ASSEMBLY AND METHOD FOR PERFORMING IN-SITU ENDPOINT DETECTION WHEN BACKSIDE MILLING SILICON BASED DEVICES

Abstract

An assembly for monitoring a semiconductor device under test comprising a mill configured to mill the device, a sensor configured to measure an electrical characteristic of the device, and a computer configured to determine the amount of strain in the device from the electrical characteristic when the mill is milling the device and detect an endpoint of milling at a circuit within the device. In use the endpoints of the milling process of the semiconductor device are detected measuring an electrical characteristic of the device with a sensor during milling determining the amount of strain in the device from the electrical characteristic and detecting an endpoint of the milling process within the device based on the amount of strain.

Claims

1-20. (canceled)

21. An assembly for monitoring a semiconductor device during milling comprising: a micromill configured to mill the semiconductor device, a signal generator configured to send a signal to the semiconductor device; a sensor configured to sense a power draw signal from the semiconductor device, and a computer configured to: segment the power draw signal into a feature vector; determine one or more second order effects of the power draw signal from the feature vector; determine a strain on the semiconductor device based on the one or more second order effects using one or more machine learning algorithms; and detect a milling endpoint when the strain on the semiconductor device reaches a predetermined threshold.

22. The assembly of claim 21, wherein the feature vector is a set of discrete values that represent the power draw signal.

23. The assembly of claim 22, wherein the feature vector is transformed into a frequency or a time independent domain.

24. The assembly of claim 23, wherein the feature vector is transformed into the frequency using at least one of a discrete Fourier transform, a fast Fourier transform, a cosine transform, a Hilbert transform, a real cepstrum, a wavelet coefficients, or combinations thereof.

25. The assembly of claim 22, wherein the feature vector is transformed to reduce dimensionality on the feature vector.

26. The assembly of claim 21, wherein the computer is further configured to stop milling the semiconductor device when the milling endpoint has been detected.

27. The assembly of claim 21, wherein a value of the predetermined threshold is selected to avoid irreversible damage to the semiconductor device.

28. The assembly of claim 21, further comprising: a socket, the socket configured to: immobilize the semiconductor device with respect to the micromill; and allow access to a backside of the semiconductor device for milling.

29. The assembly of claim 28, wherein the signal generator is connected to the socket when the socket is placed in the micromill.

30. The assembly of claim 21, wherein the sensor is a current or a voltage sensor.

31. An assembly for monitoring a semiconductor device during milling of the semiconductor device, the assembly comprising: a sensor configured to sense a power draw of the semiconductor device during the milling in a mill and provide a power draw signal representative of an electrical characteristic; and a computer configured to: segment the power draw signal into a feature vector; determine one or more second order effects of the power draw signal from the feature vector; determine a strain on the semiconductor device based on the one or more second order effects using one or more machine learning algorithms; and stop the milling before a circuit within the semiconductor device is damaged based on the strain on the semiconductor device.

32. A method of milling a semiconductor device, the method comprising: receiving a power draw signal from a sensor configured to monitor a power draw of the semiconductor device; segmenting the power draw signal into a feature vector; determining one or more second order effects of the power draw signal from the feature vector; determining a strain on the semiconductor device based on the one or more second order effects using one or more machine learning algorithms; and detecting a milling endpoint when the strain on the semiconductor device reaches a predetermined threshold.

33. The method of claim 32, wherein the feature vector is a set of discrete values that represent the power draw signal.

34. The method of claim 33, further comprising: transforming the feature vector into a frequency or a time independent domain.

35. The method of claim 34, wherein the feature vector is transformed into the frequency using at least one of a discrete Fourier transform, a fast Fourier transform, a cosine transform, a Hilbert transform, a real cepstrum, a wavelet coefficients, or combinations thereof.

36. The method of claim 33, wherein the feature vector is transformed to reduce dimensionality on the feature vector.

37. The method of claim 32, further comprising: stopping the milling of the semiconductor device when the milling endpoint has been detected.

38. The method of claim 32, wherein a value of the predetermined threshold is selected to avoid irreversible damage to the semiconductor device.

39. The method according to claim 32 further comprising: sending a signal to the semiconductor device to generate the power draw signal.

40. The method according to claim 39 wherein the signal is sent to the semiconductor device by a signal generator.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

[0013] The disclosure may be more completely understood in consideration of the following description of various illustrative embodiments in connection with the accompanying drawings.

[0014] FIG. 1A shows a schematic view of an overall assembly for milling semiconductor devices in accordance with a preferred embodiment of the invention. The assembly is shown with a signal generator, a micromill with a socket, an oscilloscope and a data analysis system with a trigger connection to the micromill.

[0015] FIG. 1B shows a schematic view of an assembly similar to the assembly of FIG. 1A wherein the oscilloscope has been replaced with a sensor in accordance with another preferred embodiment of the invention.

[0016] FIG. 1C shows a schematic view of an assembly similar to the assembly of FIG. 1A wherein the signal generator, the oscilloscope and the data analysis system have been combined into one unit.

[0017] FIG. 2 shows a flow of information between several components of the data analytics device.

[0018] FIG. 3 shows a micromill holding a semiconductor device.

[0019] FIG. 4 shows a cross section of a semiconductor device.

[0020] FIGS. 5A, 5B and 5C show a semiconductor device responding to different levels of applied force.

[0021] FIG. 6 is a flow chart of a milling process.

[0022] FIG. 7 is a graph showing how a semiconductor responds to different amounts of pressure.

[0023] FIG. 8 is a prior art graph showing light absorption depths of silicon at various wavelengths of light.

[0024] FIG. 9 is a prior art model of a load versus strain at various thicknesses of a semiconductor device.

DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0025] The following detailed description should be read with reference to the drawings in which similar elements in different drawings are numbered the same. The detailed description and the drawings, which are not necessarily to scale, depict illustrative embodiments and are not intended to limit the scope of the disclosure. The illustrative embodiments depicted are intended only as exemplary. Selected features of any illustrative embodiment may be incorporated into an additional embodiment unless clearly stated to the contrary. While the disclosure is amenable to various modifications and alternative forms, specifics thereof have been shown by way of example in the drawings and will be described in detail. It should be understood, however, that the intention is not to limit aspects of the disclosure to the particular illustrative embodiments described. On the contrary, the intention is to cover all modifications, equivalents, and alternatives falling within the spirit and scope of the disclosure.

[0026] As used in this specification and the appended claims, the singular forms a, an and the include plural referents unless the content clearly dictates otherwise. As used in this specification and the appended claims, the term or is generally employed in its sense including and/or unless the content clearly dictates otherwise.

[0027] In the description of embodiments disclosed herein, any reference to direction or orientation is merely intended for convenience of description and is not intended in any way to limit the scope of the present invention. Relative terms such as lower, upper, horizontal, vertical,, above, below, up, down, top and bottom as well as derivative thereof (e.g., horizontally, downwardly, upwardly, etc.) should be construed to refer to the orientation as then described or as shown in the drawing under discussion. These relative terms are for convenience of description only and do not require that the apparatus be constructed or operated in a particular orientation. Terms such as attached, affixed, connected, coupled, interconnected, and similar refer to a relationship wherein structures are secured or attached to one another either directly or indirectly through intervening structures, as well as both movable or rigid attachments or relationships, unless expressly described otherwise.

[0028] As used throughout, any ranges disclosed herein are used as shorthand for describing each and every value that is within the range. Any value within the range can be selected as the terminus of the range.

[0029] As best seen in FIG. 1A, an assembly 100 is comprised of a signal generator 110, a socket 120 configured to receive a semiconductor device 125, which may include an integrated circuit, a mill or micromill 130, an electrical signal recording device such as an oscilloscope 140 and a data analysis computer 150. Preferably micromill 130 includes a pressure-based grinding system. Semiconductor device 125 is connected to oscilloscope 140 and signal generator 110 via socket 120. More specifically, a first communication line 160 is provided from signal generator 110 to socket 120 and a second communication line 170 is provided from oscilloscope 140 to socket 120. Oscilloscope 140 is also connected to data analysis computer 150 through a third communication line 180. Data analysis computer 150 can interface via a trigger, represented by line 190 with micromill 130.

[0030] Socket 120 may be any socket which allows for temporary placement and electrical connection to signal generator 110 and oscilloscope 140. Preferably, socket 120 immobilizes semiconductor device 125 with respect to micromill 130 and prevents contamination of electrical contacts by milling fluid. Socket 120 may be semi-permanent via solder or conductive attachment. Socket 120 may allow back side access or front side access to semiconductor device 125. In some embodiments semiconductor device 125 is soldered or otherwise permanently wired to lead wires represented by communication line 170 which are connected to oscilloscope 140 and signal generator 110. In these cases, socket 120 is not needed. In some embodiments, socket 120 will be replaced by a printed circuit board (not shown) to which semiconductor device 125 is soldered. Preferably, the printed circuit board is coated with a material after soldering to provide chemical and electrical passivation. Parylene, PDMS, or another common encapsulates are used for this purpose.

[0031] Signal generator 110 may be any clock-generating device that can be used to apply a clock or power waveform 200 to semiconductor device 125 either directly or through a specified socket 120. Signal generator 110 could be as simple as a single board microcontroller (i.e. Arduino), FPGA, development board, single board computer, or complex as a professional-grade signal generator. One or more signals 200 are applied to semiconductor device 125. In a most preferred embodiment, a single signal is sent to a clock of semiconductor device 125 causing power draw signal 210 which is measured by oscilloscope 140.

[0032] Oscilloscope 140 may be replaced with any electrical signal recording device such as a side-channel or secondary-effects measurement device. As milling occurs, semiconductor device 125 will continue to produce undisturbed primary effects; indeed, if primary effects are disturbed, milling has caused device failure and sample semiconductor device 125 is no longer usable. As shown in FIG. 1B, oscilloscope 140 has been replaced by a box 240 representing a number of secondary effect measurement devices which may be used instead of oscilloscope 140. For example, a current sensor whose output is digitized via a high speed analog to digital converter could be used. Alternatively, a voltage sensor could be used in place of a current sensor. A thermal sensor could also be employed. Finally, an RF antenna and signal chain could be used to look for variations in RF energy related to strain on the die in semiconductor device 125. Data analysis computer 150 may be any PC or miniature computer; standalone signal analyzer; or signal processing unit built-in to the oscilloscope or any other signal measurement device.

[0033] In some embodiments, as represented by box 350 shown in FIG. 1C, the functions provided by signal generator 110, oscilloscope 140, and data analysis computer 150 may be implemented in the same controller 350. Controller 350 may take many forms including a field-programmable gate array, a microprocessor development board including a printed circuit board containing a microprocessor and minimal support logic, a personal computer, or a controller integrated within the micromill electronics. Further details of these components are found in U.S. Pat. No. 10,054,624, incorporated herein by reference. Regardless of whether the computer is one separate unit or combined with other units, the overall arrangement will preferably follow the logic shown in FIG. 2 Namely a sensor 410 measures power draw from semiconductor device 125. An analog conditioner 420 receives signal 200 as an analog signal power versus time. Digitizer 430 and digital conditioner 440 segments signal 200 to produce a digital vector representing the signal which is then processed by various machine learning algorithms 450 to determine the second order effects on the power draw and how the second order effects relate to strain in semiconductor device 125. Algorithms 450 can then determine when to send trigger signal 180 to micromill 130 to prevent damage to semiconductor device 125. Dependent on the desired impact, trigger 190 will act differently on micromill 130. In some configurations, trigger 190 will increase or reduce applied pressure, increase or decrease milling bit rotational speed, increase or decrease feed velocity in the horizontal plane, increase or decrease feed velocity in the vertical dimension, increase or decrease stage angle, increase or decrease spindle/collet/milling bit angle, start or stop milling, turn on and off the mill or any combination of the above specified actions.

[0034] FIG. 3 is a schematic representation of the micromill 130 which includes a tilt table 514 for supporting and oscillating supported semiconductor device 125 in the X and Y directions. Two drive motors are disposed in a base element 520 serve to oscillate table 514 as is well known in the art. Both the speed and the amplitude of the oscillations in the X and Y directions are independently adjustable via controller 522. A tool 524 is rotated at an adjustable speed by an element 528 which is movable in the Z direction. The speed of rotation is adjustable via input from controller 522. A precision Z motor 530 controls the feed velocity normal to the milling surface. Micromill 130 uses a pressure sensor 540 to determine force which is applying to spinning tool 524. This method of contact is known as end milling, as the end of the milling tool 524 removes the material. This pressure ranges from 0-1000 g of force. A similar mill is described in U.S. Pat. No. 6,620,369, incorporated herein by reference. Alternative embodiments of the micromill include additional milling axes, an edge mill (as opposed to an end mill), a mill employing Laser Milling, Laser Assisted Chemical Etching, or Focused or broad ion beam milling.

[0035] FIG. 4 shows a cross section of semiconductor device 125 with a portion of epoxy layer 610 completely removed. Also, a portion of silicon layer 620 is removed leaving a remaining thickness 630 between a milling surface 640 and a die or circuit 650. Circuit 650 is connected to a power contact or pin 660 and a ground contact or pin 670 by conductors 680 and 690 respectively.

[0036] Assembly 100 employs the inherent nature of the semiconductor materials to detect damaging strain in a semiconductor device 700 that includes an integrated circuit 705. As shown in FIGS. 5A-5C there is shown circuit of a semiconductor device 700 under the application of a force shown by an arrow 710. In FIG. 5A no force is applied and an input signal to semiconductor device 700 is faithfully produced as an output signal 720. In FIG. 5B as force 710 is applied the behavior of circuit 705 begins to change. If the pressure is reduced in response to increasing strain semiconductor device 700 will return to normal operation shown in FIG. 5A. If the pressure is not reduced in response to increasing strain, thinning of the silicon by milling will result in device 700 breaking as seen in FIG. 5C. The large deformation cause collisions between metal layers in circuit 705, resulting in shorts and therefore large increases in current draw.

[0037] As shown in FIG. 6, a method 800 of in-situ monitoring of a principle component analysis of the power draw of semiconductor device 125 is employed to detect a change in strain. Method 800 is employed to remove backside silicon layer 620 in FIG. 4 from semiconductor device 125 to exposed integrated circuit 650, thus allowing circuit 650 to be analyzed for faults using the optical techniques described above.

[0038] Initially, at step 810, semiconductor device 125 mounted on table of micromill 130, as shown in FIG. 3. Semiconductor device 125 may be, for example a PIC16 microcontroller manufactured by the Microchip company having at least a power pin connection 660 and a ground pin connection 670. Next, signal generator 110 and oscilloscope 140 are connected to semiconductor device 125. Preferably, signal generator 110 is connected to power pin connection 660 and oscilloscope 140 is connected to ground pin 670. Next micromill 130 is used to remove protective epoxy layer 610 on semiconductor device 125.

[0039] At step 820, micromill 130 is controlled to mill backside silicon layer 620 of semiconductor 125 at the same time, signal generator 110 is adjusted to send a signal to semiconductor device 125. Preferably, a block signal 200 is sent but various different types periodic signal may be employed. Oscilloscope 140 measures a response signal from semiconductor device 125, caused by block signal 200 at ground pin 670. While silicon layer 620 of semiconductor device is being milled, the response signal is continuously measured. Also, a force is applied to semiconductor device 125. As micromill 130 mills through silicon layer 620 and gets closer to the actual circuit or die 650 within semiconductor device 125, the strain on die 650 increases and causes a change in power draw signal 210. The force can come from the milling performed by micromill 130 or be generated internally within semiconductor device 125. The force within die 650 is also caused by the die's heterogeneous structure having two disparate materials bonded together which create stress when cyclically loaded and that in turn creates strain. As the milling process removes more and more silicon, semiconductor device 125 becomes more flexible allowing it to strain more. While not wishing to be bound by theory, it is believed that the change in power draw signal 210 is caused by the change in carrier density in the actual semiconductor junctions that change as a function of strain. The power draw is correlated to the strain, under load, of the integrated circuit. Assembly 100 is used to monitor the in-situ strain signals through power analysis while semiconductor device 125 is undergoing backside milling. A typical integrated circuit 650 in semiconductor device 125 contains many field effect transistors. It is theorized that this measurable change in the power draw, or more precisely the measurable second order effect on the power draw is due to the change in electron mobility of the field effect transistors when strained. Regardless, once micromill 130 reaches the endpoint of the milling process when micromill 130 has removed enough of the silicon covering the die, the milling is stopped. Typically the mill is stopped before the silicon is completely removed otherwise the device will not function.

[0040] Next at step 830 a power analysis is performed using machine learning techniques to determine the normal power signature of semiconductor device 125 under minimal strain. Preferably, power draw signal 210 is measured over time when semiconductor device 125 subject to clock signal 200, although numerous other values such as voltage over time or current over time could be measured. Power draw signal 210 may be segmented to convert a measured analog power signal into a set of discrete values that represent the power signal. The segmented signal may be referred to as a feature vector. The feature vector can be transformed into the frequency or a different time independent domain. In one embodiment, each feature vector is transformed with a discrete fourier transform or fast fourier transform. In alternative embodiments, each feature vector may be transformed with a discrete cosine transform, Hilbert transform, real cepstrum, wavelet coefficients, or a hybrid of several different transforms. The dimension of the feature vector can be reduced using essentially any known dimension reduction technique. Preferably, principal component analysis is conducted to reduce dimensionality on the feature vector. Principal component analysis transforms the feature vectors into a space where the greatest variance between samples is in the first dimension, the next greatest variance in the next dimension and so on. By organizing the feature vectors by greatest variance, dimensions where the least variance between samples occurs can be discarded in order to enable comparisons in a lower dimensional space with conventional distance metrics. Although the current embodiment implements principal component analysis, other non-linear analysis techniques may be employed instead such as self organizing maps or other manifold based learning algorithms. In one embodiment, principal component analysis on the feature vector to reduce dimensionality of the feature vector includes organizing the feature vector by variance and discarding dimensions where the variance is below a threshold. In another embodiment, principal component analysis on the feature vector to reduce dimensionality of the feature vector includes organizing the feature vector by variance and discarding all but a predefined number of dimensions that have the highest variance. In addition, a clustering analysis may be conducted of the vectors to increase accuracy of determining the strain in the semiconductor device. Further details of these machine learning techniques are found in U.S. Pat. No. 10,054,624 and U.S. Patent Application Publication No. 2018/0307654, both incorporated herein by reference.

[0041] At step 840 the device is monitored while under load. Strain is a function of the cross-sectional profile of homogeneous material. As the thickness goes to zero the strain increases as t.sup.3, where t is the thickness and all other factors being equal. By monitoring the change for a sharp increase in the electrical characteristics as a function of strain, the endpoint is detected and the load on the semiconductor device can be reduced when the milling endpoint is detected to avoid irreversible damage. The triggers sent from the computer can thereby be controlled to stop milling when the milling endpoint has been detected. This technique has been demonstrated using a PIC16 microcontroller. In this demonstration, shown in FIG. 7, the device was damaged do to overloading (800 mg) but this is significantly higher than standard milling profiles. As shown in FIG. 7 the strain is reversible after loading. Therefore, the strain can be detected at lower loading values and the milling stopped before irreversible strain is reached.

[0042] Having thus described several illustrative embodiments of the present disclosure, those of skill in the art will readily appreciate that yet other embodiments may be made and used within the scope of the claims hereto attached. Numerous advantages of the disclosure covered by this document have been set forth in the foregoing description. For example with the disclosed method a single integrated device may be tested by backside milling method without risking destruction of the device. This contrasts with the prior art method which tended to destroy several identical devices before successfully measuring one of the devices. It will be understood, however, that this disclosure is, in many respects, only illustrative. Changes may be made in details, for example the method looks for significant changes in trend to indicate device stress. This trend identification can be performed through multiple methods. The present embodiment uses principle component analysis on a current waveform. Changes of the principal components over time signify the stress of the sample. Thresholding, rate of change, and other similar algorithms may also be used to indicate critical stress, and thus the stoppage condition. The same algorithm can be applied to other types of sensors. Other algorithm types can include any algorithm that produces trend information from a sample, such as lossy compression, neural networks, curve fitting, and machine learning. The data can be preprocessed in the frequency domain, e.g., fourier transforms, and wavelets, or the time domain, e.g., down sampling, and smoothing filters. Because the nature of internal stresses in modern integrated circuits this method may also be used to determine alterations to a circuit undergoing non-contact-based milling. This may include a focused ion beam, chemical removal, or a laser-based system which does not exert direct force on the backside of the device. The disclosure's scope is, of course, defined in the language in which the appended claims are expressed.