Training method for semiconductor process prediction model, semiconductor process prediction device, and semiconductor process prediction method

This application claims the benefit of People's Republic of China application Serial No. 202210070302.0, filed Jan. 21, 2022, the disclosure of which is incorporated by reference herein in its entirety.

The disclosure relates in general to a training method for a process prediction model, a process prediction device, and a process prediction method, and more particularly to a training method for a semiconductor process prediction model, a semiconductor process prediction device, and a semiconductor process prediction method.

With the high development of semiconductor technology, various complex semiconductor components are constantly being introduced. In the semiconductor manufacturing process, a wafer needs to go through thousands of processes to produce the final product. Therefore, researchers need to perform various tests on semi-finished products. The electrical measurement data is a very important test, which can test the electrical parameters of NMOS, PMOS, resistor, contact resistor or internal connection on the wafer.

The electrical measurement data can be used to determine whether there is an abnormality, and can also be used to estimate the yield of the final product. The electrical measurement data has a certain effect on the improvement of yield. However, if abnormalities can be detected earlier in the process, it will be more helpful to improve the yield. In order to detect anomalies as early as possible, researchers expect that the electrical measurement data can also be predicted.

The disclosure is directed to a training method for a semiconductor process prediction model, a semiconductor process prediction device, and a semiconductor process prediction method. During the training process, the samples having physical defects are not considered to improve the accuracy of the semiconductor process prediction model. After improving the accuracy of the semiconductor process prediction model, the electrical measurement data can be accurately predicted for the wafers without physical defects, and the wafers with physical defects can also be found to be abnormal, which is very helpful to improve the yield.

According to one embodiment, a training method for a semiconductor process prediction model. The training method includes the following steps. A semiconductor process is performed on a plurality of samples. A plurality of process data of the samples are obtained. A plurality of electrical measurement data of the samples are obtained. Some of the samples having at least one physical defect are filtered out according to the process data. The semiconductor process prediction model is trained according to the process data and the electrical measurement data of the samples after filtering.

According to another embodiment, a semiconductor process prediction device is provided. The semiconductor process prediction device includes a process data receiving unit, an electrical data receiving unit, a filtering unit and a semiconductor process prediction model. The process data receiving unit is configured to obtain a plurality of process data of a plurality of samples which are performed a semiconductor process. The electrical data receiving unit is configured to obtain a plurality of electrical measurement data of the samples. The filtering unit is configured to filter out some of the samples having at least one physical defect according to the process data. The semiconductor process prediction model is trained according to the process data and the electrical measurement data of the samples after filtering.

According to an alternative embodiment, a prediction method of a semiconductor process is provided. The prediction method includes the following steps. A semiconductor process is performed on at least one wafer. A plurality of process data of the wafer is obtained. Whether the wafer has at least one physical defect is determined according to the process data. If the wafer has the physical defect, a prediction is terminated. If the wafer does not have the physical defect, the prediction is performed, via a semiconductor process prediction model according to the process data, to predict an electrical measurement data.

FIG. 1 shows a schematic diagram of a semiconductor process according to an embodiment.

FIG. 2 shows a block diagram of a semiconductor process prediction device 1000 according to an embodiment.

FIG. 3 shows a flowchart of the training method of the semiconductor process prediction model according to an embodiment.

FIG. 4 shows an example of performing the steps in FIG. 3 .

FIG. 5 is a flowchart illustrating a prediction method of a semiconductor process according to an embodiment.

FIG. 6 shows an example of the steps in FIG. 5 .

FIG. 7 illustrates another example of the steps in FIG. 5 .

In the following detailed description, for purposes of explanation, numerous specific details are set forth in order to provide a thorough understanding of the disclosed embodiments. It will be apparent, however, that one or more embodiments may be practiced without these specific details. In other instances, well-known structures and devices are schematically shown in order to simplify the drawing.

Please refer to FIG. 1 , which shows a schematic diagram of a semiconductor process according to an embodiment. In the semiconductor manufacturing process, a wafer WF 1 needs to go through thousands of processes to produce the final product. Researchers need to perform various tests on semi-finished products. After some circuits and component structures have built through several processes, an electrical measurement can be performed. The electrical measurement can test the electrical parameters of NMOS, PMOS, resistor, contact resistor or internal connection on the wafer WF 1 . Electrical measurement data WT 1 can be used to determine whether there is any abnormality in these processes, and can also be used to estimate the yield of the final product. After confirming the electrical measurement data WT 1 , the wafer WF 1 may need to continue with other stages of the process. Then, after several processes are completed, the electrical measurement is performed again.

In order to detect abnormality as early as possible, a semiconductor process prediction model is provided in the present embodiment. Before the process at this stage is completed (that is, before the electrical measurement is performed), the semiconductor process prediction model can predict the electrical measurement data to detect abnormalities in advance. Before the process at this stage is completed (that is, before the electrical measurement is performed), the semiconductor process prediction model can predict the electrical measurement data to detect abnormalities in advance.

Please refer to FIG. 2 , which shows a block diagram of a semiconductor process prediction device 1000 according to an embodiment. The semiconductor process prediction device 1000 includes a process data receiving unit 120 , an electrical data receiving unit 130 , a filtering unit 140 , a semiconductor process prediction model 150 and a determination unit 230 . The functions of the components are outlined below. The process data receiving unit 120 and the electrical data receiving unit 130 are used to obtain data. The process data receiving unit 120 and the electrical data receiving unit 130 are, for example, wired network transmission modules, wireless network transmission modules or LTE transmission modules. The filtering unit 140 is used for data filtering. The semiconductor process prediction model 150 is a machine learning model. The determination unit 230 is used for data judgment. The filtering unit 140 , the semiconductor process prediction model 150 and/or the determination unit 230 are, for example, a circuit, a chip, a circuit board, a code, a computer program product, or a recording medium storing the code.

The semiconductor process prediction model 150 can predict the electrical measurement data. During performing the training method of the semiconductor process prediction model 150 in this embodiment, some of the samples having physical defects will be filtered out to ensure that non-process factors will not affect the accuracy of the prediction results. Moreover, during performing the execution method of the semiconductor process prediction model 150 in this embodiment, the prediction will be made only when the wafer to be tested has no physical defect, so as to ensure the accuracy of the prediction result. The following describes the training method first, and then the execution method.

Please refer to FIGS. 2 to 4 . FIG. 3 shows a flowchart of the training method of the semiconductor process prediction model 150 according to an embodiment. FIG. 4 shows an example of performing the steps in FIG. 3 . In step S 110 , as shown in FIG. 4 , the semiconductor process is performed on a plurality of samples WF 21 , WF 22 , WF 23 , WF 24 , etc. The semiconductor process is, for example, etching, deposition, annealing, ion implantation, etc. In one embodiment, the samples WF 21 , WF 22 , WF 23 , WF 24 , . . . may be the same lot belonging to the same product. Alternatively, the samples WF 21 , WF 22 , WF 23 , WF 24 , . . . can be different batches belonging to the same product. Or, the samples WF 21 , WF 22 , WF 23 , WF 24 , . . . can be different products manufactured by different process sequences and different process.

Then, in step S 120 , the process data receiving unit 120 obtains a plurality of process data MT 21 ( 1 ), MT 21 ( 2 ), . . . , MT 22 ( 1 ), MT 22 ( 2 ), . . . , MT 23 ( 1 ), MT 23 ( 2 ), . . . , MT 24 ( 1 ), MT 24 ( 2 ), . . . of the samples WF 21 , WF 22 , WF 23 , WF 24 , etc. The process data MT 21 ( 1 ), MT 21 ( 2 ), . . . , MT 22 ( 1 ), MT 22 ( 2 ), . . . , MT 23 ( 1 ), MT 23 ( 2 ), . . . , MT 24 ( 1 ), MT 24 ( 2 ), . . . are, for example, physical measurement data. The physical measurement data is, for example, the measurement data (metrology data), such as width and thickness, detected by an optical microscope, an electron microscope or an ion microscope.

In another embodiment, the process data MT 21 ( 1 ), MT 21 ( 2 ), . . . , MT 22 ( 1 ), MT 22 ( 2 ), . . . , MT 23 ( 1 ), MT 23 ( 2 ), . . . , MT 24 ( 1 ), MT 24 ( 2 ), . . . further include equipment setting data or equipment detecting data. The equipment setting data is, for example, the temperature set in the equipment, the pressure set in the equipment, the processing time set in the equipment, the gas used in the equipment, the gas flow set in the equipment, and so on. The equipment detecting data is, for example, the temperature detected by the equipment, the pressure detected by the equipment, the wavelength of light measured by the equipment, and so on.

Then, in step S 130 , as shown in FIG. 4 , the electrical data receiving unit 130 obtains a plurality of electrical measurement data WT 21 , WT 22 , WT 23 , WT 24 , etc. of the samples WF 21 , WF 22 , WF 23 , WF 24 , etc. The electrical measurement data WT 21 , WT 22 , WT 23 , WT 24 , etc., such as Wafer Acceptance Test (WAT), which are used to test the electrical parameters of NMOS, PMOS, resistor, contact resistor or internal connection in the wafers WF 21 , WF 22 , WF 23 , WF 24 , etc.

Then, in step S 140 , as shown in FIG. 4 , the filtering unit 130 filter out some of the samples WF 21 , WF 22 , WF 23 , WF 24 , . . . having at least one physical defect according to the process data MT 21 ( 1 ), MT 21 ( 2 ), . . . , MT 22 ( 1 ), MT 22 ( 2 ), . . . , MT 23 ( 1 ), MT 23 ( 2 ), . . . , MT 24 ( 1 ), MT 24 ( 2 ), etc. The physical defect is, for example, a particle pt, a scratch sc, or a crack cr, etc. which are detected by an optical microscope, an electron microscope or an ion microscope. For example, the process data MT 22 ( 2 ) is the particle pt. When the particle pt is located between two wires M 1 , M 2 , it will cause a short circuit. Therefore, the particles pt will have a significant impact on the electrical measurement data WT 22 . The process data MT 22 ( 1 ) is the scratch sc. The scratch sc may destroy the element structure. Therefore, the scratch sc will have a significant impact on the electrical measurement data WT 22 . The process data MT 23 ( 2 ) is the crack cr. The crack cr may damage the wiring or cause wafer breaking. Therefore, the crack cr also has a significant effect on the electrical measurement data WT 23 . The filtering unit 130 filters out the samples WF 22 , WF 23 having physical defects according to the scratch sc, the particle pt, the crack cr in the process data MT 22 ( 1 ), MT 22 ( 2 ), MT 23 ( 2 ), and keeps the samples WF 21 , WF 24 , etc. That is to say, the process data MT 22 ( 1 ), MT 22 ( 2 ), . . . , MT 23 ( 1 ), MT 23 ( 2 ), . . . and the electrical measurement data WT 22 , WT 23 of the samples WF 22 , WF 23 are filtered out.

Then, in step S 150 , as shown in FIG. 4 , the semiconductor process prediction model 150 is trained according to the process data MT 21 ( 1 ), MT 21 ( 2 ), . . . , MT 24 ( 1 ), MT 24 ( 2 ), etc. and the electrical measurement data WT 21 , WT 24 , etc. of the samples WF 21 , WF 24 , etc. after filtered. During training, the samples WF 22 , WF 23 having physical defects are not considered. The physical defects are accidental events, not normal events in the process. Therefore, after filtering out the samples WF 22 , WF 23 having physical defects, the prediction results of the semiconductor process prediction model 150 will not be biased by the accidental events.

Please refer to FIGS. 5 to 6 . FIG. 5 is a flowchart illustrating a prediction method of a semiconductor process according to an embodiment. FIG. 6 shows an example of the steps in FIG. 5 . In step S 210 , as shown in FIG. 6 , a semiconductor process is performed on at least one wafer WF 3 . The semiconductor processes is, for example, etching, deposition, annealing, or ion implantation etc. In one embodiment, the wafer WF 3 may be performed the process which is the same as the process performed on the training samples of the semiconductor process prediction model 150 . Alternatively, the wafer WF 3 can be performed the process which is fine-tuned from the process performed on the training samples of the semiconductor process prediction model 150 . Or, the wafer WF 3 can be performed the process which is different from the process performed on the training samples of the semiconductor process prediction model 150 .

Then, in step S 220 , as shown in FIG. 6 , the process data receiving unit 120 obtains a plurality of process data MT 3 ( 1 ), MT 3 ( 2 ), etc. of the wafer WF 3 . The process data MT 3 ( 1 ), MT 3 ( 2 ), etc. are, for example, physical measurement data. The physical measurement data is, for example, the measurement data (metrology data), such as width and thickness, detected by an optical microscope, an electron microscope or an ion microscope.

In another embodiment, the process data MT 3 ( 1 ), MT 3 ( 2 ), etc. may further include an equipment setting data or an equipment detecting data. The equipment setting data is, for example, the temperature set in the equipment, the pressure set in the equipment, the processing time set in the equipment, the gas used in the equipment, the gas flow set in the equipment, and so on. The equipment detecting data is, for example, the temperature detected by the equipment, the pressure detected by the equipment, the wavelength of light measured by the equipment, and so on.

Then, in step S 230 , as shown in FIG. 6 , the determination unit 230 determines whether the wafer WF 3 has at least one physical defect according to the process data MT 3 ( 1 ), MT 3 ( 2 ), etc. The physical defects are, for example, the particle pt (shown in FIG. 4 ), the scratch sc (shown in FIG. 4 ), or the crack cr (shown in FIG. 4 ) detected by the optical microscope, the electron microscope or the ion microscope. As shown in FIG. 6 , the wafer WF 3 has no physical defect, so the process proceeds to the step S 240 .

In step S 240 , the semiconductor process prediction model 150 performs the prediction to predict the electrical measurement data WT 3 ′ according to the process data MT 3 ( 1 ), MT 3 ( 2 ), etc. The training process of the semiconductor process prediction model 150 does not consider the samples having physical defects. The physical defects are accidental events, not normal events in the process. Therefore, after filtering out the samples having physical defects, the predictions of the semiconductor process prediction model 150 will not be biased by the accidental events. Therefore, the semiconductor process prediction model 150 can accurately predict the electrical measurement data WT 3 ′ in this step.

Then, in step S 250 , whether the electrical measurement data WT 3 ′ is abnormal is determined. If the electrical measurement data WT 3 ′ is abnormal, then the process proceeds to the step S 260 .

In step S 260 , an abnormal elimination operation is executed. The abnormal elimination operation is, for example, the machine inspection, the machine parameter adjustment or the recipe adjustment to avoid the occurrence of a large number of defective products.

Through the above prediction method, the wafer WF 3 without physical defects can be accurately predicted the electrical measurement data WT 3 ′, so that when the electrical measurement data WT 3 ′ is abnormal, the machine detection, the machine parameter adjustment or the process recipe adjustment can be executed to avoid the occurrence of a large number of defective products.

Please refer to FIG. 5 and FIG. 7 . FIG. 7 illustrates another example of the steps in FIG. 5 . In step S 210 , as shown in FIG. 7 , a semiconductor process is performed on at least one wafer WF 4 . The semiconductor process is, for example, etching, deposition, annealing, ion implantation, etc. In one embodiment, the wafer WF 4 may be performed the process which is the same as the process performed on the training samples of the semiconductor process prediction model 150 . Alternatively, the wafer WF 4 can be performed the process which is fine-tuned from the process performed on the training samples of the semiconductor process prediction model 150 . Or, the wafer WF 4 can be performed the process which is different from the process performed on the training samples of the semiconductor process prediction model 150 .

Then, in step S 220 , as shown in FIG. 7 , the process data receiving unit 120 obtains a plurality of process data MT 4 ( 1 ), MT 4 ( 2 ), etc. of the wafer WF 4 . The process data MT 4 ( 1 ), MT 4 ( 2 ), etc. are, for example, physical measurement data. The physical measurement data is, for example, the measurement data (metrology data), such as width and thickness detected by an optical microscope, an electron microscope or an ion microscope.

In another embodiment, the process data MT 4 ( 1 ), MT 4 ( 2 ), eta. may further include the equipment setting data or the equipment detecting data. The equipment setting data is, for example, the temperature set in the equipment, the pressure set in the equipment, the processing time set in the equipment, the gas used in the equipment, the gas flow set in the equipment, and so on. The equipment detecting data is, for example, the temperature detected by the equipment, the pressure detected by the equipment, the wavelength of light measured by the equipment, and so on.

Then, in step S 230 , as shown in FIG. 7 , the determination unit 230 determines whether the wafer WF 4 has at least one physical defect according to the process data MT 4 ( 1 ), MT 4 ( 2 ), etc. The physical defect is, for example, the particle pt (shown in FIG. 4 ), the scratch sc (shown in FIG. 4 ), the crack cr (shown in FIG. 4 ) detected by the optical microscope, the electron microscope or the ion microscope. As shown in FIG. 7 , the wafer WF 4 has the physical defect, i.e., the particles pt, so the process proceeds to step S 260 .

In step S 260 , an abnormal elimination operation is executed. The abnormal elimination operation is, for example, the fixture adjustment or the carrier adjustment, to avoid the occurrence of a large number of defective products.

Through the above prediction method, the wafer WF 4 with physical defects can also be found abnormal, and the fixture adjustment and the vehicle adjustment can be executed immediately to avoid the occurrence of a large number of defective products.

Through the above embodiment, in the training process of the semiconductor process prediction model 150 , the samples having physical defects are not considered. The physical defects are accidental events, not normal events in the process. Therefore, after filtering out the samples having physical defects, the prediction of the semiconductor process prediction model 150 will not be biased by the accidental events.

After improving the accuracy of the semiconductor process prediction model 150 , the electrical measurement data can be accurately predicted for the wafers without physical defects, and the wafers with physical defects can also be found to be abnormal, which is very helpful to improve the yield.

It will be apparent to those skilled in the art that various modifications and variations can be made to the disclosed embodiments. It is intended that the specification and examples be considered as exemplary only, with a true scope of the disclosure being indicated by the following claims and their equivalents.