Techniques for training neural networks for use in in situ monitoring during polishing and polishing system - Patents.com

Abstract

A method for polishing a substrate includes polishing a conductive layer on the substrate at a polishing station, monitoring the layer with an in-situ eddy current monitoring system to generate a plurality of measured signal values ​​for a plurality of different locations on the layer, generating thickness measurements for the locations, and detecting a polishing endpoint or modifying polishing parameters based on the thickness measurements. The conductive layer is formed of a first material having a first conductivity. The generating includes calculating an initial thickness value based on the plurality of measured signal values, and processing the initial thickness value through a neural network trained using training data obtained by measuring a calibration substrate having a conductive layer formed of a second material having a second conductivity lower than the first conductivity, for the generated adjusted thickness value.

Description

[Technical field]

[0001] The present disclosure relates to the reconstruction of measurements made near the edge of a substrate using in situ monitoring while the substrate is being polished. [Background technology]

[0002] Integrated circuits are typically formed on a substrate (eg, a semiconductor wafer) by sequentially depositing conductive, semiconductive or insulating layers onto the silicon wafer and then processing the layers.

[0003] Some manufacturing processes include depositing a fill layer on a non-planar surface and planarizing the fill layer until the non-planar surface is exposed. For example, a conductive fill layer can be deposited on a patterned insulating layer to fill trenches or holes in the insulating layer. The fill layer is then polished until the raised pattern of the insulating layer is exposed. After planarization, portions of the conductive layer left between the raised pattern of the insulating layer form vias, plugs, and lines that provide conductive paths between thin film circuits on the substrate. In addition, planarization may be used to planarize the substrate surface for lithography.

[0004] Chemical mechanical polishing (CMP) is one accepted method of planarization. This planarization method typically requires that a substrate be mounted on a carrier head. The exposed surface of the substrate is placed against a rotating polishing pad. The carrier head exerts a controllable load on the substrate, forcing it against the polishing pad. A polishing fluid, such as a slurry containing abrasive polishing particles, is supplied to the surface of the polishing pad.

[0005] During semiconductor processing, it may be important to determine one or more characteristics of a substrate or a layer on a substrate. For example, during a CMP process, it is important to know the thickness of a conductive layer so that the process can be terminated at the correct time. Several methods may be used to determine substrate characteristics. For example, optical sensors may be used to perform in situ monitoring of a substrate during chemical mechanical polishing.

[0006] Alternatively (or additionally), an eddy current sensing system may be used to induce eddy currents in a conductive region on the substrate and determine parameters such as the local thickness of the conductive region. Summary of the Invention

[0007] In one embodiment, a method of polishing a substrate includes polishing a conductive layer on the substrate at a polishing station, monitoring the layer during polishing at the polishing station with an in-situ eddy current monitor system to generate a plurality of measured signal values ​​for a plurality of different locations on the layer, generating an estimate of a thickness of the location for each of the plurality of different locations, and at least one of detecting a polishing endpoint or modifying a polishing parameter based on each thickness estimate. The conductive layer is formed of a first material having a first conductivity. The generating includes calculating an initial thickness value based on the plurality of measured signal values, and processing the initial thickness value through a neural network trained using training data obtained by measuring a calibration substrate having a conductive layer formed of a second material having a second conductivity lower than the first conductivity for the generated adjusted thickness value.

[0008] In another aspect, a method of training a neural network for use with an in-situ monitor system includes scanning a sensor of the in-situ monitor system over a calibration substrate having a conductive layer formed of a first material having a first conductivity, the monitor system generating a first set of sensor measurements. Ground-truth measurements of the thickness of the conductive layer of the calibration substrate are obtained to provide a first thickness profile. The first thickness profile is scaled based on the first conductivity and a target second conductivity greater than the first conductivity to provide a modified second thickness profile that corresponds to a thickness profile that would be generated if the conductive layer were formed of a material of the second conductivity. The neural network is trained to convert the sensor measurements from the in-situ monitor system into thickness measurements of the layer formed of the second material. The training is performed using training data including the modified training profile and calibration thickness values ​​based on the first set of sensor measurements from the conductive layer formed of the first material.

[0009] Certain implementations may include one or more of the following advantages: An in-situ monitoring system, e.g., an eddy current monitoring system, may generate a signal as a sensor scans across the substrate. The system may compensate for distortions in the portion of the signal that corresponds to the substrate edge. The signal may be used for endpoint and / or closed-loop control of polishing parameters, e.g., carrier head pressure, thus providing improved within-wafer non-uniformity (WIWNU) and water-to-wafer non-uniformity (WTWNU). An in-situ monitoring system may generate more accurate thickness measurements in very thin layers, e.g., during a clearing phase.

[0010] The details of one or more embodiments are set forth in the accompanying drawings and the description below. Other aspects, features, and advantages will become apparent from the description and drawings, and from the claims. [Brief description of the drawings]

[0011] [Figure 1A] 1 is a schematic side view in partial cross section of a chemical mechanical polishing station including an eddy current monitoring system; [Figure 1B] FIG. 2 is a schematic plan view of a chemical mechanical polishing station. [Diagram 2] 2 is a schematic top view of a substrate being scanned by a sensor head of the polishing apparatus; [Diagram 3] 4 is a schematic graph of a static formula for determining a thickness of a substrate based on a measured signal. [Figure 4] 4 is a schematic graph of measurement signals acquired while monitoring locations on a substrate. [Diagram 5] 1 is an exemplary neural network. [Figure 6] FIG. 1 is a schematic diagram of two calibration wafers of different materials. [Figure 7] FIG. 2 is a flow diagram of an example process for polishing a substrate. [Figure 8] FIG. 1 is a flow diagram of an exemplary process for generating a thickness estimate using a neural network. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0012] Like reference symbols in the various drawings indicate like elements.

[0013] The polishing apparatus can use an in-situ monitoring system, such as an eddy current monitoring system, to detect the thickness of the outer layer being polished on the substrate. During polishing of the outer layer, the in-situ monitoring system can determine the thickness of different locations of the layer on the substrate. The thickness measurements can be used to trigger a polishing endpoint and / or to adjust processing parameters of the polishing process in real time. For example, the substrate carrier head can adjust the pressure on the backside of the substrate to increase or decrease the polishing rate at the locations of the outer layer. The polishing rate can be adjusted so that the locations of the layer have substantially the same thickness after polishing. The CMP system can adjust the polishing rate so that polishing of multiple locations of the layer is completed at approximately the same time. Such profile control is sometimes referred to as real-time profile control (RTPC).

[0014] In-situ monitoring systems can suffer from signal distortion of measurements close to the substrate edge. For example, eddy current monitoring systems can generate magnetic fields. Near the substrate edge, the signal can be artificially low because the magnetic field only partially overlaps the conductive layer of the substrate. If the eddy current sensor head is large, the signal can drop off sooner and be further reduced because the larger sensor head requires more time and space to pass through the edge region, resulting in lower spatial resolution. A technique for compensating for distortions in the substrate is to use a neural network to generate a modified signal based on the measurement signal generated by the in-situ monitoring system.

[0015] However, near the end of the polishing operation, e.g., at or near the clearing to expose the underlayer, the thickness of the conductive layer becomes thinner and thinner, and the neural network may not be able to reconstruct the signal reliably. Without being limited to a particular theory, the neural network may be trained for edge reconstruction using a calibration substrate of a different material having a relatively thick conductive layer. Such training may not be suitable for configuring the neural network to reconstruct the layer thickness for the thin conductive layer at the substrate edge. In particular, the thin layer on the calibration wafer may include discontinuities, e.g., areas where the underlayer is exposed. As a result, it may not be possible to obtain a ground-truth measurement from the calibration wafer, e.g., using a four-point probe, and as a result, reliable training data for the thin layer may not be available.

[0016] However, measurements can be taken by collecting signals using a calibration wafer having a thicker layer composed of a material with a lower conductivity, and then scaled based on the relative conductance ratio between the two calibration wafers to provide training data corresponding to a thinner layer of higher conductivity. The neural network can be trained using this training data to enhance its ability to generate substrate edge thickness estimates for very thin layers made of highly conductive materials.

[0017] 1A and 1B show an example of a polishing apparatus 100. The polishing apparatus 100 includes a rotatable, disk-shaped platen 120 on which a polishing pad 110 is positioned. The platen is operable to rotate about an axis 125. For example, a motor 121 can turn a drive shaft 124 to rotate the platen 120. The polishing pad 110 can be a dual-layer polishing pad having an outer polishing layer 112 and a softer backing layer 114.

[0018] The polishing apparatus 100 may include a port 130 for dispensing a polishing fluid 132, such as a slurry, onto the polishing pad 110. The polishing apparatus may also include a polishing pad conditioner that abrades the polishing pad 110 to maintain the polishing pad 110 in a consistent abrasive state.

[0019] The polishing apparatus 100 includes at least one carrier head 140. The carrier head 140 is operable to hold the substrate 10 against the polishing pad 110. The carrier head 140 can have independent control of polishing parameters, such as the pressure associated with each substrate.

[0020] In particular, carrier head 140 may include a retaining ring 142 for retaining substrate 10 beneath a flexible membrane 144. Carrier head 140 also includes a plurality of independently controllable, pressurizable chambers defined by the membrane, e.g., three chambers 146a-146c, that can apply independently controllable pressures to associated zones on flexible membrane 144 and thus substrate 10. For ease of illustration, only three chambers are illustrated in FIG. 1, however, there may be one or two chambers, or more than three chambers, e.g., five chambers.

[0021] Carrier head 140 is suspended from a support structure 150, e.g., a carousel or track, and is connected by a drive shaft 152 to a carrier head rotation motor 154 so that the carrier head can rotate about axis 155. Optionally, carrier head 140 can oscillate laterally, e.g., on a slider on the carousel 150 or track, or by rotational oscillation of the carousel itself. In operation, the platen rotates about its central axis 125 and the carrier head rotates about its central axis 155 and translates laterally across the upper surface of the polishing pad.

[0022] Although only one carrier head 140 is shown, more carrier heads can be provided to hold additional substrates so that the surface area of ​​the polishing pad 110 can be used efficiently.

[0023] The polishing apparatus 100 also includes an in-situ monitor system 160 that generates a series of time-varying values ​​that are dependent on the thickness of a layer on the substrate. The in-situ monitor system 160 includes a sensor head from which measurements are generated, and due to relative motion between the substrate and the sensor head, the measurements will be taken at different locations on the substrate.

[0024] The in-situ monitoring system 160 may be an eddy current monitoring system. The eddy current monitoring system 160 includes a drive system for inducing eddy currents in a conductive layer on the substrate and a sensing system for detecting the eddy currents induced in the conductive layer by the drive system. The monitoring system 160 includes a core 162 positioned in the recess 128 to rotate with the platen, at least one coil 164 wound around a portion of the core 162, and a drive and sense circuit 166 connected to the coil 164 by wiring 168. The combination of the core 162 and the coil 164 may provide a sensor head. In some implementations, the core 162 protrudes above the top surface of the platen 120, for example, into the recess 118 in the bottom of the polishing pad 110.

[0025] Drive and sense circuitry 166 is configured to apply an oscillating electrical signal to coil 164 and measure the resulting eddy currents. Various configurations of the drive and sense circuitry and the coil or coils are possible, as described, for example, in U.S. Patent Nos. 6,924,641, 7,112,960, and 8,284,560, and U.S. Patent Publication Nos. 2011-0189925 and 2012-0276661. The drive and sense circuitry 166 could be located in the same recess 128 or in a different portion of the platen 120, or could be located outside the platen 120 and coupled to components within the platen through a rotary electrical union 129.

[0026] In operation, the drive and sense circuitry 166 drives the coil 164 to generate an oscillating magnetic field. At least a portion of the magnetic field extends through the polishing pad 110 and into the substrate 10. If a conductive layer is present on the substrate 10, the oscillating magnetic field generates eddy currents in the conductive layer. The eddy currents cause the conductive layer to act as an impedance source that is coupled to the drive and sense circuitry 166. As the thickness of the conductive layer changes, the impedance changes, which can be detected by the drive and sense circuitry 166. Different conductive layer materials have different conductive properties that result in different signal strengths from the eddy current sensor. Copper (Cu) has a high electrical conductivity and is therefore a common material for the conductive layer. Other conductive materials (e.g., cobalt, tungsten, titanium nitride) may be used for the conductive layer based on the particular application, e.g., barrier layer, ferromagnetic memory, etc.

[0027] Alternatively or additionally, an optical monitoring system, which may function as a reflectometer or interferometer, may be fixed to the platen 120 within the recess 128. If both systems are used, the optical monitoring system and the eddy current monitoring system may monitor the same portion of the substrate.

[0028] The CMP apparatus 100 may also include a position sensor 180, such as an optical interrupter, to sense when the core 162 is under the substrate 10. For example, the optical interrupter could be attached to a fixed point on the opposite side of the carrier head 140. A flag 182 is attached to the periphery of the platen. The attachment point and length of the flag 182 is selected to interrupt the optical signal of the sensor 180 while the core 162 sweeps under the substrate 10. Alternatively or additionally, the CMP apparatus may include an encoder for determining the angular position of the platen.

[0029] A controller 190, such as a general purpose programmable digital computer, receives the intensity signals from the eddy current monitoring system 160. The controller 190 can include a processor, memory, and I / O devices, as well as output devices 192, such as a monitor, and input devices 194 (e.g., a keyboard).

[0030] Signals may be passed from the eddy current monitor system 160 through the rotating electrical union 129 to the controller 190. Alternatively, the circuitry 166 could communicate with the controller 190 by wireless signals.

[0031] As the core 162 sweeps under the substrate with each platen revolution, information regarding the thickness of the conductive layer is accumulated in situ and on a continuous real-time basis (once per platen revolution). The controller 190 can be programmed to sample measurements from the monitoring system when the substrate is generally over the core 162 (as determined by the position sensor). As polishing progresses, the thickness of the conductive layer changes and the sampled signal changes over time. The sampled signal that changes over time is sometimes referred to as a trace. Measurements from the monitoring system can be displayed on an output device 192 during polishing to allow an operator of the machine to visually monitor the progress of the polishing operation.

[0032] During operation, the CMP apparatus 100 can use the eddy current monitoring system 160 to determine when the bulk of the fill layer has been removed and / or when the underlying stop layer has been substantially exposed. Possible process control and endpoint criteria for the detector logic include local minimums or maximums, changes in slope, amplitude or slope thresholds, or combinations thereof.

[0033] The controller 190 may also be connected to a pressure mechanism that controls the pressure applied by the carrier head 140, a carrier head rotation motor 154 that controls the carrier head rotation speed, a platen rotation motor 121 that controls the platen rotation speed, or a slurry distribution system 130 to control the slurry composition delivered to the polishing pad. In addition, the computer 190 may be programmed to divide the measurements from the eddy current monitor system 160 for each sweep under the substrate into multiple sampling zones, calculate the radial position of each sampling zone, and sort the amplitude measurements into radial ranges, as discussed in U.S. Pat. No. 6,399,501. After sorting the measurements into radial ranges, information about the film thickness may be provided in real time to a closed loop controller to periodically or continuously modify the polishing pressure profile applied by the carrier head to provide improved polishing uniformity.

[0034] The controller 190 can use a correlation curve that relates the signal values ​​measured by the in-situ monitor system 160 to the thickness of the layer being polished on the substrate 10 to calculate the thickness value of the layer being polished. An example of a correlation curve 303 is shown in FIG. 3. In the coordinate system depicted in FIG. 3, the horizontal axis represents the value of the signal received from the in-situ monitor system 160, and the vertical axis represents the value for the thickness of the layer of the substrate 10. For a given signal value, the controller 190 can use the correlation curve 303 to generate a corresponding thickness value. The correlation curve 303 can be considered a "static" formula in that it predicts a thickness value for each signal value regardless of the time or location at which the sensor head acquired the signal. The correlation curve can be represented by a variety of functions, such as a polynomial function or a look-up table (LUT) combined with linear interpolation.

[0035] 1B and 2, a change in the position of the sensor head relative to the substrate 10 can result in a change in the signal from the in-situ monitoring system 160. That is, as the sensor head scans across the substrate 10, the in-situ monitoring system 160 will measure multiple regions 94, e.g., measurement spots, at different locations on the substrate 10. The regions 94 may partially overlap (see FIG. 2).

[0036] 4 illustrates a graph 420 showing a signal 401 from the in situ monitor system 160 during a single pass of the sensor head under the substrate 10. The signal 401 is composed of a series of individual measurements from the sensor head as it sweeps under the substrate. The graph 420 can be a function of measurement time or the position of the measurement on the substrate, e.g., radial position. In either case, different portions of the signal 401 correspond to measurement spots 94 at different locations on the substrate 10 scanned by the sensor head. Thus, the graph 420 shows the corresponding measured signal values ​​from the signal 401 for a given location of the substrate scanned by the sensor head.

[0037] 2 and 4, the signal 401 includes a first portion 422 corresponding to a location in the edge region 203 of the substrate 10 as the sensor head crosses the leading edge of the substrate 10, a second portion 424 corresponding to a location in the central region 201 of the substrate 10, and a third portion 426 corresponding to a location in the edge region 203 as the sensor head crosses the trailing edge of the substrate 10. The signal can also include a portion 428 corresponding to off-substrate measurements, i.e., a signal generated when the sensor head scans an area beyond the edge 204 of the substrate 10 in FIG.

[0038] The edge region 203 may correspond to a portion of the substrate where the measurement spot 94 of the sensor head overlaps the substrate edge 204. The central region 201 may include an annular anchor region 202 adjacent the edge region 203 and an interior region 205 surrounded by the anchor region 202. The sensor head may scan these regions on its path 210 and generate a sequence of measurements corresponding to the sequence of locations along the path 210.

[0039] In the first portion 422, the signal strength increases from an initial strength (typically a signal obtained when the substrate and carrier head are not present) to a higher strength by transitioning the monitoring location from initially only slightly overlapping the substrate at the edge 204 of the substrate (producing an initial low value) to a monitoring location that almost entirely overlaps the substrate (producing a high value). Similarly, in the third portion 426, the signal strength decreases as the monitoring location transitions to the edge 204 of the substrate.

[0040] As the bulk of the fill layer is removed, the metal layer in the edge region 204 of the substrate becomes thinner and discontinuities may develop as the raised pattern of the insulating layer becomes exposed. This leads to more noise in the first portion 422 and the third portion 426, as well as the second portion 424 of lower signal amplitude. In addition, when a larger eddy current sensor is used to detect a thin conductive fill layer, there is a loss in the spatial resolution of the sensor. This is because the first portion 422 and the third portion 426 become more elongated as the monitor location overlaps the edge 204 of the substrate earlier and longer, further increasing the distortion of the signal due to the sensor head overlapping the substrate edge.

[0041] Although second portion 424 is shown as flat, this is for simplicity and the actual signal in second portion 424 would likely include variations due to both noise and layer thickness variations. Second portion 424 corresponds to a monitor location scanning central region 201. Second portion 424 includes subportions 421 and 423 that result from the monitor location scanning anchor region 202 of central region 201 and subportion 427 that result from the monitor location scanning interior region 205 of central region 201.

[0042] As discussed above, the variation in signal strength in regions 422, 426 is caused in part by the measurement region of the sensor overlapping the substrate edge, rather than by inherent variations in the thickness or conductivity of the layer being monitored. As a result, this distortion in signal 401 can cause errors in the calculation of characteristic values ​​for the substrate, e.g., layer thickness, near the substrate edge. To address this issue, controller 190 can include a neural network, e.g., neural network 500 of FIG. 5, to generate modified thickness values ​​corresponding to one or more locations of substrate 10 based on the calculated thickness values ​​corresponding to these locations.

[0043] 5, a neural network 500, when properly trained, is configured to generate modified thickness values ​​that reduce and / or eliminate distortions in calculated thickness values ​​near the substrate edge. The neural network 500 receives a group of inputs 504 and processes the inputs 504 through one or more neural network layers to generate a group of outputs 550. The layers of the neural network 500 include an input layer 510, an output layer 530, and one or more hidden layers 520.

[0044] Each layer of the neural network 500 includes one or more neural network nodes. Each neural network node in a neural network layer receives one or more node input values ​​(from inputs 504 to the neural network 500 or from the outputs of one or more nodes of a previous neural network layer), processes the node input values ​​according to one or more parameter values ​​to generate activation values, and optionally applies a nonlinear transformation function (e.g., a sigmoid function or a hyperbolic tangent function) to the activation values ​​to generate an output for the neural network node.

[0045] Each node in the input layer 510 receives one of the inputs 504 to the neural network 500 as a node input value.

[0046] Inputs 504 to the neural network include initial thickness values ​​from the in situ monitor system 160 for a number of different locations on the substrate 10, such as a first initial thickness value 501, a second initial thickness value 502, etc. through an nth calculated thickness value 503. The initial thickness values ​​may be individual thickness values ​​calculated from a sequence of signal values ​​in the signal 401 using a correlation curve.

[0047] In general, the plurality of different locations include locations within the edge region 203 and anchor region 202 of the substrate 10. In some implementations, the plurality of different locations are only within the edge region 203 and anchor region 202. In other implementations, the plurality of different locations spans all regions of the substrate. In other implementations, the plurality of different locations may include locations within the edge region 203 and anchor region 202 of the second substrate 10. In some implementations, the plurality of different locations are only within the edge region 203 and anchor region 202 of the second substrate. In other implementations, the plurality of different locations spans all regions of the second substrate.

[0048] These initial thickness values ​​are received at a signal input node 544. Optionally, the input nodes 504 of the neural network 500 may also include one or more state input nodes 516 that receive one or more process state signals 504, such as measurements of wear on the pad 110 of the polishing apparatus 100.

[0049] The nodes of the hidden layer 520 and the output layer 530 are shown as receiving inputs from all nodes of the preceding layer. This is the case for a fully connected feedforward neural network. However, the neural network 500 may be a non-fully connected feedforward neural network or a non-feedforward neural network. Furthermore, the neural network 500 may include at least one of one or more fully connected feedforward layers, one or more non-fully connected feedforward layers, and one or more non-feedforward layers.

[0050] The neural network generates a group of modified thickness values ​​550 at nodes or "output nodes" 550 in the output layer 530. In some implementations, there is an output node 550 for each measurement signal from the in situ monitoring system that is fed to the neural network 500. In this case, the number of output nodes 550 may correspond to the number of signal input nodes 504 in the input layer 510.

[0051] For example, the number of signal input nodes 544 can be equal to the number of measurements in the edge regions 203 and anchor regions 202, and there can be an equal number of output nodes 550. Each output node 550 thus generates a modified thickness value corresponding to a respective measurement signal provided as an input to the signal input node 544, e.g., a first modified thickness value 551 for the first measurement signal 501, a second modified thickness value 552 for the second measurement signal 502, and an nth modified thickness value 553 for the nth measurement signal 503.

[0052] In some implementations, the number of output nodes 550 is less than the number of input nodes 504. In some implementations, the number of output nodes 550 is less than the number of signal input nodes 544. For example, the number of signal input nodes 544 can be equal to the number of measurements in the edge regions 203, or equal to the number of measurements in the edge regions 203 and the anchor regions 202. Again, each output node 550 of the output layer 530 generates a modified thickness value corresponding to a respective measurement signal provided as a signal input node 504, e.g., a first modified thickness value 551 for the first measurement signal 501, but only for signal input nodes 554 that receive signals from the edge regions 203.

[0053] The polishing apparatus 100 can use the neural network 500 to generate the modified thickness value. The modified thickness value can then be used as the determined thickness for each location in the first group of locations of the substrate, for example, the locations in the edge region (and possibly the anchor region). For example, referring again to FIG. 4, the modified thickness value for the edge region can provide the modified portion 430 of the signal 401.

[0054] For example, the controller 190 can use the neural network 500 to determine modified thickness values ​​for each of one or more locations in the edge region. Optionally, the controller 190 can use the neural network 500 to determine modified thickness values ​​for each of one or more locations in the anchor region of the substrate. In contrast, the controller 190 can use initial thickness values ​​for locations in other regions, such as the interior region 205. That is, signal values ​​from locations in other regions, such as the interior region 205, can be converted to thickness values ​​that are then used without being modified by the neural network.

[0055] In some implementations, for a modified thickness value corresponding to a given measurement location, the neural network 500 can be configured such that only input signal values ​​from measurement locations within a predetermined distance of the given location are used in determining the modified thickness value. 1 ,S 2 ,..,S M ,…S N is received and corresponds to measurements at N consecutive locations on the path 210, then the Mth location (R M The modified thickness value S' M is the modified thickness value S' M To calculate the thickness value S M-L(min1) ,…S M ,…S M+L(maxN) Only the value of L can be used for a given modified thickness value S' using measurements up to about 2-4 mm apart. M You can choose to generate measurements S M Measurements within about 1-2 mm, e.g., 1.5 mm, of the location of can be used. For example, L can be a number in the range of 0-4, e.g., 1 or 2. For example, if measurements within 3 mm are used and the spacing between measurements is 1 mm, L can be 1. If the spacing is 0.5 mm, L can be 2. If the spacing is 0.25, L can be 4. However, this can depend on the polishing equipment configuration and process conditions. Modified thickness value S' M The values ​​of other parameters, such as pad wear, could still be used in calculating .

[0056] For example, there may be a number of hidden nodes 570, or "hidden nodes" 570, in one or more hidden layers 520, equal in number to the number of signal input nodes 544, with each hidden node 570 corresponding to a respective signal input node 544. Each hidden node 570 may be disconnected (or may have a parameter value of 0 for the input node 544) from input nodes 544 that correspond to measurements at locations greater than a predetermined distance from the location of the corresponding input node's measurement. For example, the Mth hidden node may be disconnected (or may have a parameter value of 0 for the hidden node 570) from the first through (ML-1)th input nodes 544 and from the (M+L+1)th through Nth input nodes. Similarly, each output node 560 may be disconnected (or may have a parameter value of 0 for the hidden node 570) from hidden nodes 570 that correspond to altered thicknesses at locations greater than a predetermined distance from the location of the output node's measurement. For example, the Mth output node may be disconnected from the first through (ML-1)th hidden nodes 570 and the (M+L+1)th through Nth hidden nodes (or may have a parameter value of 0 for the hidden nodes).

[0057] To train a neural network, a training data set is required, which includes a number of records, each record including a group of input values ​​that may be input to an input node and a group of desired output values ​​from an output node, i.e., the output values ​​that may be desired for that group of input values.

[0058] One technique for generating this training data set is to obtain a group of calibration substrates, each with a different thickness profile. The calibration substrates are a group of substrates that may have the same dimensions and area as the substrate to be polished. The layer thicknesses may be accurately measured, for example, using a four-point probe, to generate a ground-truth measurement. Generally, the calibration substrate includes a layer of the same material as that used as the conductive fill layer to be removed during polishing, for example, a copper calibration layer for a copper conductive fill layer.

[0059] For each calibration substrate, the eddy current sensor of the polishing apparatus 100 can be scanned across the calibration substrate, and a sequence of calibration signal values ​​is generated. These calibration signal values ​​are similar to the sequence of signal values ​​obtained during polishing, but no pressure or polishing liquid is applied, so that polishing of the calibration substrate is not performed. The calibration signal values ​​can be converted to calibration thickness values ​​using a correlation curve. In addition, for each calibration substrate, ground truth thickness measurements are generated at multiple locations on the calibration substrate, for example, using a four-point probe. The signal values ​​and ground truth thickness measurements can be collected in several regions, for example, the central region 201, the anchor region 202, and the edge region 203. The calibration thickness values ​​and ground truth thickness measurements of a particular calibration substrate thus provide a record in the training data set.

[0060] Assuming that the different calibration substrates have sufficiently different thickness profiles and span a sufficient range of thicknesses, the data set can be used to train a neural network. Thus, in a polishing operation in which a first substrate is polished and monitored, the modified thickness values ​​generated by the neural network 500 can be considered to be based on thickness values ​​from measurement locations on a second substrate (the calibration substrate).

[0061] Particular problems arise in generating modified thickness values ​​at locations in edge regions where the bulk of the fill layer has been removed. As the conductive fill layer is removed and the raised pattern of the insulating layer is exposed, a larger eddy current sensor may be required to detect the thin conductive fill layer. Exposing the insulating layer results in a decrease in signal quality, as the eddy current sensor may span an area that includes both conductive and insulating layer areas. A larger eddy current sensor will cover a larger area, thereby covering more of both the conductive and insulating layer areas. This may cause a decrease in the slope of the first portion of the signal 422 and an increase in noise. There is further signal reduction as the path of the eddy current sensor 210 head attempts to measure the signal from the edge regions 203. Thus, these regions may have a high variability between the measured initial thickness value and the actual. Training data is needed to train the neural network to accurately generate modified thickness values, particularly at locations in edge regions where the bulk of the fill layer has been removed.

[0062] However, obtaining such data and thus reconstructing the thickness of the thin conductive layer at the edge region 203 is more difficult because it is difficult to fabricate a calibration wafer of the same material and thickness as the conductive fill layer during later stages of polishing, e.g., clearing. This problem can be at least partially ameliorated by using a calibration wafer of a first material with lower conductivity but similar conductance to perform neural network training on a substrate of a second material.

[0063] Referring now to FIG. 6, the calibrated thickness values ​​and ground truth thickness measurements obtained using a four-point probe and an eddy current sensor, respectively, of a group of calibration wafers having a layer of a first material 601 can be used as training data for training a neural network, even when the neural network is then used for in situ monitoring of a group of substrates having a layer of a second material 602 of different conductivity.

[0064] In particular, to simulate a second calibration substrate having a thin layer of a second conductivity material, a first calibration substrate having a thicker layer, but less conductive, of the first conductivity material can be used, since the conductive layer of the first calibration substrate will have a similar conductance to the simulated second calibration substrate. Conductance (G) is a measure of how well a material conducts electricity, TIFF0007640562000001.tif10170, where σ is the conductivity, A is the cross-sectional area, and l is defined as the length or thickness of the material being measured.

[0065] The electrical conductivity of the first material 601 is lower than the electrical conductivity of the second material 602, e.g. In the case of TIFF0007640562000002.tif11161, the thickness 610 of the first material can be proportionally greater than the thickness 612 of the second material 611, e.g. TIFF0007640562000003.tif11169 maintains similar conductance. This allows the operator to select an approximate range of thicknesses for the conductive layers in the calibration substrate. For example, a copper layer with a thickness of 0.2-0.5 microns (ρ=5.96×10 7 If there is difficulty in obtaining a calibration substrate with a layer of 5-12.5 microns of titanium (ρ = 2.38 × 10 6 ) could be used.

[0066] For each calibration substrate having a layer of a second material, an eddy current sensor of the polishing apparatus 100 can be scanned across the calibration substrate to generate a sequence of calibration signal values. The calibration signal values ​​can be converted to calibration thickness values ​​using a correlation curve. The correlation curve assumes that the material has a conductivity related to the first material, but because the conductance of the layers is similar, the eddy current sensor will result in a calibration thickness value that appears as a thinner layer of the first material.

[0067] In addition, for each calibration substrate having a layer of a second material, calibrated thickness measurements are generated for multiple locations on the calibration substrate, for example, using a four-point probe. The calibrated thickness measurements of the calibration substrate are then scaled by the ratio of electrical conductivity between the first and second materials to generate adjusted thickness values. For example, if the substrate being polished has a copper layer and the correlation curve is similarly calculated for the copper layer (ρ=5.96x10 7 Assume that the thickness of the calibration substrate is to be monitored (ρ = 2.38x10 6 Ωm, the correct thickness measurement can be divided by a factor of about 25 (5.96x10 7 / 2.38x10 6 ), producing an adjusted thickness value that corresponds to a copper layer having the same conductance (and thus would produce the same calibrated thickness measurement using an eddy current sensor).

[0068] For each calibration substrate having a layer of a second material, the calibrated thickness value and the adjusted thickness measurements for that calibration substrate can provide a training data set record. This allows for the generation of a training data set that includes data simulating substrates having very thin layers. This allows for the training of a neural network to appropriately convert initial thickness values ​​to adjusted thickness values ​​where the layers are very thin, thus improving the accuracy and reliability of the in-situ monitor system.

[0069] Possible materials for the first and second materials include copper, aluminum, tungsten, cobalt, titanium, and titanium nitride. As mentioned above, the first material can have a higher electrical conductivity than the second material. In some implementations, the first material is copper, aluminum, or tungsten. In some implementations, the second material is cobalt, titanium, or titanium nitride.

[0070] In some embodiments, the training data set includes data acquired only from calibration substrates having a conductive layer of a different composition than the first material, hi some embodiments, the training data set includes data acquired from a first group of calibration substrates having a conductive layer of a second material, the second material being of a different composition than the first material, and data acquired from a second group of calibration substrates having a conductive layer of the first material.

[0071] 7 is a flow diagram of an example process 700 for polishing substrate 10. Process 700 can be performed by polishing apparatus 100.

[0072] The polishing apparatus 100 polishes (702) a layer on a substrate 10 and monitors (704) the layer during polishing to generate measured signal values ​​for different locations on the layer. The locations on the layer may include one or more locations in the edge region 203 of the substrate (corresponding to the region 422 / 426 of the signal 401) and one or more locations in the anchor region 202 on the substrate (corresponding to the region 421 / 423 of the signal). The anchor region 202 is spaced apart from the substrate edge 204 and in the central region 201 of the substrate, and is therefore not affected by the distortion created by the substrate edge 204. However, the anchor region 202 may be adjacent to the edge region 203. The anchor region 202 may also surround the interior region 205 of the central region 201. The number of anchor locations may depend on the measurement spot size and measurement frequency by the in situ monitor system 160. In some embodiments, the number of anchor locations may not exceed a maximum value (such as a maximum of 4).

[0073] The polishing apparatus 100 generates an initial thickness value for each of the different locations based on applying a static formula to the measurement signal for the location (706).

[0074] The inputs to the neural network 500 may be calculated thickness values ​​generated by the in situ monitor system 160 for different locations. In some embodiments, the apparatus 100 updates each measurement signal by normalizing the value of the signal. Such normalization may increase the likelihood that at least some of the inputs 504 to the neural network system 500 fall within a particular range. As a result, the quality of the neural network training and / or the accuracy of the estimations made by the neural network 500 may be increased. The normalized signal values ​​may then be made into normalized thickness values ​​by applying a static formula.

[0075] The output of the neural network 500 is a modified thickness value corresponding to each input calculated thickness value. If the input is a normalized thickness value, the modified thickness value corresponding to the calculated thickness value will also be a normalized thickness value. Therefore, the polishing apparatus 100 may need to convert such modified thickness values ​​to non-normalized values ​​before using the modified signal to estimate the thickness of the substrate.

[0076] The polishing apparatus 100 detects a polishing endpoint and / or modifies polishing parameters based on each modified thickness value (708).

[0077] 8 is a flow diagram of an exemplary process 800 for generating modified thickness values ​​using the neural network 500. The process 800 may be performed by the polishing apparatus 100.

[0078] The polishing apparatus 100 identifies anchor locations for a group of substrate locations 802 and acquires measurement signals for each location in the group of locations 804. In some embodiments, the anchor locations are spaced apart from the edge of the substrate.

[0079] The polishing apparatus 100 normalizes (806) each measured signal value based on the measured signal value of the anchor location, for example, by dividing each measured signal value at all locations by the measured signal value of the anchor location, to update the measurement signals. The polishing apparatus 100 then processes (808) the normalized measured signal values ​​through the neural network 500 to generate modified thickness values ​​for each normalized measured signal value. The polishing apparatus 100 then converts (810) the modified thickness values ​​to unnormalized signal values ​​using the measured signal values ​​of the anchor location, for example, by multiplying each measured signal value at all locations by the measured signal value of the anchor location, to update the measurement signals. The polishing apparatus 100 then uses (612) the unnormalized modified signal values ​​to generate calculated thickness values ​​for each location of the group of locations of the neural network 500.

[0080] The monitor system can be used in a variety of polishing systems. Either the polishing pad or the carrier head, or both, can move to provide relative motion between the polishing surface and the substrate. The polishing pad can be a circular (or some other shape) pad fixed to the platen, a tape running between a supply roller and a take-up roller, or a continuous belt. The polishing pad can be fixed on the platen and incrementally advanced on the platen during the polishing operation, and driven continuously on the platen during polishing. Sometimes the pad is fixed to the platen during polishing, and a fluid bearing is present between the platen and the polishing pad during polishing. The polishing pad can be a standard rough pad (e.g., polyurethane with or without fillers), a soft pad, or a fixed abrasive polishing pad.

[0081] Although the above discussion has focused on eddy current monitoring systems, the modification techniques can be applied to other types of monitoring systems that scan over the edge of a substrate, such as optical monitoring systems. Additionally, although the above discussion has focused on polishing systems, the modification techniques can be applied to other types of substrate processing systems, such as deposition or etching systems, including in situ monitoring systems that scan over the edge of a substrate.

[0082] A number of embodiments of the invention have been described. Nevertheless, it will be understood that various modifications can be made without departing from the spirit and scope of the invention. Accordingly, other embodiments are within the scope of the following claims.

Claims

1. 1. A method of polishing a substrate, comprising: polishing a conductive layer on the substrate at a polishing station, the conductive layer being formed of a first material having a first electrical conductivity; monitoring the layer during polishing at the polishing station with an in situ eddy current monitoring system to generate a plurality of measured signal values ​​for a plurality of different locations on the layer; generating a thickness measurement for the location, the thickness measurement comprising calculating an initial thickness value based on the plurality of measured signal values; generating a thickness measurement for the location, the thickness measurement comprising: processing the initial thickness value through a neural network trained using training data obtained by measuring a calibration substrate having a conductive layer formed of a second material having a second conductivity lower than the first conductivity to generate an adjusted thickness value; At least one of detecting a polishing endpoint or modifying polishing parameters based on the thickness measurement. wherein the neural network is also trained using training data obtained by measuring a calibration substrate having a conductive layer formed of the first material.

2. 2. The method of claim 1, wherein the first material and the second material are selected from the group consisting of copper, aluminum, cobalt, tungsten, titanium, and titanium nitride.

3. The method of claim 1 , wherein the first material is copper and the second material is tungsten or titanium nitride.

4. The method of claim 1 , wherein calculating an initial thickness value comprises converting the measured signal values ​​to initial thickness values ​​using a correlation curve.

5. The method of claim 4 , comprising inputting the initial thickness value into an input node of the neural network and receiving the thickness estimate from an output node.

6. 6. The method of claim 5, further comprising generating a thickness value for each location of a second plurality of different locations by generating signal values ​​with a sensor of the in situ eddy current monitoring system and converting the signal values ​​to initial thickness values ​​using the correlation curve, the initial thickness values ​​providing an estimate of the thickness without modifying the initial thickness values ​​using the neural network.

7. 1. A method of training a neural network for use in an in situ monitor system, comprising: scanning a sensor of an in situ monitoring system across a calibration substrate having a conductive layer formed of a first material having a first electrical conductivity, the monitoring system generating a first set of sensor measurements; obtaining a corrective measurement of the thickness of the conductive layer of the calibration substrate to provide a first thickness profile; increasing or decreasing the first thickness profile based on the first conductivity and a target second conductivity greater than the first conductivity to provide a modified second thickness profile that corresponds to a thickness profile that would be produced if the conductive layer were formed of a second material having the second conductivity; training a neural network to convert sensor measurements from the in-situ monitor system into thickness measurements for a layer formed of the second material, the training being performed using training data including the modified second thickness profile and calibration thickness values ​​based on the first set of sensor measurements from the conductive layer formed of the first material; Including, The method of claim 1, wherein the calibrated thickness value is calculated by receiving signal values ​​from the sensors of the in situ monitoring system and converting the signal values ​​to the calibrated thickness values ​​using a correlation curve.

8. The method of claim 7 , wherein the sensor comprises an eddy current sensor.

9. The method of claim 7 , wherein obtaining corrective measurements includes measuring the calibration substrate with a four-point probe.

10. scanning the sensor of the in situ monitoring system across a second calibration substrate having a second conductive layer formed of the second material, the monitoring system generating a second set of sensor measurements; obtaining a second corrective measurement of the thickness of the second conductive layer of the second calibration substrate to provide a second thickness profile; training the neural network to convert sensor measurements from the in situ monitor system into thickness measurements for a layer formed of the second material, the training being performed using training data including the modified second thickness profile and a second calibration thickness value based on the second set of sensor measurements from the conductive layer formed of the second material; 8. The method of claim 7, wherein the second calibrated thickness value is calculated by receiving signal values ​​from the sensor of the in situ monitoring system and converting the signal values ​​to the second calibrated thickness value using the correlation curve.

11. 8. The method of claim 7, wherein the first material and the second material are selected from the group consisting of copper, aluminum, cobalt, tungsten, titanium, and titanium nitride.

12. The method of claim 11 , wherein the second material is copper and the first material is tungsten or titanium nitride.

13. The method of claim 7 , wherein increasing or decreasing the first thickness profile comprises dividing a ground truth measurement by a ratio of the second conductivity to the first conductivity.

14. A computer program product for carrying out the method of any one of claims 1 to 13.

15. a platen for supporting the polishing pad; a carrier head for holding a substrate in contact with the polishing pad; an in-situ eddy current monitor system; A controller configured to carry out the method according to any one of claims 1 to 6. A polishing system equipped with

Images