Method and device for determining characteristics of devices produced on a wafer

Abstract

Method and device for determining characteristics of devices produced on a wafer. The invention relates to a computer-implemented method for inferring device characteristics () of devices (2) produced on a wafer (1); comprising the steps of: - providing (S2) a wafer characteristic model associating wafer positions to device characteristics () of devices (2) produced on a wafer (1), wherein the wafer characteristic model is configured to be trained by one or more wafer characteristic maps, and in particular is configured to be a Gaussian process model; - providing (S3) sample device characteristics of at least one device (2) at sample wafer positions; - inferring (S4) device characteristics () of at least one other device of the wafer (1) depending on the provided wafer characteristic model.

Description

Technical Field

[0001] This invention relates to associating features with devices manufactured on a wafer. Background Technology

[0002] Currently, in the final testing of devices (such as chips) manufactured on wafers, each device is measured or tested individually before or after device separation. Measurements / tests aim to obtain one or more device characteristics, including indications of functionality / faults or other data used to determine one or more training or calibration parameters. Calibration parameters, for example, are used to set calibration bits, which is particularly common for calibrating integrated sensors.

[0003] Even after device separation, a unique wafer ID can be used to identify the device, allowing for regression to a specific location on the wafer. Using this method, a wafer map can be reconstructed by plotting device features acquired during the measurement / testing process over the reconstructed location in the wafer map.

[0004] Due to the various process steps during device fabrication on a wafer, different patterns can be observed on the wafer pattern. These patterns originate from process variations caused by the processing equipment (such as etching losses, trench angle mismatches, etc.) as well as inherent variations on the wafer surface.

[0005] Similarly, for different wafers produced using the same devices, i.e., wafers obtained through the same mask layout and processing steps, the devices will exhibit very similar device feature patterns relative to their wafers. Summary of the Invention

[0006] According to the present invention, a method for determining one or more device features of a device manufactured on a wafer based on sample measurements is provided according to claim 1, as well as a device according to another independent claim.

[0007] Further embodiments are indicated by the dependent claims.

[0008] According to a first aspect, a method for determining device characteristics of a device manufactured on a wafer is provided, comprising the following steps:

[0009] - Provides a wafer feature model that associates wafer location with device features, the wafer location indicating the location of a device manufactured on the wafer, wherein the wafer feature model is configured to be trained by one or more wafer feature maps, and is specifically configured as a Gaussian process model;

[0010] - Provide sample device features for at least one device at a subset of sample wafer locations; and

[0011] - Determine the device characteristics of at least one other device of the (same) wafer, depending on the provided wafer feature model.

[0012] Based on the method described above, a device feature model is provided, which is trained to associate the wafer location of a device manufactured on a wafer with one or more device features. The one or more device features may include indications of functionality / faults or other data such as one or more training or calibration parameters.

[0013] Furthermore, the wafer feature model can be nonparametric, and it can be trained using sample data. Additionally, the wafer feature model can provide uncertainty values ​​for the predicted features.

[0014] In addition to wafer location, the device feature model can optionally associate one or more additional operating parameters with one or more device features.

[0015] By leveraging device features, a wafer feature map can be constructed for wafer locations. Using such a wafer feature map, sample wafer locations of the device can be determined, which will be used for measurement / testing / probing of the corresponding device features. A subset of wafer locations may include one or more sample wafer locations. The results of testing / measuring the device at the subset of wafer locations are used to determine the corresponding device features of the device at other wafer locations.

[0016] The location of a sample wafer can be determined by minimizing prediction uncertainty based on a given requirement violation. Thus, only a small number of devices can be selected for testing / measurement to obtain one or more corresponding device characteristics, allowing for the determination of corresponding device characteristics for other devices manufactured on the same wafer.

[0017] By using a device feature model, the above method allows for the acquisition of features for all devices on the same wafer or other wafers manufactured using the same devices, simply by determining one or more corresponding device features at the location of a sample wafer. Therefore, this invention utilizes similar patterns of one or more device feature distributions for device feature prediction, said similar patterns being the result of common processing steps and parameters used in wafer manufacturing.

[0018] By replacing the expensive determination of all device features (including physical measurement and testing) for all devices on a wafer with the determination of correspondences for a preselected subset of devices, measurement / testing costs can be significantly reduced. Depending on the accuracy of the wafer feature model and the given specification limits for the determined features, only a small fraction of the devices need to be tested / measured.

[0019] In addition, device characteristics may include one of the following: indications of functionality, indications of errors, indications of whether the device meets a given specification, and calibration parameters to be written / stored in a particular device.

[0020] It can be specified that the wafer feature model additionally associates environmental and / or operating conditions with device features.

[0021] According to one embodiment, a wafer feature model can be trained by providing multiple processed wafers including the same devices, wherein, in particular, the processed wafers include one or more corner lot wafers.

[0022] It can be specified that the sample device features of the at least one device are obtained at a selected discrete subset of wafer locations, the selected discrete subset of wafer locations maximizing the likelihood of all samples on the wafer within a given specification limit.

[0023] According to one embodiment, the following formula is used to maximize all samples on the wafer within a given specification limit. Internal likelihood is used to select a discrete subset of wafer locations.

[0024]

[0025] in

[0026]

[0027] and It concerns each sample meeting the specification limits. The lower bound of the minimum probability, where It is the set of all chip locations. It is a modeling feature, and These are the nominal characteristics that a device at a specific location i must satisfy.

[0028] Furthermore, the limit violation v can be specified as

[0029]

[0030] A subset of wafer locations is selected to minimize the probability of expected limit violations across all devices on the wafer.

[0031] A subset of wafer locations can be specified to be updated using the following fetch function:

[0032]

[0033] Here, wafer locations s are selected for the optimized subset of wafer locations, and an acquisition function is evaluated for all possible combinations of wafer locations, where the subset of wafer locations with the lowest expected limit violation is selected.

[0034] Furthermore, based on prior technical knowledge of the manufacturing steps using the processing equipment, prior knowledge of the corresponding pattern can be used to provide a device feature model.

[0035] According to another aspect, a computer-implemented method for manufacturing devices on a wafer is provided, comprising the following steps:

[0036] - Control the production of devices on one or more wafers;

[0037] -The device characteristics are determined using the methods described above.

[0038] The control of device production depends on the determined device characteristics.

[0039] Furthermore, the corresponding device characteristics can be written into the device's memory, specifically as corresponding calibration parameters. Additionally or alternatively, devices may be classified or rejected based on the corresponding device characteristics. Furthermore, it can be stipulated that device production be halted based on the corresponding device characteristics.

[0040] According to another aspect, a system is provided for determining device characteristics of a device manufactured on a wafer, wherein the system is configured to perform the following steps:

[0041] - Provides a wafer feature model that associates wafer location with device features, the wafer location indicating the location of a device manufactured on the wafer, wherein the wafer feature model is configured to be trained by one or more wafer feature maps, and is specifically configured as a Gaussian process model;

[0042] - Provide sample device features for at least one device at the sample wafer location;

[0043] - Determine the device characteristics of at least one other device of the wafer, depending on the provided wafer feature model.

[0044] According to another aspect, a system for manufacturing devices on a wafer is provided, wherein the system is configured to perform the following steps:

[0045] - Control the production of devices on one or more wafers;

[0046] - Provides a wafer feature model that associates wafer location with device features, the wafer location indicating the location of a device manufactured on the wafer, wherein the wafer feature model is configured to be trained by one or more wafer feature maps, and is specifically configured as a Gaussian process model;

[0047] - Provide sample device features for at least one device at the sample wafer location;

[0048] - Determine the device characteristics of at least one other device of the wafer, depending on the provided wafer feature model.

[0049] The system is further configured to control device production based on the determined device characteristics. Attached Figure Description

[0050] The embodiments are described in more detail with reference to the accompanying drawings, in which:

[0051] Figure 1 An exemplary arrangement of devices on a wafer is shown;

[0052] Figure 2 An exemplary wafer diagram showing device features of an example accelerometer sensor is provided, where the device features are calibration values; and

[0053] Figure 3 A flowchart illustrating a method for determining device characteristics using a device feature model is shown. Detailed Implementation

[0054] The invention is described in more detail using a manufacturing example of an accelerometer manufactured on a wafer comprising multiple identical sensor devices. Figure 1 An exemplary arrangement of such a device 2 on wafer 1 is shown.

[0055] The production of accelerometers requires numerous processing steps at the wafer level, such as photolithography, etching, and annealing, to apply processing parameters. Due to the characteristics of these processing steps and variations in the process, different patterns of device characteristics are exposed on the wafer. These device characteristics can include functional indications, error indications, indications of whether the device meets given specifications, and calibration parameters to be stored in the specific device.

[0056] For the same product on multiple chips, even with normal variations in process parameters, the patterns are similar.

[0057] For a given example of an accelerometer sensor, the functional output is the acceleration value determined according to the following formula. :

[0058]

[0059] in as well as These are the inherent constant offset and the temperature-dependent offset, respectively. as well as It is the inherent and temperature-dependent response to external stimuli (ACC, which is the actual applied acceleration), such as the acceleration that the sensor should detect. Additionally, calibration parameters... and This must be determined and written into the calibration memory of each device to enable the sensor's functional output. Basically corresponds to the true physical value of acceleration After the original chip is manufactured, the measurement process is used to determine calibration parameters so that the behavior of sensor device 2 meets the following criteria:

[0060] .

[0061] The actual acceleration applied and given sensor specification limits In this case.

[0062] Generally speaking, for non-calibration parameters (such as test results) or as device characteristics... The criteria for any property can be defined as:

[0063]

[0064] in Corresponding to the measured characteristic value, and Corresponding to the desired device characteristic values, and Corresponding to the specification limits.

[0065] exist Figure 2 The text describes features as exemplary devices (i.e.) An example of a wafer feature map, which is one of the calibration parameters.

[0066] In the following text, combined with Figure 3 The flowchart details a method for controlling device production on a wafer by determining device characteristics. This method can be executed on a data processing system with access to a database storing training data for constructing wafer feature models.

[0067] In step S1, training data for generating the wafer feature model is generated by providing multiple processed wafers, including accelerometers of the same type and variants. When the wafer may also include one or more corner batch wafers, standard processing steps are used to produce the wafer, which are wafers for which the manufacturing process has been intentionally slightly modified to simulate more extreme process variations.

[0068] A wafer feature map is constructed based on the obtained device features. The wafer feature map essentially associates the location of a device on the wafer with one or more device features, which in this example may include calibration parameters. Specifically, a separate wafer feature map is constructed for each device feature. Figure 2As shown, exemplary wafer feature maps (one for each device feature) form the basis for generating the corresponding wafer feature model as input to the training dataset. The training dataset associates the device's wafer location (x and y coordinates) on the wafer with optional environmental conditions (such as temperature, humidity, etc.) and operating conditions (such as the actual acceleration in a given example of an accelerometer) with the device features to be reflected by the wafer feature model.

[0069] The method is further described below regarding individual device characteristics.

[0070] In step S2, a wafer feature model is generated, trained from typical wafer feature maps with corresponding uncertainties, such that wafer location, environment, and operating conditions are associated with corresponding device features based on the wafer feature model. Furthermore, the wafer feature model is nonparametric and should further provide uncertainty values ​​for the predicted features. A preferred model type is a Gaussian process, which can be trained to output a smooth function, model uncertainty, and allows prior knowledge to be incorporated, for example, via a parametric mean function, such as by maximizing the logarithmic marginal likelihood. The Gaussian process can be trained using the dataset provided in step S1.

[0071] Instead of using a Gaussian process model, a multifidelity Gaussian process model can be used, where the original measured samples can be used directly, instead of the mean function applied to them.

[0072] One advantage of using Gaussian processes is that they automatically output the uncertainty of the predicted features of the modeling device.

[0073] Basically, for efficient chip testing / measurement, it is necessary to start from the assembly of all devices on the chip. Select a subset of devices with the same wafer. ( ), this subset of devices It indicates the specific device to be tested / measured to obtain the corresponding actual device characteristics, while for devices that are not tested / measured, device characteristics are inferred. A discrete subset corresponding to wafer location and optional environmental and additional conditions (such as, for a given example, the actual applied acceleration Acc and operating temperature). Subset The following formula should be used to maximize all samples on the wafer within the specification limits. Internal likelihood

[0074]

[0075] in

[0076] and It concerns each sample meeting the specification limits. The lower bound of the minimum probability.

[0077] This corresponds to the relevant feature in this example. This feature is a calibration parameter that must be identified for each device on the wafer, and it needs to be evaluated to see if it complies with the limits.

[0078] .

[0079] Then the violation can be specified as

[0080] .

[0081] Since the violation will be different for different electrical responses, it can be applied to all predicted values. The specified violation.

[0082] Depend on Marking .

[0083] Generally, this is achieved by maximizing all samples on the wafer within a given specification limit according to the following formula. The internal likelihood is used to select the discrete subset of wafer locations.

[0084]

[0085] in

[0086] and It concerns each sample meeting the specification limits. The lower bound of the minimum probability, where It is the set of all chip locations. It is a modeling feature, and These are the nominal characteristics that a device at a specific location i should satisfy.

[0087] Furthermore, the limit violation v can be specified as

[0088]

[0089] Selecting a subset of chip locations To minimize the probability of expected limit violations across all devices on the chip.

[0090] In the next step S3, for efficient wafer sample testing / measurement, the size is selected. subset of Selecting a subset To minimize the expected limit violation v across all devices on the wafer. Therefore, based on the Bayesian optimization algorithm on a subset of samples... The above specifies a function 'a' to acquire when new measurements are added to the Gaussian process and it is assumed that their corresponding values ​​are... When the values ​​are distributed according to the Gaussian process prediction, the sample subset will be after the expected improvement of the Gaussian process posterior. The violation of the limit v should be minimized.

[0091] For a given environmental and / or operating condition (acceleration and temperature in this example), the Gaussian process prediction for wafer device characteristics is as follows. ,in It is the actual subset that has been obtained so far. .

[0092] Gaussian process prediction based on Using sample location New measurements at the location Perform an update. The function used to retrieve this information is denoted as:

[0093]

[0094] In order to select an optimized subset of sample devices The acquisition function 'a' must be evaluated for all possible combinations of sample devices (indicated by their wafer location), where a subset with the lowest expected limit violation is selected. Since the Gaussian process can predict the characteristics of untested / unmeasured devices, the test / measurement time can be significantly reduced compared to the test / measurement time for all devices on the wafer.

[0095] In the next step S4, for the subset The device samples are used to determine the characteristics. The wafer feature model is updated, and the expected limit violations across the wafer can be re-evaluated. If the results are appropriate, testing can be stopped; otherwise, the newly measured device features of this subset can be added to the existing subset. And the process can continue with step S3.

[0096] In step S5, device production on wafer 1 is controlled according to device characteristics. If the device characteristic is a calibration parameter, it can be stored in the calibration setting memory of the corresponding device.

[0097] Furthermore, device characteristics may include test information, such as electrical parameters. Depending on the test information, corresponding devices associated with device characteristics can be classified or rejected. For example, a rejection threshold can be defined, where a device is rejected when the test information exceeds the rejection threshold.

[0098] Furthermore, production control may include halting device production based on the corresponding device characteristics. Device characteristics may include test information. If the aggregated test information of device 2 on wafer 1 exceeds a given interruption threshold, production can be stopped.

[0099] Furthermore, it is possible to achieve prior knowledge of the manufacturing steps using processing equipment, such as the radial or angular correlation / etching loss of known trench angles, to realize the device feature pattern of the corresponding wafer feature map. This allows prior knowledge to be incorporated into the creation of the wafer feature model.

[0100] Additionally, active learning allows for the identification of which wafer locations sample portions can be placed to achieve low-uncertainty predictions by minimizing the number of sample portions. To achieve this, active learning iteratively selects the sample currently with the highest predictive uncertainty and updates the model after measuring it.

Claims

1. Computer-implemented method for determining a device feature y of a device (2) produced on a wafer (1); comprising the following steps: - providing (S2) a wafer feature model associating wafer positions to device features y, the wafer positions indicating positions of devices (2) produced on a wafer (1), wherein the wafer feature model is configured to be trained by one or more wafer feature maps and is configured as a Gaussian process model; - providing (S3) sample device features of at least one device (2) at sample wafer positions; - determining (S4) a device feature y of at least one other device of the wafer (1) depending on the provided wafer feature model, wherein the sample device feature y of the at least one device is obtained at a selected discrete subset x of wafer locations samp , the selected discrete subset x of wafer locations samp is selected by means of active learning and maximizes the likelihood that all samples on the wafer are inside the given specification limit.

2. The method according to claim 1, wherein the device feature y comprises one of: an indication of functionality, an indication of error, an indication whether a device meets a given specification, and a calibration parameter to be stored into a particular device.

3. The method according to claim 1 or 2, wherein the wafer feature model additionally associates environmental and / or operating conditions to the device feature y.

4. The method of any one of claims 1 to 2, wherein, The wafer feature model is trained by providing a plurality of processed wafers comprising identical devices (2), wherein the processed wafers comprise one or more corner lot wafers.

5. The method according to any one of claims 1 to 2, wherein the wafer feature model is non-parametric and provides an uncertainty value for the predicted feature y.

6. The method of any of claims 1-2, wherein the selected discrete subset of wafer positions x is selected by maximizing the likelihood of all samples on the wafer being within a given specification limit s according to the formula lim samp ​ p(d(y)≤s lim )≥p min wherein And p min is a lower bound on the minimum probability that each sample meets a specification limit s lim , where x is a set of all wafer positions, y mod is a modeled feature, and y des is a nominal feature that should be met for a device at a particular position i.

7. The method according to claim 6, wherein the limit violation v can be specified as where a subset of wafer locations x is selected samp to minimize the probability of expected extreme violations over all devices on the wafer, and where s acc is a specification limit.

8. The method according to claim 7, wherein the subset of wafer positions is updated by means of an acquisition function: where x is the optimized subset of wafer positions samp Selecting wafer positions, evaluating the acquisition function for all possible combinations of wafer positions, where the subset of wafer positions x with the lowest expected limit violation is selected samp where the Gaussian process prediction of wafer device features is where is the actual subset x that has been acquired so far samp , y samp is a new measurement at the subset of wafer positions x samp .

9. The method of claim 6, wherein, The device feature model is provided by using a prior for the corresponding pattern based on prior technical knowledge of the manufacturing steps using the processing equipment.

10. Computer-implemented method for device production on a wafer, comprising the following steps: - controlling (S5) the production of devices (2) on one or more wafers (1); - determination of a device feature y of a device by means of the method according to any one of claims 1 to 9, wherein the production of devices (2) is controlled depending on the determined device feature y.

11. The method of claim 10, wherein, The production of devices (2) is controlled by one of: - the respective device feature y is written into a memory of the device (2) as a respective calibration parameter, - devices are classified or rejected depending on the respective device feature y, and - the production of devices is stopped depending on the respective device feature y of a device.

12. System for determining a device feature y of a device (2) produced on a wafer (1); wherein the system is configured to perform the following steps: - providing a wafer feature model associating wafer positions to device features y, the wafer positions indicating positions of devices produced on a wafer, wherein the wafer feature model is configured to be trained by one or more wafer feature maps and is configured as a Gaussian process model; - providing sample device features y of at least one device at sample wafer positions; - determining a device feature y of at least one other device of the wafer (1) depending on the provided wafer feature model, wherein The sample device features y of the at least one device are obtained at a selected discrete subset x of wafer locations samp The selected discrete subset x of wafer locations is selected by means of active learning and maximizes the likelihood that all samples on the wafer are within the given specification limit. samp ​ 13. System for production of devices (2) on a wafer (1), wherein the system is configured to perform the following steps: - controlling the production of devices (2) on one or more wafers (1); - determination of a device feature y of a device by means of the method according to any one of claims 1 to 9, wherein the production of devices (2) is controlled depending on the determined device feature y. - providing a wafer feature model associating wafer positions with device feature y, the wafer positions being indicative of positions of devices produced on a wafer (1), wherein the wafer feature model is configured to be trained by one or more wafer feature maps, and is configured as a Gaussian process model; - providing a sample device feature y of at least one device at a sample wafer position; - determining a device feature y of at least one other device of the wafer (1) depending on the provided wafer feature model, wherein a sample device feature y of the at least one device is obtained at a selected discrete subset x of wafer locations samp a selected discrete subset x of wafer locations samp is selected by means of active learning and maximizes the likelihood that all samples on the wafer are inside the given specification limit, wherein the system is further configured to control production of devices (2) in dependence on the determined device feature.

14. A computer program product comprising instructions which, when executed by a computer, cause the computer to carry out the steps of the method of any one of claims 1 to 9.

15. A machine-readable storage medium comprising instructions which, when executed by a data processing unit, cause the data processing unit to carry out the steps of the method of any one of claims 1 to 9.

Images