher than the curvature threshold, the processor 104 may, for example, determine that the real-time processing quality is abnormal, and then provide corresponding warning messages for reference by relevant personnel and take corresponding improvement measures.

In summary, the embodiment of the present invention can determine the corresponding real-time processing quality according to the processing status signal generated by the wafer processing equipment when the wafer is processed. In this way, all wafers in processing can be inspected in real time. In addition, the embodiment of the present invention can also provide a warning message when it is determined that the real-time processing quality of the wafer is abnormal, so that relevant personnel can make corresponding adjustments accordingly, so that the wafer can meet the specification requirements.

Although the present invention has been disclosed as above by the embodiments, they are not intended to limit the present invention. Any person with ordinary knowledge in the relevant technical field can make some changes and modifications without departing from the spirit and scope of the present invention. Therefore, the protection scope of the present invention shall be defined by the scope of the attached patent application.

100: electronic device 102: storage circuit 104: processor 300: wafer processing equipment 310: turntable 320: grinding wheel 321: spindle 322: grinding part 330: wafer S210~S230: steps A, B, C: reference point , :Radius , :Speed

FIG. 1 is a schematic diagram of an electronic device according to an embodiment of the present invention. FIG. 2 is a flow chart of a method for real-time wafer processing quality estimation according to an embodiment of the present invention. FIG. 3 is a schematic diagram of a wafer processing equipment according to an embodiment of the present invention.

S210~S230: Steps

Claims (5)

A method for real-time wafer processing quality estimation, suitable for an electronic device, comprises: obtaining a processing state signal generated by a wafer processing device when performing a processing operation on a wafer, wherein the processing state signal comprises one of a first vibration signal, a second vibration signal and a sound signal, and the step of obtaining the processing state signal generated by the wafer processing device when performing the processing operation on the wafer comprises: obtaining the first vibration signal from a first accelerometer installed on a grinding wheel spindle of the wafer processing device; obtaining the second vibration signal from a second accelerometer installed on a turntable spindle of a turntable, wherein the turntable carries the wafer; or obtaining the sound signal from a microphone device, wherein the microphone device is used to collect the sound signal generated by the wafer processing device when performing the processing operation on the wafer. The sound generated during the processing operation; and obtaining at least one real-time processing quality of the processing operation by inputting the processing state signal into at least one machine learning model, including: inputting the processing state signal into a first machine learning model, wherein the first machine learning model provides real-time surface roughness in response to the processing state signal; inputting the processing state signal into a second machine learning model, wherein the second machine learning model provides real-time total thickness variation in response to the processing state signal; inputting the processing state signal into a third machine learning model, wherein the third machine learning model provides real-time curvature in response to the processing state signal; and inputting the processing state signal into a fourth machine learning model, wherein the fourth machine learning model provides real-time curvature in response to the processing state signal. The method as claimed in claim 1, wherein the processing operation includes a grinding process performed by the grinding wheel of the wafer processing equipment on the wafer located on the turntable. The method as claimed in claim 1 further includes: providing a warning message in response to determining that at least one real-time processing quality is abnormal. The method as claimed in claim 1, wherein each of the machine learning models is trained based on a hyperparameter combination selected from a plurality of first hyperparameter combinations by random search and 5-fold cross validation. An electronic device includes: a storage circuit storing a program code; and a processor coupled to the storage circuit and accessing the program code to execute: obtaining a processing state signal generated by a wafer processing device when performing a processing operation on a wafer, wherein the processing state signal includes one of a first vibration signal, a second vibration signal and a sound signal, and obtaining the processing state signal generated by the wafer processing device when performing a processing operation on a wafer. The processing state signal generated when the wafer is subjected to the processing operation includes: obtaining the first vibration signal from a first accelerometer mounted on a grinding wheel spindle of the wafer processing equipment; obtaining the second vibration signal from a second accelerometer mounted on a turntable spindle of a turntable, wherein the turntable carries the wafer; or obtaining the sound signal from a microphone device, wherein the microphone device is used to collect the sound generated by the wafer processing equipment during the processing operation on the wafer; and obtaining at least one real-time processing quality of the processing operation by inputting the processing state signal into at least one machine learning model, including: inputting the processing state signal into the first machine learning model, wherein the first machine learning model provides real-time surface roughness in response to the processing state signal; inputting the processing state signal into the second machine learning model; The processing state signal is input into a third machine learning model, wherein the second machine learning model provides real-time total thickness variation in response to the processing state signal; the processing state signal is input into a third machine learning model, wherein the third machine learning model provides real-time curvature in response to the processing state signal; and the processing state signal is input into a fourth machine learning model, wherein the fourth machine learning model provides real-time curvature in response to the processing state signal.

Images