Probe control device and probe control method

Abstract

To provide a prober control device and a prober control method that can more accurately predict the tip position of a probe needle.SOLUTION: A prober controlling device that brings a probe needles (35) into contact with an inspection target for electrical characteristics includes: an input data acquisition unit (42) that acquires input data including temperature data of at least one of a probe card (26) and a card holder (25); a prediction unit (46) that predicts a tip position of the probe needle on the basis of the input data acquired by the input data acquisition unit, using a prediction model (47) that receives the input data as input and outputs the tip position of the probe needle; and a determination unit (50) that determines whether or not to execute prediction by the prediction unit on the basis of the input data used as training data for machine learning of the prediction model and the input data acquired by the input data acquisition unit, before the prediction by the prediction unit.SELECTED DRAWING: Figure 5

Description

Technical Field

[0001] The present invention relates to a prober control device for a prober, a prober control method, and a prober used for inspecting electrical characteristics of semiconductor chips formed on a wafer.

Background Art

[0002] A plurality of semiconductor chips each having the same electric element circuit are formed on a surface of a wafer. Each semiconductor chip is inspected for electrical characteristics by a wafer test system before being individually cut by a dicer. This wafer test system includes a prober and a tester (see Patent Documents 1 to 4).

[0003] While holding a wafer on a wafer chuck, the prober relatively moves a probe card having probe needles and the wafer chuck, thereby bringing the probe needles into electrical contact with electrode pads of the semiconductor chips. The tester supplies various test signals to the semiconductor chips via terminals connected to the probe needles, and receives and analyzes signals output from the semiconductor chips to test whether the semiconductor chips operate normally.

[0004] Semiconductor chips are used in a wide range of applications and are used over a wide temperature range. Therefore, inspection of semiconductor chips needs to be performed, for example, at room temperature (normal temperature), high temperature, and low temperature. For this reason, the wafer chuck of the prober is provided with a temperature adjustment unit such as a heater mechanism, a chiller mechanism, or a heat pump mechanism, and the temperature adjustment unit heats or cools the wafer held on the wafer chuck.

[0005] During this process, the temperature of each part of the prober other than the wafer chuck gradually changes to approach the temperature of the wafer chuck. As a result, each part deforms due to thermal expansion from heating or contraction from cooling, and the relative position between the probe needle and the semiconductor chip changes with this deformation. Consequently, when the probe needle and wafer are moved relative to each other to inspect the semiconductor chip, there is a risk of probing errors occurring where the probe needle does not properly contact the semiconductor chip.

[0006] Therefore, Patent Document 1 discloses a prober that attaches a temperature sensor to a probe card having probe needles, and corrects the height position of the wafer chuck when the probe needles are brought into contact with a semiconductor chip based on the measurement results of the temperature sensor. In the prober described in Patent Document 1, by pre-determining the relationship between the temperature of the probe card and the amount of displacement of the probe needles in the height direction, the amount of correction for the height position of the wafer chuck can be determined from the measurement results of the temperature sensor.

[0007] Patent Document 2 discloses a prober equipped with temperature sensors on the probe card and the X-direction moving stage, which contacts the semiconductor chip with the probe needle when the temperature of a predetermined part of the prober is stable based on the measurement results of the temperature sensors. According to the prober described in Patent Document 2, the preheating time for heating the wafer and probe card can be shortened.

[0008] Patent Document 3 discloses a prober that attaches temperature sensors to a wafer chuck, a card holder for holding probe cards, and a head stage for holding the card holder, and corrects the contact position between the probe needle and the semiconductor chip based on the measurement results of each temperature sensor. In this prober of Patent Document 3, the relationship between the temperature of the wafer chuck and the card holder and the position of the probe needle is determined in advance, and a predictive model showing the change in the position of the probe needle due to changes in each temperature is generated. As a result, the prober of Patent Document 3 can correct the contact position between the probe needle and the semiconductor chip by referring to the predictive model based on the temperature measurement results of each temperature sensor.

[0009] Patent Document 4 discloses a prober that measures the temperature of both the probe card and the card holder, and predicts the tip position of the probe needle by referring to a predictive model that shows the relationship between the temperatures of both and the tip position of the probe needle displaced by the thermal deformation of both, based on the temperature measurement results of both. According to the prober described in Patent Document 4, the probe needle can be made to contact the semiconductor chip efficiently and stably. [Prior art documents] [Patent Documents]

[0010] [Patent Document 1] Japanese Patent Publication No. 2006-173206 [Patent Document 2] Japanese Patent Publication No. 2005-228788 [Patent Document 3] Japanese Patent Publication No. 2007-311389 [Patent Document 4] Japanese Patent Publication No. 2018-117095 [Overview of the project] [Problems that the invention aims to solve]

[0011] In the probers described in Patent Documents 3 and 4 above, a prediction model is generated in advance. However, if the prediction model is generated when the temperature inside the prober is unstable, such as immediately after a temperature change in the wafer chuck, a discrepancy will occur between the predicted value of the probe needle tip position by the prediction model and the measured value of the probe needle tip position. As a result, probing errors may occur. Therefore, in this case, it is necessary to continue generating (learning) the prediction model for a long time until the temperature inside the prober stabilizes.

[0012] Furthermore, even if a prediction model is used that is generated under stable conditions within the prober, if temperature sensor drift occurs over a long period of time, or if the tip position of the probe needle is displaced due to temperature fluctuations in areas within the prober where temperature measurements are not being taken, a discrepancy may occur between the predicted value and the measured value of the probe needle tip position.

[0013] This invention has been made in view of these circumstances, and aims to provide a prober control device, a prober control method, and a prober that can more accurately predict the tip position of the probe needle. [Means for solving the problem]

[0014] A probe control device for achieving the object of the present invention includes a wafer chuck for holding a wafer on which a plurality of semiconductor chips are formed, a probe card having probe needles, a card holder for holding the outer circumference of the probe card and facing the wafer, and a relative movement unit for moving the wafer chuck relative to the probe needles, and a probe control device for driving the relative movement unit of the probe to bring the probe needles into contact with the semiconductor chips, comprising: an input data acquisition unit for acquiring input data including temperature data of at least one of the probe card and the card holder; a prediction unit for predicting the tip position of the probe needles based on the input data acquired by the input data acquisition unit using a prediction model that takes the input data as input and the tip position of the probe needles as output; and a decision unit that determines whether or not to perform a prediction by the prediction unit based on the input data used as training data for machine learning of the prediction model and the input data acquired by the input data acquisition unit before the prediction by the prediction unit.

[0015] This probe control device allows for determining, before the prediction unit makes a prediction, whether the prediction model used for the prediction can accurately predict the tip position of the probe needle based on the current input data.

[0016] In another embodiment of the present invention, the probe control device includes the following processes: for each parameter of the input data, the determination unit calculates the difference between the input data acquired by the input data acquisition unit and the input data used as training data; and for each parameter, it calculates the root of the sum of the squares of the differences and determines whether or not the prediction unit can perform a prediction based on whether or not there is at least one root of the sum of the squares of the differences for each parameter that falls within a predetermined range.

[0017] In another aspect of the present invention, a probe control device includes a needle position acquisition unit that acquires the tip position of the probe needle when the determination unit determines no, and a retraining unit that retrains a prediction model using training data which is the sum of the input data acquired by the input data acquisition unit and the tip position of the probe needle acquired by the needle position acquisition unit. The needle position acquisition unit, the retraining unit, the input data acquisition unit, and the determination unit operate repeatedly until the determination unit determines yes. This makes it possible to accurately predict the tip position of the probe needle.

[0018] In another aspect of the present invention, the probe control device excludes the oldest input data and the tip position of the probe needle corresponding to the input data from the training data, and then retrains the prediction model based on the training data. This reduces the influence of drift from the input data acquisition unit (temperature sensor).

[0019] In another embodiment of the present invention, a probe control device is provided in which, when a determination unit determines that a probe is permissible, a prediction unit predicts the tip position of the probe needle, and a movement control unit controls a relative movement unit based on the tip position of the probe needle predicted by the prediction unit to bring the probe needle into contact with the semiconductor chip. This makes it possible to make correct contact between the probe needle and the semiconductor chip.

[0020] In a prober control device according to another aspect of the present invention, an input data acquisition unit acquires, as input data, alignment data including at least any one of, in addition to temperature data, a chip size of a semiconductor chip, a position of a wafer, and a positional relationship between a first camera used for detecting a semiconductor chip and a second camera used for detecting a probe needle.

[0021] A prober for achieving the object of the present invention comprises: a wafer chuck that holds a wafer on which a plurality of semiconductor chips are formed; a probe card having probe needles; a card holder that holds an outer periphery of the probe card and causes the probe card to face the wafer; a relative movement unit that relatively moves the wafer chuck with respect to the probe needles; and the prober control device described above.

[0022] A prober control method for achieving the object of the present invention is a prober control method for driving a relative movement unit of a prober to bring probe needles into contact with semiconductor chips, wherein the prober comprises a wafer chuck that holds a wafer on which a plurality of semiconductor chips are formed, a probe card having probe needles, a card holder that holds an outer periphery of the probe card and causes the probe card to face the wafer, and a relative movement unit that relatively moves the wafer chuck with respect to the probe needles. The method comprises: an input data acquisition step of acquiring input data including temperature data of at least one of the probe card and the card holder; a prediction step of predicting a tip position of the probe needle using a prediction model that receives the input data and outputs the tip position of the probe needle, based on the input data acquired in the input data acquisition step; and a determination step, performed before the prediction step, of determining whether execution of the prediction step is allowable based on input data used as teacher data for machine learning of the prediction model and the input data acquired in the input data acquisition step.

Effects of the Invention

[0023] The present invention can more accurately predict the tip position of a probe needle. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] [Figure 1] This is a schematic diagram of a prober used in a wafer test system to inspect the electrical characteristics of multiple semiconductor chips formed on a wafer. [Figure 2] This is a perspective view of the Prova. [Figure 3] This is a top view of a wafer held in a wafer chuck. [Figure 4] This is an explanatory diagram showing an example of temperature measurement points for a card holder and probe card using a temperature sensor. [Figure 5] This is a functional block diagram showing the functions of the prober's control unit. [Figure 6] This is an explanatory diagram showing an example of training data used for machine learning of prediction models by the prediction model generation unit. [Figure 7] This flowchart shows the flow of how a probe needle makes contact with a semiconductor chip using a prober. [Figure 8] In a comparative example where the decision unit does not perform decision and prediction model retraining, the graph shows the predicted and measured values ​​of the probe needle tip position (see symbol VIIIA), and the graph shows the difference between the predicted and measured values ​​of the probe needle tip position (see symbol VIIIB). [Figure 9] In this embodiment, the graph shows the predicted and measured values ​​of the probe needle tip position (see symbol IXA), and the graph shows the difference between the predicted and measured values ​​of the probe needle tip position (see symbol IXB). [Modes for carrying out the invention]

[0025] [Prover configuration] Figure 1 is a schematic diagram of a prober 10 used in a wafer test system for inspecting the electrical characteristics of multiple semiconductor chips 9 (see Figure 3) formed on a wafer W. Figure 2 is an external perspective view of the prober 10.

[0026] As shown in Figures 1 and 2, the prober 10 comprises a base 12, a Y-stage 13, a Y-movement unit 14, an X-stage 15, an X-movement unit 16, a Zθ-stage 17, a Zθ-movement unit 18, a wafer chuck 20, a support column 23 (see Figure 2), a head stage 24 (see Figure 2), a card holder 25, a probe card 26, a wafer alignment camera 29, upper and lower stages 30, a needle alignment camera 31, a cleaning plate 32, and a temperature sensor 34. Note that the external configuration of the prober 10 is not limited to the examples shown in Figures 1 and 2 and can be changed as appropriate.

[0027] The Y-stage 13 is supported on the upper surface of the base 12 via the Y-movement section 14 so as to be movable in the Y-axis direction.

[0028] The Y-movement unit 14 includes, for example, a guide rail provided on the upper surface of the base 12 and parallel to the Y-axis, a slider provided on the lower surface of the Y-stage 13 and engaging with the guide rail, and an actuator such as a motor that moves the Y-stage 13 in the Y-axis direction. This Y-movement unit 14 moves the Y-stage 13 in the Y-axis direction on the base 12.

[0029] An X-stage 15 is supported on the upper surface of the Y-stage 13 via an X-movement unit 16 so as to be movable in the X-axis direction. The X-movement unit 16 includes, for example, a guide rail provided on the upper surface of the Y-stage 13 and parallel to the X-axis, a slider provided on the lower surface of the X-stage 15 and engaging with the guide rail, and an actuator such as a motor that moves the X-stage 15 in the X-axis direction. This X-movement unit 16 moves the X-stage 15 on the Y-stage 13 in the X-axis direction.

[0030] The upper surface of the X-stage 15 is provided with a Zθ stage 17 and upper and lower stages 30. The Zθ stage 17 is provided with a Zθ moving part 18. A wafer chuck 20 is held on the upper surface of the Zθ stage 17 via the Zθ moving part 18.

[0031] The Zθ moving unit 18 includes, for example, a lifting mechanism that moves the Zθ stage 17 in the Z-axis direction (vertical direction) and a rotation mechanism that rotates the Zθ stage 17 around the Z-axis. Therefore, the Zθ moving unit 18 moves the wafer chuck 20, which is held on the upper surface of the Zθ stage 17, in the Z-axis direction and rotates it around the Z-axis.

[0032] The wafer W is held on the upper surface of the wafer chuck 20 by various holding methods such as vacuum suction. The wafer chuck 20 is also provided with a temperature control unit 20a for adjusting the temperature of the wafer W. Known mechanisms such as a heater mechanism, a chiller mechanism, and a heat pump mechanism can be used as this temperature control unit 20a. The temperature control unit 20a adjusts the temperature of the wafer W held in the wafer chuck 20.

[0033] The wafer chuck 20 is supported so as to be movable in the XYZ directions and so as to be rotatable around the Z axis, via the Y stage 13, Y movement unit 14, X stage 15, X movement unit 16, Zθ stage 17, and Zθ movement unit 18 described above. This allows the wafer W held in the wafer chuck 20 and the probe needle 35 described later to move relative to each other. That is, the Y stage 13 and Y movement unit 14, the X stage 15 and X movement unit 16, and the Zθ stage 17 and Zθ movement unit 18 function as relative movement units of the present invention.

[0034] Figure 3 is a top view of a wafer W held in a wafer chuck 20. As shown in Figure 3, multiple semiconductor chips 9 are formed on the wafer W. In addition, multiple electrode pads 9a are formed on each semiconductor chip 9.

[0035] Returning to Figures 1 and 2, the support column 23 is provided on the upper surface of the base 12 and supports the head stage 24 above the Y stage 13, X stage 15, and Zθ stage 17 (hereinafter simply referred to as stages 13, 15, and 17). This fixes the head stage 24 to the base 12 via the support column 23.

[0036] A card holder 25 is held in the center of the head stage 24. The card holder 25 has a holding hole 25a that holds the outer circumference of the probe card 26, and the probe card 26 is held in this holding hole 25a. As a result, the probe card 26 is held in a position facing the wafer W via the head stage 24 and the card holder 25.

[0037] The probe card 26 has probe needles 35 arranged according to the arrangement of electrode pads 9a on the semiconductor chip 9 to be tested. These card holders 25 and probe cards 26 are replaced according to the type of semiconductor chip 9.

[0038] The probe card 26 is provided with a connection terminal (not shown) electrically connected to the probe needle 35, to which a tester (not shown) is connected. The tester supplies various test signals to the electrode pads 9a of the semiconductor chip 9 via the connection terminal of the probe card 26 and the probe needle 35, and also receives and analyzes the signals output from the electrode pads 9a to test whether the semiconductor chip 9 is functioning correctly. The configuration of the tester and the test method are known technologies, so a detailed explanation is omitted.

[0039] The wafer alignment camera 29 corresponds to the first camera of the present invention and photographs the semiconductor chip 9 of the wafer W held in the wafer chuck 20. Based on the image captured by this wafer alignment camera 29, the position of the electrode pad 9a of the semiconductor chip 9 to be inspected can be detected. The installation position and structure of the wafer alignment camera 29 are not particularly limited, but in this embodiment, as disclosed in Japanese Patent Application Publication No. 2003-303865, an installation position and structure (spot light irradiation optical system) that can measure the relative distance between it and the needle alignment camera 31 described later is employed.

[0040] The upper and lower stages 30 are equipped with a needle alignment camera 31 and a cleaning plate 32 at positions approximately opposite to the head stage 24, etc. The upper and lower stages 30 also have a lifting mechanism (not shown) that is movable in the Z-axis direction, allowing adjustment of the Z-axis position of the needle alignment camera 31 and the cleaning plate 32. The needle alignment camera 31 and the cleaning plate 32 are supported so as to be movable in the XYZ axes via the Y-stage 13 and Y-movement unit 14, the X-stage 15 and X-movement unit 16, and the upper and lower stages 30. This allows relative movement between the needle alignment camera 31 and the cleaning plate 32 and the probe needle 35.

[0041] The needle alignment camera 31 corresponds to the second camera of the present invention and photographs the probe needle 35 of the probe card 26. Based on the image of the probe needle 35 captured by this needle alignment camera 31, the position of the probe needle 35 can be detected. Specifically, the XY coordinates of the tip position of the probe needle 35 are detected from the position coordinates of the needle alignment camera 31, and the Z coordinate of the tip position of the probe needle 35 is detected from the focal position of the needle alignment camera 31.

[0042] When inspecting semiconductor chips 9 on wafer W with the prober 10 configured as described above, each time the probe card 26 is replaced or each predetermined number of semiconductor chips 9 are inspected, the stages 13, 15, and 17 are driven to move the needle alignment camera 31 relative to the position where the probe needle 35 will be photographed, and then the probe needle 35 is photographed with the needle alignment camera 31. Based on the image captured by the needle alignment camera 31, the tip position of the probe needle 35 is detected as described above.

[0043] Furthermore, with the wafer W to be inspected held in the wafer chuck 20, each stage 13, 15, and 17 is driven to move the wafer alignment camera 29 relative to the shooting position of the wafer W, and then the semiconductor chip 9 of the wafer W is photographed with the wafer alignment camera 29. Based on the image taken by the wafer alignment camera 29, the position of the electrode pad 9a of the semiconductor chip 9 to be inspected is detected.

[0044] Then, stages 13, 15, and 17 are driven to electrically contact the probe needle 35 with the electrode pad 9a of the semiconductor chip 9 to be inspected first. In this state, the semiconductor chip 9 to be inspected first is inspected using a tester (not shown). The remaining semiconductor chips 9 to be inspected are then inspected in the same manner. Note that the specific inspection method for the semiconductor chips 9 is publicly known, so a detailed explanation is omitted here (see, for example, Patent Document 4).

[0045] The temperature sensors 34 are positioned opposite the lower surfaces of the card holder 25 and the probe card 26, for example, on the side of the Zθ stage 17 and the side of the upper and lower stages 30, respectively. Therefore, each temperature sensor 34 is held by the stages 13, 15, 17, and 30 so as to be movable relative to the card holder 25 and the probe card 26.

[0046] The temperature sensor 34 is a non-contact temperature sensor, for example, using a radiant energy detection method, and measures the temperature of the card holder 25 and the probe card 26 in a non-contact manner. The card holder 25 and the probe card 26 are thermally deformed by the temperature of the wafer chuck 20, and the tip position of the probe needle 35 is displaced as a result of this thermal deformation. Therefore, by measuring the temperature of the card holder 25 and the probe card 26 with the temperature sensor 34, the tip position [displacement (direction of displacement, amount of displacement)] of the probe needle 35 can be predicted (see Patent Document 4 above).

[0047] Figure 4 is an explanatory diagram showing an example of temperature measurement points on the card holder 25 and probe card 26 by the temperature sensor 34. Note that the probe needle 35 is not shown in Figure 4. As shown in Figure 4, the temperature sensor 34 measures the temperature at multiple locations on both the card holder 25 and the probe card 26, including multiple temperature measurement points P1 to P5 on the probe card 26 and multiple temperature measurement points P6 to P13 on the card holder 25, in order to detect the temperature distribution of both. Note that the temperature measurement points P1 to P13 in Figure 4 are examples, and their positions and numbers may be changed as appropriate.

[0048] The temperature sensor 34 measures the temperature at each temperature measurement point P1 to P13 under the control of the control unit 40 (see Figure 5), which will be described later, and outputs the temperature data, which is the result of the temperature measurement, to the control unit 40. When measuring the temperature at each temperature measurement point P1 to P13, the stages 13, 15, 17, and 30 are driven under the control of the control unit 40, which will be described later, so that the temperature sensor 34 is positioned to measure the temperature at each temperature measurement point P1 to P13. In other words, the temperature sensor 34 is moved relative to the card holder 25 and the probe card 26. This makes it possible to measure the temperature at each temperature measurement point P1 to P13 at a fixed point.

[0049] <Functions of the control unit> Figure 5 is a functional block diagram showing the functions of the control unit 40 of the prober 10. In Figure 5, only the functions related to contact control between the probe needle 35 and the wafer W (electrode pad 9a of the semiconductor chip 9) are shown among the functions of the control unit 40, and other functions are omitted from the illustration as they are known technology.

[0050] As shown in Figure 5, the control unit 40 corresponds to the prober control device of the present invention and provides overall control of each part of the prober 10. The control unit 40 may be built into the main body of the prober 10, or it may be provided separately from the main body.

[0051] The control unit 40 is composed of an arithmetic unit such as a personal computer, and includes an arithmetic circuit composed of various processors and memory. These various processors include CPUs (Central Processing Units), GPUs (Graphics Processing Units), ASICs (Application Specific Integrated Circuits), and programmable logic devices [e.g., SPLDs (Simple Programmable Logic Devices), CPLDs (Complex Programmable Logic Devices), and FPGAs (Field Programmable Gate Arrays)]. The various functions of the control unit 40 may be implemented by a single processor, or by multiple processors of the same or different types.

[0052] Furthermore, the control unit 40 is connected to the wafer alignment camera 29, needle alignment camera 31, and temperature sensor 34, as well as the alignment data measurement unit 38 and storage unit 39, via various communication interfaces (not shown).

[0053] The alignment data measurement unit 38 controls the wafer alignment camera 29 and the needle alignment camera 31, etc., to measure alignment data. The alignment data, along with the temperature data described above, is used to predict the tip position (displacement) of the probe needle 35. This alignment data includes, for example, the three-dimensional chip size of the semiconductor chip 9 to be inspected, the three-dimensional position of the wafer W, and the three-dimensional relative distance between the wafer alignment camera 29 and the needle alignment camera 31 (hereinafter abbreviated as camera relative distance). The camera relative distance indicates the positional relationship between the wafer alignment camera 29 and the needle alignment camera 31.

[0054] Specifically, the alignment data measurement unit 38 measures the chip size (expansion) of the semiconductor chip 9 based on the image of the wafer W (semiconductor chip 9) captured by the wafer alignment camera 29. The alignment data measurement unit 38 also measures the position of the wafer W based on the image of a specific pattern (not shown) of the semiconductor chip 9 captured by the wafer alignment camera 29. Furthermore, as disclosed in Japanese Patent Application Publication No. 2003-303865, the alignment data measurement unit 38 measures the camera relative distance using the wafer alignment camera 29, a needle alignment camera 31, and an optical system (not shown) that irradiates spot light. The alignment data measurement unit 38 then outputs alignment data, including the chip size of the semiconductor chip 9, the position of the wafer W, and the camera relative distance, to the control unit 40.

[0055] The memory unit 39 stores a control program (not shown in the diagram) for operating the control unit 40, the inspection results of the semiconductor chip 9 by the prober 10, and the training data 56 (also called training data) used for machine learning of the prediction model 47, which will be described later.

[0056] When inspecting a semiconductor chip 9 to be inspected within the wafer W, the control unit 40 executes a control program (not shown) read from the storage unit 39, thereby functioning as an input data acquisition unit 42, a needle position acquisition unit 44, a prediction unit 46, a prediction model generation unit 48, a determination unit 50, and a movement control unit 52.

[0057] The input data acquisition unit 42 performs temperature measurement at each temperature measurement point P1 to P13 using the temperature sensor 34 and measurement of alignment data using the alignment data measurement unit 38 before contact control is performed by bringing the probe needle 35 into contact with the semiconductor chip 9 to be inspected (hereinafter simply referred to as "before contact control"), and before the generation and retraining of the prediction model 47 described later. As a result, the input data acquisition unit 42 acquires input data including temperature data from each temperature measurement point P1 to P13 from the temperature sensor 34 and alignment data from the alignment data measurement unit 38 at each of the above timings.

[0058] Furthermore, the input data acquisition unit 42 outputs the input data acquired before contact control to the prediction unit 46 and the determination unit 50 described later, and outputs the input data acquired before the generation of the prediction model 47 and before retraining to the prediction model generation unit 48 described later.

[0059] The needle position acquisition unit 44, after the replacement of the probe card 26, after the inspection of a predetermined number of semiconductor chips 9, and before the generation and retraining of the prediction model 47 described later, performs photography of the probe needle 35 using the needle alignment camera 31 to acquire images of the probe needle 35 from the needle alignment camera 31, and acquires the tip position of the probe needle 35 based on these images.

[0060] Furthermore, the needle position acquisition unit 44 outputs the tip position of the probe needle 35 acquired before the generation and retraining of the prediction model 47 to the prediction model generation unit 48 (described later), and outputs the tip position of the probe needle 35 acquired after the replacement of the probe card 26, etc., to the movement control unit 52 (described later).

[0061] The prediction unit 46 predicts the tip position of the probe needle 35 before contact control is performed and when the decision unit 50 (described later) determines that the prediction by the prediction unit 46 is permissible. Specifically, the prediction unit 46 predicts the tip position of the probe needle 35 based on the input data acquired by the input data acquisition unit 42 (temperature data for each temperature measurement point P1 to P13, alignment data) and refers to the pre-generated prediction model 47 (described later), and outputs this prediction result to the movement control unit 52. The tip position of the probe needle 35 predicted by the prediction model 47 also includes the amount of variation (correction amount) from the tip position of the probe needle 35 acquired by the needle position acquisition unit 44.

[0062] The prediction model 47 is a pre-trained model generated by the prediction model generation unit 48 (described later) using machine learning (supervised learning) with a multiple regression model (also known as a multiple regression equation or multiple regression analysis). The prediction model 47 takes multiple input data (temperature data and alignment data for each temperature measurement point P1 to P13) as explanatory variables and outputs a predicted value for the tip position of the probe needle 35, which is the objective variable.

[0063] The prediction model generation unit 48 generates a prediction model 47 before inspecting the semiconductor chip 9 on the wafer W for the product. First, the prediction model generation unit 48 measures the training data 56 (input data, alignment data, and the tip position of the probe needle 35) using the wafer W for the product or an identical test wafer W (for prediction model creation).

[0064] Figure 6 is an explanatory diagram showing an example of training data 56 used for machine learning of the prediction model 47 by the prediction model generation unit 48. In order to avoid complexity in the diagram, Figure 6 only shows alignment data for one direction (in this case, the Y direction) and the tip position of the probe needle 35 in one direction.

[0065] As shown in Figure 6 and the previously described Figure 5, the prediction model generation unit 48 controls the temperature sensor 34, the alignment data measurement unit 38, and the movement control unit 52 (described later) to perform the measurement of input data (temperature measurement of each temperature measurement point P1 to P13, and measurement of alignment data) for a predetermined time. This provides temperature data T1 to T13 for each temperature measurement point P1 to P13. In addition, alignment data is obtained, including the chip size D1 of the semiconductor chip 9 (change from the start of the lot), the wafer position D2 of the wafer W (variation from the start of the lot), and the camera relative distance D3 (variation from the start of the lot).

[0066] The prediction model generation unit 48 controls the needle positioning camera 31, the needle position acquisition unit 44, and the movement control unit 52 (described later) in accordance with the measurement timing of the input data described above, to measure the tip position of the probe needle 35. This provides the tip position Y[α] (where α is an arbitrary natural number) of the probe needle 35 for each measurement timing of the input data.

[0067] In this way, the prediction model generation unit 48 acquires multiple training data 56, which include the input data and the tip position Y[α] of the probe needle 35 corresponding to the input data. As will be described in more detail later, in this embodiment, since the prediction model 47 can be retrained, it is sufficient to acquire only the minimum number of training data 56 necessary for machine learning the prediction model 47. For example, in this embodiment, training data 56 is acquired using one wafer W.

[0068] Next, the prediction model generation unit 48 generates a prediction model 47 for predicting the tip position of the probe needle 35 from the input data, based on multiple training data 56, namely the explanatory variables, which are the input data (T1~T13, D1~D3) and the target variable, which is the tip position Y[α] of the probe needle 35, using machine learning with a multiple regression model. The specific method for generating this prediction model 47, i.e., the machine learning algorithm using the multiple regression model, is a publicly known technique, so a detailed explanation is omitted here. As a result, the prediction unit 46 can predict the current tip position of the probe needle 35 from the current input data.

[0069] Furthermore, the machine learning algorithm used to generate the prediction model 47 is not limited to multiple regression models; for example, other known machine learning algorithms such as convolutional neural networks (CNNs) may also be used.

[0070] Furthermore, the prediction model generation unit 48 stores the training data 56 (or input data only) used for machine learning of the prediction model 47 in the storage unit 39. The training data 56 stored in the storage unit 39 is used by the decision unit 50, described later, to determine whether or not to perform the prediction by the prediction unit 46.

[0071] Furthermore, the prediction model generation unit 48, as will be described in more detail later, operates after the prediction model 47 has been generated and when the decision unit 50, described later, decides "no" to perform the prediction by the prediction unit 46, and retrains the prediction model 47.

[0072] Returning to Figure 5, the decision unit 50 operates before contact control (before prediction by the prediction unit 46) and compares the current (latest) input data acquired by the input data acquisition unit 42 with the input data of the training data 56 in the storage unit 39 to determine whether or not to perform prediction by the prediction unit 46 (simply referred to as prediction feasibility determination).

[0073] If the current input data is not significantly different from the input data of the training data 56 used in the machine learning of the prediction model 47, then the prediction model 47 is already trained, having completed machine learning on the training data 56 corresponding to the current input data. Therefore, when the prediction unit 46 predicts the tip position of the probe needle 35 using the prediction model 47 based on the current input data, it can accurately predict the tip position of the probe needle 35.

[0074] On the other hand, if the current input data deviates significantly from the input data of the training data 56, the prediction model 47 is in an untrained state, having not undergone machine learning with the training data 56 corresponding to the current input data. Therefore, even if the prediction unit 46 attempts to predict the tip position of the probe needle 35 using this untrained prediction model 47, it cannot accurately predict the tip position of the probe needle 35.

[0075] Therefore, the decision unit 50 compares the current input data with the input data of the training data 56 in the storage unit 39 to determine whether the prediction model 47 has been trained or not for the current input data, thereby determining whether prediction is possible.

[0076] First, the determination unit 50 calculates the difference between the current input data and the input data of the training data 56 for each parameter of the input data (temperature data T1 to T13, chip size D1, wafer position D2, camera relative distance D3). Next, the determination unit 50 calculates the square root of the sum of squares of the differences for each parameter and determines whether the prediction model 47 is in a trained state or an untrained state based on whether there is at least one parameter whose square root of the sum of squares falls within a certain range (below a threshold).

[0077] The determination method (judgment method) by the determination unit 50 will be explained in detail below. To avoid complicating the explanation, it will be assumed here that the input data consists only of temperature data T1 to T13.

[0078] Given that the number of input data parameters is m and the number of training iterations is N, the trained input data (explanatory variables) is expressed by the following equation [Equation 1]. The trained tip position Y[α] of the probe needle 35 (dependent variable) is expressed by the following equation [Equation 2]. The function obtained by the machine learning algorithm using a multiple regression model, i.e., the predictive model 47, is expressed by the following equation [Equation 3].

[0079]

number

[0080]

number

[0081]

number

[0082] When the "current input data" obtained during the prediction of the tip position of the probe needle 35 is represented by X[T] as shown in equation [Equation 4] below, the Euclidean distance D[s] between this current input data X[T] and the input data X[s] of any sth training data 56 is expressed by equation [Equation 5] below.

[0083]

number

[0084]

number

[0085] The determination unit 50 determines a predetermined threshold D for all Euclidean distances D[s] (s=1, 2, ..., N). th A comparison is made with the threshold D th If there is at least one Euclidean distance D[s] that is less than [equation 3], it is determined that the prediction model 47 (equation 3) is in a trained state, and it is decided to perform a prediction using this prediction model 47. In this case, the prediction unit 46 described above predicts the target variable, the tip position Y[T] of the probe needle 35, using the prediction model 47 shown in equation 3, based on the explanatory variable, the current input data X[T].

[0086] On the other hand, the determination unit 50 determines the threshold D th If there are no Euclidean distances D[s] less than the given value, it is determined that the prediction model 47 is in an untrained state, and the prediction by this prediction model 47 is not performed. In this case, the prediction model generation unit 48 receives the decision result input from the decision unit 50 and functions as the retraining unit of the present invention to retrain the prediction model 47.

[0087] When the prediction model generation unit 48 retrains the prediction model 47, it controls the needle position acquisition unit 44 and the movement control unit 52 (described later) to acquire the tip position Y[T] of the probe needle 35, which is the target variable corresponding to the current input data X[T]. At this time, the prediction model generation unit 48 may also control the temperature sensor 34, alignment data measurement unit 38, and movement control unit 52 to remeasure the input data.

[0088] Next, the prediction model generation unit 48 adds the current input data X[T] and the tip position Y[T] of the probe needle 35 to the training data 56 in the storage unit 39 to create new training data 56. At this time, in order to reduce the effect of drift of the temperature sensor 34, it is preferable for the prediction model generation unit 48 to exclude the oldest data (X[1], Y[1]) from the training data 56 (explanatory variable, target variable) in the storage unit 39, as shown in equations [Equation 6] and [Equation 7] below.

[0089]

number

[0090]

number

[0091] Then, the prediction model generation unit 48 performs machine learning using a multiple regression model based on the training data 56 stored in the memory unit 39, that is, the input data (explanatory variables) shown in equation [Equation 6] above and the tip position of the probe needle 35 (dependent variable) shown in equation [Equation 7] above, to retrain the prediction model 47. As a result, a new prediction model 47 (function) is obtained as shown in equation [Equation 8] below.

[0092]

number

[0093] Once the retraining of the prediction model 47 is complete, the input data acquisition unit 42 acquires the input data, and then the decision unit 50 determines whether or not a prediction is possible.

[0094] From here on, until the decision unit 50 determines that prediction is possible, the needle position acquisition unit 44, the prediction model generation unit 48, the input data acquisition unit 42, and the decision unit 50 operate repeatedly, thereby repeatedly acquiring the tip position of the probe needle 35, updating the training data 56 in the storage unit 39, retraining the prediction model 47, acquiring input data, and determining whether prediction is possible. As a result, the prediction unit 46 can always predict the tip position of the probe needle 35 using the trained prediction model 47.

[0095] The movement control unit 52 drives each of the stages 13, 15, and 17 via the Y movement unit 14, the X movement unit 16, and the Zθ movement unit 18. Based on the captured image input from the wafer alignment camera 29, the movement control unit 52 acquires the position of the semiconductor chip 9 (electrode pad 9a) to be inspected on the wafer W held in the wafer chuck 20. The movement control unit 52 also acquires the tip position of the probe needle 35 (a value measured when the probe card 26 is replaced, etc.) from the needle position acquisition unit 44.

[0096] Then, when inspecting a wafer W for a product, the movement control unit 52 drives each stage 13, 15, and 17 to move the wafer W relative to the probe needle 35, thereby bringing the probe needle 35 into sequential contact with the semiconductor chips 9 on the wafer W that are to be inspected. At this time, based on the prediction result of the prediction unit 46 for the tip position of the probe needle 35, the movement control unit 52 drives each stage 13, 15, and 17 to correct the contact position of the probe needle 35 with respect to each semiconductor chip 9 that is to be inspected. As a result, even if the tip position of the probe needle 35 is displaced due to thermal deformation of the card holder 25 and probe card 26, the probe needle 35 can be brought into contact with the semiconductor chips 9 that are to be inspected at each corrected contact position corresponding to the displaced tip position.

[0097] [Operation of this embodiment] Figure 7 is a flowchart illustrating the flow of the method by which the probe needle 35 contacts the semiconductor chip 9 using the probe 10 configured as described above, which corresponds to the probe control method of the present invention. It is assumed that a prediction model 47 has been generated in advance, the training data 56 used for its machine learning is stored in the storage unit 39, and the tip position of the probe needle 35 is also acquired by the needle position acquisition unit 44.

[0098] When a wafer W for a product is held in the wafer chuck 20, the wafer alignment camera 29 takes an image of the semiconductor chip 9 on the wafer W. Then, the movement control unit 52 determines the position of the semiconductor chip 9 (electrode pad 9a) to be inspected based on the image taken by the wafer alignment camera 29.

[0099] Furthermore, the input data acquisition unit 42 causes the temperature sensor 34 to measure the temperature of each temperature measurement point P1 to P13 and the alignment data measurement unit 38 to measure alignment data. As a result, the input data acquisition unit 42 acquires the current input data, which includes the temperature data of each temperature measurement point P1 to P13 and the alignment data (step S1, corresponding to the input data acquisition step of the present invention).

[0100] Once the acquisition of the current input data is complete, the determination unit 50 operates and compares the current input data acquired by the input data acquisition unit 42 with the input data of the training data 56 in the storage unit 39 to determine whether prediction is possible (step S2, corresponding to the determination step of the present invention). Specifically, the determination unit 50 compares each of the Euclidean distances D[s] (s=1, 2, ..., N) represented by the above [Equation 5] with the threshold D th Based on the results of this comparison, the predictive model 47 determines whether the current input data has been learned or not.

[0101] If the determination unit 50 determines that prediction is not possible (NO in step S3), the prediction model generation unit 48 controls the needle position acquisition unit 44 and the movement control unit 52, etc., to acquire the tip position of the probe needle 35 corresponding to the current input data (step S4). At this time, the input data may be reacquired.

[0102] Next, the prediction model generation unit 48 updates the training data 56 in the storage unit 39 by adding the current input data and the tip position Y of the probe needle 35 to the training data 56, as shown in equations [6] and [7] above, and excluding the oldest data (step S5). Then, the prediction model generation unit 48 retrains the prediction model 47 based on the new training data 56 in the storage unit 39 to generate a new prediction model 47 (step S6).

[0103] Once the retraining of the prediction model 47 is complete, the input data acquisition unit 42 acquires the input data again (step S2), and the decision unit 50 makes a decision on whether or not to make a prediction based on this input data (step S3). The processes of steps S4 to S6, step S1, and step S2 are repeated until the decision unit 50 determines that the prediction is possible.

[0104] If the determination unit 50 determines that prediction is possible (YES in step S3), the prediction unit 46 predicts the tip position of the probe needle 35 by referring to the prediction model 47 based on the input data acquired in the most recent step S1 (step S7, corresponding to the prediction step of the present invention). The prediction unit 46 then outputs the prediction result of the tip position of the probe needle 35 to the movement control unit 52.

[0105] Next, the movement control unit 52 controls the movement of each stage 13, 15, and 17 based on the predicted position of the tip of the probe needle 35 input from the prediction unit 46 and the previously determined position of the semiconductor chip 9 to be inspected, so that the probe needle 35 comes into contact with the semiconductor chip 9 to be inspected (step S8). After this contact, the semiconductor chip 9 is inspected by a tester (not shown) (step S9).

[0106] The same inspection process is then carried out for the remaining semiconductor chips 9 to be inspected. In this process, steps S1 to S7 may be repeated each time a predetermined number of semiconductor chips 9 are inspected or after a predetermined amount of time has elapsed.

[0107] [Effects of this embodiment] As described above, in this embodiment, the decision unit 50 compares the current input data acquired before contact control with the input data of the training data 56 to determine whether prediction is possible or not. If the decision unit 50 determines that prediction is not possible, it retrains the prediction model 47, thereby enabling more accurate prediction of the tip position of the probe needle 35 than in the conventional method.

[0108] Figure 8 shows a graph (see symbol VIIIA) showing the predicted value PV and the measured value MV of the tip position of the probe needle 35, and a graph (see symbol VIIIB) showing the difference between the predicted value PV and the measured value MV of the tip position of the probe needle 35, in a comparative example where the determination unit 50 does not perform determination and retraining of the prediction model 47. Figure 9 shows a graph (see symbol IXA) showing the predicted value PV and the measured value MV of the tip position of the probe needle 35, and a graph (see symbol IXB) showing the difference between the predicted value PV and the measured value MV of the tip position of the probe needle 35, in this embodiment.

[0109] The graphs in Figures 8 and 9 show the time evolution of the predicted value PV and the measured value MV of the tip position of the probe needle 35 in any one direction (in this case, the Y direction) within the XYZ directions, as well as the time evolution of the difference between them, when the temperature of the wafer chuck 20 is set to 200°. In addition, the symbol WA in Figures 8 and 9 indicates the machine learning range in which machine learning was performed.

[0110] As shown in Figure 8, in the comparative example, drift occurred in the temperature sensor 34, or displacement of the tip position of the probe needle 35 occurred due to temperature fluctuations in areas within the prober 10 where temperature measurement was not performed. As a result, a discrepancy occurred between the predicted value PV and the measured value MV of the tip position of the probe needle 35 beyond the machine learning range WA, and it was confirmed that the difference value gradually increased.

[0111] In contrast, as shown in Figure 9, in this embodiment, even after the machine learning range WA, the determination unit 50 determines whether prediction is possible and the prediction model 47 is retrained, resulting in a near-perfect match between the predicted value PV and the measured value MV of the probe needle tip position, and a reduction in the difference value. As a result, in this embodiment, even if drift occurs in the temperature sensor 34 or if the tip position of the probe needle 35 is displaced due to temperature fluctuations in areas within the prober 10 where temperature measurement is not performed, the prediction model 47 is retrained, allowing for a more accurate prediction of the probe needle tip position.

[0112] Furthermore, in this embodiment, by enabling retraining of the prediction model 47, it becomes unnecessary to continue the initial generation (machine learning) of the prediction model 47 for a long time, thus reducing the workload required to generate the prediction model 47.

[0113] [others] In the above embodiment, the prediction model generation unit 48 performs both the generation and retraining of the prediction model 47. However, the generation of the prediction model 47 may be performed by the manufacturer of the prober 10 or another prober 10. In this case, a retraining unit that performs only the retraining of the prediction model 47 may be provided in the control unit 40 instead of the prediction model generation unit 48.

[0114] In the above embodiment, the determination unit 50 performs the prediction feasibility determination using the above formula [Equation 5], but this determination method is not particularly limited. For example, for each parameter of the input data, it may be determined whether the current input data is included between the maximum and minimum values ​​of the training data 56 input data (hereinafter referred to as the maximum / minimum range), and the prediction feasibility determination may be performed based on whether the current input data is included in the maximum / minimum range for all parameters.

[0115] In the above embodiment, temperature data of the card holder 25 and probe card 26, and alignment data are measured as input data, but only temperature data may be measured. Also, in the above embodiment, temperature data of both the card holder 25 and probe card 26 are measured as input data, but the temperature data of at least one of the card holder 25 and probe card 26 may be measured. Furthermore, in the above embodiment, the chip size of the semiconductor chip 9, the position of the wafer W, and the camera relative distance are measured as alignment data, but at least one of these may be measured.

[0116] In the above embodiment, a non-contact temperature sensor 34 is used, but a contact temperature sensor 34 may also be used. [Explanation of symbols]

[0117] 9...Semiconductor chip, 9a...Electrode pad, 10...Probe, 12...Base, 13...Y stage, 14...Y movement unit, 15...X stage, 16...X movement unit, 17...Zθ stage, 18...Zθ movement unit, 20...Wafer chuck, 20a...Temperature control unit, 23...Support column, 24...Head stage, 25...Card holder, 25a...Holding hole, 26...Probe card, 29...Wafer alignment camera, 30...Upper and lower stages, 31...Needle alignment camera, 32...Cleaning plate, 34...Temperature sensor, 35...P Lobe needle, 38... Alignment data measurement unit, 39... Storage unit, 40... Control unit, 42... Input data acquisition unit, 44... Needle position acquisition unit, 46... Prediction unit, 47... Prediction model, 48... Prediction model generation unit, 50... Decision unit, 52... Movement control unit, 56... Training data, D... Euclidean distance, D1... Chip size, D2... Wafer position, D3... Camera relative distance, Dth... Threshold, MV... Measured value, P1~P13... Temperature measurement points, PV... Predicted value, T1~T13... Temperature data, W... Wafer, WA... Machine learning range

Claims

1. In a prober control device that drives a relative movement unit to move a probe card having a probe needle and a card holder that faces the object to be inspected for electrical characteristics, the probe needle comes into contact with the object to be inspected, An input data acquisition unit that acquires input data including temperature data from at least one of the probe card and the card holder, Based on the input data acquired by the input data acquisition unit, a prediction unit predicts the tip position of the probe needle using a prediction model that takes the input data as input and outputs the tip position of the probe needle. Before the prediction unit makes a prediction, a decision unit determines whether or not to perform a prediction by the prediction unit based on the input data used as training data for the machine learning of the prediction model and the input data acquired by the input data acquisition unit. A probe control device equipped with [a specific feature].

2. In a prober control method, a probe card having a probe needle and a card holder that faces the probe card toward an object to be inspected for electrical characteristics are driven by a relative movement unit to move the probe needle toward the object to be inspected, An input data acquisition step of acquiring input data including temperature data of at least one of the probe card and the card holder, A prediction step in which, based on the input data acquired in the input data acquisition step, the tip position of the probe needle is predicted using a prediction model that takes the input data as input and the tip position of the probe needle as output, Prior to the prediction step, a decision step is performed to determine whether or not to perform the prediction step based on the input data used as training data for the machine learning of the prediction model and the input data acquired in the input data acquisition step. A probe control method having the following.

Images