Maintenance of modules for light sources in semiconductor photolithography

Abstract

The light source includes one or more modules configured together to generate a light beam for semiconductor photolithography. The apparatus for maintaining the light source comprises a prediction unit and an aggregation unit. The prediction unit is configured to receive a plurality of partial characteristic data sets, each partial characteristic data set being associated with a specific performance criterion of the light source, and to evaluate, for each partial characteristic data set, whether a failure mode is detected in the light source in a prediction increment. The aggregation unit is configured to receive the plurality of evaluations from the prediction unit and to determine, for each failure mode, which one or more performance criteria are related to the failure mode based on the associated partial characteristic data set of the failure mode.

Description

[Technical field]

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application claims priority to U.S. application 63 / 293,453, filed December 23, 2021, entitled "Module Maintenance for Light Sources in Semiconductor Photolithography," which is incorporated by reference in its entirety into this specification.

[0002] The disclosed subject matter relates to maintenance of light sources such as those used in integrated circuit photolithography manufacturing processes. [Background technology]

[0003]

[0003] Light, such as laser radiation used in semiconductor photolithography, is typically supplied by a system called a light source. These light sources generate radiation as a train of pulses at a predefined repetition rate, for example, in the range of about 500 Hz to about 6 kHz. Moreover, such light sources have a predicted useful life, conventionally measured as the number of pulses that can be projected and generated before repair or replacement is required, typically expressed in billions of pulses.

[0004]

[0004] One system for generating light such as laser radiation at frequencies useful for semiconductor photolithography (e.g., at deep ultraviolet (DUV) wavelengths) uses a master oscillator power amplifier (MOPA) dual gas discharge chamber configuration. This configuration has two chambers: a master oscillator chamber (MO chamber) and a power amplifier chamber (PA chamber). These chambers and many other system components can be considered modules, and the entire light source can be considered a collection of multiple modules. Each module typically has a lifetime that is shorter than the lifetime of the entire light source. Thus, during the lifetime of the light source, the health of the individual modules is evaluated to determine whether they should be repaired or replaced. Summary of the Invention

[0005]

[0005] In some general aspects, an apparatus maintains a light source including one or more modules configured together to generate a light beam for semiconductor photolithography. The apparatus includes a prediction unit and an aggregation unit. The prediction unit is configured to receive a plurality of partial feature data sets, each partial feature data set being associated with a specific performance criterion of the light source, and to evaluate, for each partial feature data set, whether a failure mode is detected in the light source in a prediction increment. The aggregation unit is configured to receive a plurality of evaluations from the prediction unit and to determine, for each failure mode, which one or more performance criteria are associated with the failure mode based on the associated partial feature data set of the failure mode.

[0006]

[0006] Implementations may include one or more of the following features: For example, the prediction unit may include or have access to multiple models, each model generated by machine learning using a unique partial feature dataset; Each of the models in the prediction unit may have the same adaptation structure; Each of the models in the prediction unit may have a different adaptation structure; At least one of the models in the prediction unit may have a different adaptation structure than a adaptation structure of another model in the prediction unit.

[0007]

[0007] The plurality of partial feature data sets may include a partial feature data set for a bandwidth of the light beam, a partial feature data set for a wavelength of the light beam, and a partial feature data set for an energy of the light beam. The plurality of partial feature data sets may include a partial feature data set for a long-term bandwidth of the light beam, a partial feature data set for a short-term bandwidth of the light beam, a partial feature data set for a bandwidth error event, a partial feature data set for a long-term wavelength of the light beam, a partial feature data set for a short-term wavelength of the light beam, a partial feature data set for a wavelength error event, a partial feature data set for a long-term energy of the light beam, a partial feature data set for a short-term energy of the light beam, and a partial feature data set for an energy error event.

[0008]

[0008] The aggregation unit may be configured to determine which one or more performance criteria are related to a failure mode based on the associated partial feature data set by aggregating the evaluations from the prediction units and prioritizing the failure reasons associated with each detected failure mode. The aggregation unit may be configured to aggregate the evaluations from the prediction units based on the performance criteria associated with each partial feature data set. The aggregation unit may be configured to label a performance criterion for further evaluation if at least one failure mode associated with a performance criterion is detected. The aggregation unit may be configured to label a performance criterion for further evaluation if at least one failure mode associated with a performance criterion is detected, regardless of whether another failure mode associated with another performance criterion is detected.

[0009]

[0009] The prediction unit being configured to evaluate whether a failure mode is detected in the light source at the prediction increment may include evaluating whether a failure mode is detected in a specific module of the light source based on a model trained for the specific module. The specific module of the light source may include a master oscillator module. The prediction unit being configured to evaluate whether a failure mode is detected in the light source at the prediction increment may include evaluating whether a failure mode is detected in one or more specific modules of the light source based on a plurality of models trained for the specific module, respectively. The one or more specific modules of the light source may include one or more of a master oscillator module, a power amplifier module, a line narrowing module, a spectral feature analysis module, and a pulse stretcher module. The prediction unit being configured to evaluate whether a failure mode is detected in the light source at the prediction increment may include evaluating each partial feature dataset separately according to a model generated by machine learning using the partial feature dataset, and outputting a score on the probability of failure for each partial feature dataset evaluation. The score may correspond to a certainty score, a binary output, or a function ranging from 0 to 1.

[0010]

[0010] The prediction unit may be configured to evaluate, for each partial feature data set, whether a failure mode is detected in the light source at a prediction increment by performing an evaluation after each evaluation increment. The evaluation increment may be 100 million pulses.

[0011]

[0011] In another general aspect, an apparatus maintains a light source including one or more modules configured together to generate a light beam for semiconductor photolithography. The apparatus includes a pipeline unit, a prediction unit, and an aggregation unit. The pipeline unit is configured to receive a plurality of partial feature data sets, each partial feature data set being associated with a specific performance criterion of the operating light source, and to classify each of the received partial feature data sets into a feature category based on the associated specific performance criterion. The prediction unit is configured to receive the plurality of partial feature data sets and an assigned category for each partial feature data set, and to evaluate whether a failure mode is detected in the light source in a prediction increment. The aggregation unit is configured to determine which one or more performance criteria are associated with the failure mode based on the assigned category of each partial feature data set.

[0012]

[0012] The implementation may include one or more of the following features. For example, the prediction unit configured to evaluate whether a failure mode is detected in the light source in the prediction increment may include evaluating whether a failure mode is detected in a particular module of the light source based on a plurality of models trained for the particular module. Each model may be grown by machine learning by supplying a respective partial feature dataset to train the model. The apparatus may also include an alert generation unit configured to prescribe a maintenance action for the particular module if the evaluation determines that a failure mode is detected in the particular module.

[0013]

[0013] The prediction unit being configured to evaluate whether a failure mode is detected in the light source in a prediction increment may include evaluating each partial feature data set separately according to a model generated by machine learning using the partial feature data set, and outputting a score on the probability of failure for each partial feature data set evaluation. The aggregation unit being configured to determine which one or more performance criteria are associated with the failure mode based on the assigned category of each partial feature data set may include analyzing scores from partial feature data sets associated with the same assigned category. The score may be a certainty score, a binary output, or a function ranging from 0 to 1.

[0014]

[0014] The prediction unit being configured to evaluate whether a failure mode is detected in the light source at a prediction increment can include determining whether at least one module in the light source does not have at least a minimum probability of operating without failure at a prediction increment. The prediction increment can be measured as a number of pulses of the light beam.

[0015]

[0015] Prior to each evaluation, the pipeline unit may be further configured to aggregate data of at least one partial feature data set over one or more lookback increments preceding the evaluation in use. The prediction unit configured to evaluate whether a failure mode is detected in the light source in the prediction increment may include evaluating each partial feature data set separately according to a model generated by machine learning using the partial feature data set, and using the aggregated partial feature data set when the partial feature data set is aggregated over one or more lookback increments in use. Each lookback increment may be selected from a set of lookback increments assumed to be related to how the performance criterion changes as use progresses. The set of assumed lookback increments may include a first lookback increment of about 2 billion pulses, a second lookback increment of about 1 billion pulses, and a third lookback increment of about 100 million pulses. One or more of the partial feature data sets may be associated with a performance metric that is tracked at relatively short lookback increments, and one or more of the partial feature data sets may be associated with a performance metric that is tracked at relatively long lookback increments.

[0016]

[0016] In use, the pipeline unit may be configured to aggregate one or more partial feature datasets based on the same lookback increment, and the prediction unit may be configured to receive multiple partial feature datasets by receiving an aggregation of partial feature datasets that are tracked using the same lookback increment.

[0017]

[0017] In another general aspect, a method is performed for maintaining a light source including one or more modules configured together to generate a light beam for semiconductor photolithography, the method including receiving a plurality of partial feature data sets, each partial feature data set being associated with a specific performance criterion of the light source, evaluating for each partial feature data set whether a failure mode is detected in the light source at the predicted increment, and determining for each detected failure mode which one or more performance criteria are associated with the failure mode based on the failure mode's associated partial feature data set.

[0018]

[0018] Implementations may include one or more of the following features. For example, the method may further include, after receiving the plurality of partial feature data sets, classifying each of the received partial feature data sets into a feature category based on an associated specific performance criterion. And, determining which one or more performance criteria are associated with the failure mode may include determining which one or more performance criteria are associated with the failure mode based on the assigned category of each partial feature data set.

[0019]

[0019] In a prediction increment, evaluating whether a failure mode is detected in the light source may include evaluating each partial feature data set separately according to a model generated by machine learning using the partial feature data set, and outputting a score for each partial feature data set evaluation on a probability of failure. In a prediction increment, evaluating whether a failure mode is detected in the light source may include performing an evaluation after each evaluation increment.

[0020]

[0020] Determining which one or more performance criteria are associated with a failure mode based on the associated partial feature data set of the failure mode may include aggregating the evaluation and prioritizing the failure reasons associated with each detected failure mode. The evaluation may be aggregated by aggregating the evaluation based on the performance criteria associated with each partial feature data set. The method may further include labeling the performance criterion for further evaluation if at least one failure mode associated with the performance criterion is detected. The method may also further include labeling the performance criterion for further evaluation if at least one failure mode associated with the performance criterion is detected, regardless of whether another failure mode associated with another performance criterion is detected.

[0021]

[0021] In a prediction increment, evaluating whether a failure mode is detected in the light source may include evaluating whether a failure mode is detected in a particular module of the light source based on a model trained for the particular module. In a prediction increment, evaluating whether a failure mode is detected in the light source may include evaluating whether a failure mode is detected in one or more particular modules of the light source based on a plurality of models each trained for the particular module.

[0022]

[0022] The method may further comprise, prior to each evaluation, aggregating, in use, data of at least one partial feature dataset across one or more look-back increments preceding the evaluation. In use, aggregating data of at least one partial feature dataset across one or more look-back increments may include, in use, aggregating one or more partial feature datasets based on a same look-back increment, and receiving a plurality of partial feature datasets may include receiving an aggregation of partial feature datasets tracked using a same look-back increment.

[0023]

[0023] In another general aspect, a method is performed for maintaining a light source including one or more modules configured together to generate a light beam for semiconductor photolithography. The method includes receiving a plurality of partial feature data sets, each partial feature data set being associated with a specific performance criterion of the light source, evaluating for each partial feature data set whether a failure mode is detected in the light source at the prediction increment, including evaluating each partial feature data set separately according to a model generated by machine learning using the partial feature data set, and determining, for each detected failure mode, which one or more performance criteria are associated with the failure mode based on the failure mode's associated partial feature data set. The determining includes aggregating the evaluations and prioritizing the failure reasons associated with each detected failure mode.

[0024]

[0024] Implementations may include one or more of the following features. For example, the evaluations may be aggregated by aggregating the evaluations based on a performance criterion associated with each partial feature data set. The method may further include, if at least one failure mode associated with a performance criterion is detected, labeling the performance criterion for further evaluation, regardless of whether another failure mode associated with another performance criterion is detected. [Brief description of the drawings]

[0025]

[0025] The accompanying drawings, which are incorporated in and form a part of this specification, illustrate the invention and, together with the description, serve to further explain the principles of the invention and to enable one of ordinary skill in the relevant art to make and use the invention. One or more of the drawings may not be to scale.

[0026] [Figure 1A]

[0026] A block diagram of a maintenance device configured to maintain a module of a light source that generates a light beam for use by an output device and provide advance notification of module failure, the maintenance device including a prediction unit and an aggregation unit. [Figure 1B]

[0027] FIG. 2 is a block diagram showing an implementation of a maintenance device including a pipeline unit. [Figure 2A]

[0028] FIG. 1C is a block diagram of a training pipeline illustrating the deployment of machine learning models accessed and used by the maintenance device of FIGS. 1A and 1B. [Figure 2B]

[0029] FIG. 2 is a block diagram of an operational pipeline illustrating how machine learning models are accessed by a prediction unit during operation of the light source of FIGS. 1A and 1B. [Figure 3A]

[0030] FIG. 1 illustrates a maintenance procedure in terms of use according to some implementations. [Figure 3B]

[0031] FIG. 13 illustrates an example of three lookback increments in evaluation points for three partial feature datasets according to some implementations. [Figure 4]

[0032] FIG. 1B is a block diagram illustrating an implementation of the light source of FIG. 1A. [Diagram 5]

[0033] FIG. 1B is a block diagram illustrating an implementation of the output device of FIG. 1A. [Figure 6]

[0034] 1C is a flow chart showing a procedure performed by the maintenance device of FIG. 1A or 1B to maintain a module of a light source. [Figure 7A]

[0035] 1 is a table showing operational data used by a maintenance device that has been parsed into nine partial feature datasets, each of which is associated with a specific performance metric of a light source and associated with a machine learning model, according to some implementations. [Figure 7B]

[0036] 7B is a table showing the set of ratings obtained at the evaluation points for each machine learning model in FIG. 7A . DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0027]

[0037] Referring to FIG. 1A, light source 100 generates an amplified light beam 105 for use by output device 110, which may be a photolithography device. Light source 100 includes a plurality of modules 115-m, where m is a set of numbers from 1 to a positive integer greater than or equal to 2. In the example shown in FIG. 1A, m=1, 2, 3, 4, 5, and the plurality of modules 115-m includes modules 115-1, 115-2, 115-3, 115-4, 115-5, and light source 100 as a whole can be considered a collection of such modules 115-1, 115-2, 115-3, 115-4, 115-5. Although five modules (i.e., 115-1, 115-2, 115-3, 115-4, 115-5) are shown in FIG. 1A, light source 100 may include fewer than five modules or may include more than five modules. Each module 115-m typically has a lifetime that is shorter than the lifetime of the entire light source 100. Thus, during the lifetime of the light source 100, the health of the individual modules 115-m is evaluated to determine whether any of the modules 115-m should be repaired or replaced. The timing of maintenance of the modules 115-m is determined according to a maintenance policy implemented by the maintenance device 120.

[0028]

[0038] In an example in which light source 100 is a dual stage deep ultraviolet light source (as described below with reference to FIG. 4), modules 115-m include a master oscillator (MO) chamber module, a power amplifier (PA) chamber module, a line narrowing module (LNM), a line analysis module (LAM), a bandwidth analysis module (BAM), and an optical pulse stretcher (OPuS) module. In general, any module for which data regarding the performance of the module is collected can be considered a module 115-m.

[0029]

[0039] The maintenance policy implemented by the maintenance device 120 is a predictive maintenance policy designed to monitor the condition of the equipment in service (i.e., modules 115-m) and predict when and which modules 115-m will fail. The future behavior / condition of the machine components is approximated, allowing the maintenance tasks to be optimized (e.g., predictive health monitoring). Thus, machine downtime and maintenance costs can be significantly reduced while ensuring the least frequent maintenance possible. The maintenance policy implemented by the maintenance device utilizes predictive tools based on historical data to allow advance detection of impending failures and allows timely pre-failure intervention.

[0030]

[0040] The maintenance device 120 employs machine learning (ML) models that utilize rational, logical, and physics-based datasets that together form the overall feature dataset 140 to derive information about the causality of failures and thus help determine which modules will fail in the future. Each dataset in the feature dataset 140 is referred to as a partial feature dataset 140-sf, where sf represents a specific measure of performance (also called performance criterion) of the light source 100. Thus, in the example of FIG. 1A, there are eight partial feature datasets 140-1, 140-2, 140-3, ... 140-8. The feature dataset 140 may include fewer or more than eight partial feature datasets. Each partial feature dataset 140-sf targets a specific aspect of the performance (i.e., a specific performance criterion) of the light source 100. In other words, each partial feature dataset 140-sf is curated to model a specific aspect of the light source 100. Each partial feature data set 140-sf corresponds to a set of data associated with the operation of the light source 100. Furthermore, each machine learning model is based on a partial feature training set generated using machine learning and corresponding to a particular partial feature data set 140-sf. Thus, each model is separately trained with a unique partial feature training set corresponding to a particular partial feature data set 140-sf that is fed into the particular model to evaluate the light source 100 during normal operation of the light source 100 (i.e., while the light source 100 is generating the amplified light beam 105 used by the output device 110).

[0031]

[0041] During normal operation, the maintenance device 120 evaluates or calculates a probability for each model and can use this information to determine whether a particular module 115-m in the light source 100 will fail in the future. The maintenance device 120 can make this determination for a particular module 115-m due to a pipeline formed between each performance aspect of the light source and each evaluation via the partial feature training set used to train the model. In this way, the maintenance device 120 can not only determine that a failure will occur somewhere in the light source 100 in the future, but also indicate how the light source 100 should be adjusted to prevent that failure, and in particular can target which module 115-m needs action. The maintenance device 120 can pinpoint problems in the light source 100 by relying on a pipeline in which each machine learning model is trained on a unique partial feature training set, and by using the machine learning model on the partial feature data set 140-sf corresponding to the unique partial feature training set used to train the model.

[0032]

[0042] During operation, some modules 115-m may affect certain performance aspects of the light source 100 more than other modules 115-m. For example, using the example of a dual stage deep ultraviolet light source 400 (as described below with reference to FIG. 4), the operation of the MO chamber module 461 affects some performance aspects of the light source 100, such as the bandwidth, wavelength, and energy of the amplified light beam 105. This is because the MO chamber module 461 is the first module in the light source 400 to generate light energy, and other modules in the light source are optically downstream from the MO chamber module 461. For example, if the amplified light beam 105 output from the light source 400 is not generating enough light energy and the output light beam from the MO chamber module 461 has an energy below a threshold energy, the MO chamber module 461 may be considered not to be operating at an optimal or acceptable level. However, if the output light beam from the MO chamber module 461 and the output light beam from the PRA chamber module 466 both have acceptable energy levels, but the amplified light beam 405 output from the light source 400 is not producing enough energy, then it can be assumed that there is a problem with the module (optically) following the PRA chamber module 466 (such as one of modules 476, 477, or 478). On the other hand, the operation of the OpuS module 477 has only a negligible effect on the energy of the amplified light beam 105.

[0033]

[0043] To this end, a particular ML model can be trained for a particular module 115-m of light source 100. For example, an ML module associated with wavelength can be trained for MO chamber module 461 (FIG. 4). As another example, a model ML associated with bandwidth can be trained for MO chamber module 461 (FIG. 4). And an ML model associated with energy (of amplified light beam 105) can be trained for MO chamber module 461 (FIG. 4).

[0034]

[0044] The maintenance device 120 uses multiple machine learning models to evaluate the light source 100, as opposed to a single machine learning model. Furthermore, as described above, a pipeline is created between the training of each machine learning model and the use of the machine learning model after each training. In this way, the average prediction accuracy (in the evaluation) is improved and the performance metrics required by the output device 110 can be better met. Furthermore, the use of multiple machine learning models can reduce bias against key features and results, without the need to narrow the feature selection, modify the underlying data, or parameter model to capture less prominent features. In addition, the use of multiple machine learning models in this way reduces errors in the evaluation and reduces variability when new data is evaluated by the maintenance device 120.

[0035]

[0045] The maintenance device 120 includes, among other things, a prediction unit 122 and an aggregation unit 124. The prediction unit 122 has access to a number of machine learning models (either stored in the prediction unit 122 or stored within the maintenance device 120). The prediction unit 122 evaluates each of the partial feature data sets 140-sf. Specifically, the prediction unit 122 evaluates a particular partial feature data set 140-sf using a machine learning model trained based on a partial feature training set corresponding to the particular partial feature data set 140-sf. The aggregation unit 124 receives the evaluations from the prediction unit 122 and interprets these evaluations to aggregate the underlying model predictions and prioritize the failure reasons to determine the root cause of the failure.

[0036]

[0046] The maintenance apparatus 120 includes one or more of digital electronic circuitry, computer hardware, firmware, and software. The maintenance apparatus 120 also includes memory 128, which may be read-only memory and / or random access memory. Suitable storage devices for tangibly embodying computer program instructions and data include all forms of non-volatile memory, including, by way of example, semiconductor memory devices such as EPROM, EEPROM, and flash memory devices, magnetic disks such as internal hard disks and removable disks, magneto-optical disks, and CD-ROM disks. The maintenance apparatus 120 may also include one or more input devices 121 (such as a keyboard, touch screen, microphone, mouse, handheld input device, etc.) and one or more output devices 123 (such as a speaker or monitor). The maintenance apparatus may also include components that enable wireless communication, including Bluetooth, NFC, and Wi-Fi. The maintenance apparatus 120 may also include one or more programmable processors and one or more computer program products tangibly embodied in machine-readable storage devices for execution by the one or more programmable processors. The one or more programmable processors may each execute a program of instructions to operate on input data and generate appropriate output to perform a desired function. Typically, the processor receives instructions and data from a memory. Any of the above may be supplemented by or incorporated into specially designed ASICs (Application Specific Integrated Circuits). The maintenance device 120 may be centralized throughout the light source 100 or may be partially or wholly distributed and may communicate with other controllers for controlling other aspects of the light source 100.

[0037]

[0047] 1B, an implementation 120B of the maintenance device 120 includes a pipeline unit 126. The pipeline unit 126 is configured to receive the feature set 140 (including a plurality of partial feature data sets 140-sf) and classify each partial feature data set 140-sf into a feature category based on a specific performance criterion associated with each partial feature data set 140-sf. For example, if the performance criterion is the long-term bandwidth of the amplified light beam 105, the feature category is the bandwidth of the amplified light beam 105, if the performance criterion is the short-term wavelength of the amplified light beam 105, the feature category is the wavelength of the amplified light beam 105, if the performance criterion is an error in the energy of the amplified light beam 105, the feature category is the energy of the amplified light beam 105.

[0038]

[0048] The pipeline unit 126 communicates the partial feature data set 140-sf with their associated feature categories to the prediction unit 122. The prediction unit 122, the aggregation unit 124, and the pipeline unit 126 can each communicate with each other. The prediction unit 122, the aggregation unit 124, and the pipeline unit 126 can each access a memory stored in the maintenance device 120B. Furthermore, the prediction unit 122, the aggregation unit 124, and the pipeline unit 126 can each include one or more of digital electronic circuitry, computer hardware, firmware, and software. The prediction unit 122, the aggregation unit 124, and the pipeline unit 126 can also each include one or more programmable processors and one or more computer program products tangibly embodied in a machine-readable storage device for execution by the one or more programmable processors, and the one or more programmable processors can each execute a program of instructions to operate on input data and generate appropriate outputs to perform a desired function. Generally, these processors receive instructions and data from a memory 128 .

[0039]

[0049] The memory accessed by units 122, 124, 126 may use multiple machine learning models and may store training data sets (sub-feature training sets).

[0040]

[0050] The maintenance apparatus 120B may also include an alert generation unit 125 configured to prescribe a maintenance action for a particular module 115-m when the aggregation unit 124 determines that one or more assessments from the prediction unit 122 are relevant to that particular module 115-m.

[0041]

[0051] A training pipeline 250A is shown in FIG. 2A, and an operational pipeline 250B is shown in FIG. 2B. The training pipeline 250A illustrates the development of an ML model that is accessed and used by the maintenance machine 120 to provide advance notification of module failures for modules 115-m currently installed in the light source 100. The training pipeline 250A includes training data 251A acquired during data acquisition by a metrology machine, where the metrology machine includes a number of metrology units, each configured to acquire information about a previous light source 100P. For example, many instances of the same module were installed and subsequently removed in a previous light source 100P that has been deployed in semiconductor manufacturing plants worldwide over a period of time, and data collected during the life of each individual module is stored centrally. The historical data of these removed modules can be used as the data 251A.

[0042]

[0052] In general, each ML model (ML-sf in block 254A) is the output of a machine learning algorithm executed on the training data 251A and represents what has been learned by the machine learning algorithm. The prediction unit 122 can execute these ML algorithms, or another unit separate from the maintenance device 120 can be used to execute these ML algorithms for training. Each ML model ML-sf is stored in the memory 128 after execution of the machine learning algorithm on the data 251A. The ML model ML-sf represents an algorithm-specific data structure (which may include rules and / or numbers) that the maintenance device 120 needs to eventually make a prediction (operation pipeline 250B) about the light source 100 during operation to generate the amplified light beam 105. For example, the data structure of the ML model ML-sf may consist of a vector or matrix of coefficients with specific values, a tree of if-then statements with specific values, or a graph structure with a vector or matrix of weights with specific values. Each of these types of ML models ML-sf has its own adaptation data structure that these ML models ML-sf adapt based on the algorithm used in training. Each ML model ML-sf stored in the memory 128 contains both data and instructions for using this data to recognize certain types of patterns and make predictions. Once trained, each ML model ML-sf is stored in the memory 128, and the maintenance apparatus 120 (in the operation pipeline 250B) uses each ML model ML-sf to make inferences about the data (feature set 140) that have not been seen before and to make predictions about those data (feature set 140). Each ML model ML-sf contains both the fitting data structure and instructions on how to use this data, so that the maintenance apparatus 120 can make predictions about new data (for each partial feature data set 140-sf).

[0043]

[0053] Suitable ML models ML-sf that can be trained include Random Forest, AdaBoost, Support Vector Machine (SVM), and Artificial Neural Network (ANN). Other ML models ML-sf may be used. In some implementations, each and every ML model ML-sf stored in memory 128 and used in operational pipeline 250B has the same adaptation data structure. In this case, for example, all ML models ML-sf may be Random Forest models. In other implementations, one or more of the ML models ML-sf stored in memory 128 and used in operational pipeline 250B may have a unique adaptation data structure or a different adaptation data structure than other ML models ML-sf stored in memory 128. Thus, for example, ML model ML-1 may be a Random Forest model, while models ML-2 and ML-3 are ANN models.

[0044]

[0054] Also, in some implementations, the trained ML model ML-sf uses many examples of all modules in the population of removed modules. In a scheme where modules are evaluated in increments of 100 million pulses starting with 10 billion light beam pulses until the end of life, for example, a module that fails at 32.55 billion pulses will yield 226 examples (of data) that can be used to train the model. Thus, from a population of hundreds of modules, a training data set with tens of thousands of examples can be created, which is favorable from a machine learning perspective in the sense that it can be implemented using many different machine learning models. It is also favorable in that the accuracy is improved for larger training data sets.

[0045]

[0055] The raw training data 251A may be pre-processed into a form from which condition indicators can be extracted. Block 252A thus corresponds to feature identification or extraction, including identification of features (or metrics) that change in predictable ways as modules age and deteriorate.

[0046]

[0056] Block 253A corresponds to the identified partial feature training sets 241-sf (where sf represents each unique performance metric of the previous light source 100P, and sf has an integer value from 1 to N). Each partial feature training set 241-sf is used to train a respective machine learning model ML-sf, which provides a predicted increment estimate. The set of machine learning models ML-sf is represented by block 254A. In operational use (such as in operational pipeline 250B), each machine learning model ML-sf predicts an outcome based on the model prepared and trained on this past or historical input data (in training pipeline 250A) and its output behavior (partial feature training set 241-sf). Here, training involves aligning the estimates with real-world data, as described below.

[0047]

[0057] Each partial feature training set 241-sf targets a particular performance aspect (a unique performance criterion) of the previous light source 100P. Each partial feature training set 241-sf is curated to model a particular aspect of the previous light source 100P. Examples of partial feature training sets 241-sf are described below.

[0048]

[0058] In one example, one particular performance criterion of the previous light source 100P is the long-term bandwidth of the amplified light beam 105 generated by the previous light source 100P. The bandwidth of the amplified light beam 105 is the width of the intensity spectrum of the light beam 105, which can be expressed in terms of the wavelength or frequency of the light beam 105. Any suitable mathematical construct (i.e., metric) related to the spectral details of the light beam 105 can be used to estimate the bandwidth of the light beam 105. For example, the full width of the spectrum at a fraction of the maximum peak intensity (X), called FWXM, can be used to estimate the light beam bandwidth. As another example, the width of the spectrum with a fraction of the integrated spectral intensity (Y), called EY, can be used to estimate the light beam bandwidth. Other metrics for estimating the bandwidth are also suitable. Furthermore, the bandwidth is measured over a "long-term" scale of use. This means that each of the metrics is evaluated looking back over a longer period of use (and then averaged by any suitable calculation). For example, the long-term bandwidth may be measured or estimated using one particular mathematical construct based on a retrospective evaluation of a large number (e.g., 2 billion) of pulses of the light beam 105 generated by the previous light source 100P. Thus, one sub-feature training set 241-sf (e.g., training set 241-1) may include a set of metrics (set [EY lt , F.W.X.M. lt , ...ME lt ]), where ME is any other metric that can be used, and "lt" represents a retrospective evaluation on a long-term usage scale.

[0049]

[0059] As another example, another performance criterion of the previous light source 100P is the short-term bandwidth of the amplified light beam 105 generated by the previous light source 100P. In this particular criterion, the bandwidth is measured on a "short-term" scale of use. This means that each of the metrics is evaluated by looking back over a shorter period of use (and then averaged by any suitable calculation). For example, the short-term bandwidth can be measured or estimated using one particular mathematical construction based on a look-back evaluation of a small number (e.g., 100 million) of pulses of the light beam 105 generated by the previous light source 100P. Thus, one partial feature training set 241-sf (e.g., training set 241-2) includes a set of metrics (set [EY st , F.W.X.M. st , ...ME st ]), where ME is any other metric that can be used, and "st" represents a retrospective evaluation on a short-term usage scale.

[0050]

[0060] Another performance criterion envisaged for the previous light source 100P for which a partial feature training set 241-sf (such as partial feature training set 241-3) can be formed includes data on the number of error events (BQ) related to bandwidth, which can be measured on a short-term or long-term usage scale.

[0051]

[0061] Other possible performance criteria of the previous light source 100P from which the respective partial feature training set 241-sf (such as those identified as partial feature training sets 241-4, 241-5, 241-6, not explicitly shown in FIG. 2A) can be formed include data on long-term energy (of the amplified light beam 105 output from the previous light source 100P), short-term energy (of the amplified light beam 105 output from the previous light source 100P), and data on the number of error events (BQ) related to the energy. Another possible performance criterion of the previous light source 100P from which the partial feature training set 241-sf (such as partial feature training set 241-7, not explicitly shown in FIG. 2A) can be formed corresponds to the detection of a very low energy of the amplified light beam 105 output from the previous light source 100P that caused a system (light source) shutdown in the previous usage environment.

[0052]

[0062] Other performance criteria envisioned for the previous light source 100P that may form each partial feature training set 241-sf (such as those that may be identified as partial feature training sets 241-8, 241-9, 241-10, not explicitly shown in FIG. 2A) include data regarding the long-term wavelength (of the amplified light beam 105 output from the previous light source 100P), the short-term wavelength (of the amplified light beam 105 output from the previous light source 100P), and the number of error events (BQ) associated with the wavelength.

[0053]

[0063] Other performance criteria envisioned for the previous light source 100P from which each partial feature training set 241-sf (such as those which may be identified as partial feature training sets 241-11 and 241-12, not explicitly shown in FIG. 2A) may be formed include data regarding long-term voltage and power amplifier signatures, and data regarding the number of error events (BQ) associated with such signatures.

[0054]

[0064] Further performance criteria of the previous light source 100P from which the respective partial feature training set 241-sf can be formed include data on the specific modules 115-m in the light source 100. For example, with additional reference to FIG. 4, other performance criteria of the previous light source designed like the dual stage light source 400 from which the respective partial feature training set 241-sf can be formed include data on the energy, wavelength, and bandwidth of the output light beam from the MO system 461 or data on the energy, wavelength, and bandwidth of the output light beam from the PRA system 465. Other performance criteria of the previous light source designed like the dual stage light source 400 from which the respective partial feature training set 241-sf can be formed include data on the pressure of the MO chamber module 461 or data on the pressure of the PRA chamber module 466. The specific modules 115-m in the light source 100 and the specific modules in the light source 400 may be related to each other or coupled to each other such that the operation of one of the modules of the light source can affect the operation of one or more other modules. As an example, if the MO chamber module 461 is not operating at an optimal or acceptable level, this may affect the operation of the PRA chamber module 466, which may then operate outside of acceptable levels to compensate for the poor performance of the MO chamber module 461. In this manner, some data in the partial feature training set 241-sf specifically associated with a particular module may affect the operation of other modules. For example, if the MO chamber module 461 is operating outside of its acceptable levels, the pressure of the PRA chamber module 466 may change. In another example, the energy of the output light beam from the PRA chamber module 466 may affect the operation of the MO chamber module 461.As another example, failures associated with the PRA WEB module 467 are not relevant to the bandwidth performance of the light beam, so it does not make sense to use data from the PRA WEB module 467 to train an ML model using the bandwidth partial feature training set 241-sf.

[0055]

[0065] As described above, the training data 251A acquired for the partial feature training set 241-sf is acquired from a plurality of metrology apparatuses 202P, each associated with a previous light source 100P. Such a metrology apparatus 202P includes a plurality of metrology units, each configured to acquire information about the previous light source 100P. Thus, in the above example, the wavelength data can be acquired from a line center analysis module ("LAM"), which can include, for example, an etalon spectrometer for precision wavelength measurement and a coarser resolution grating spectrometer through which the amplified light beam 105 (or a precursor light beam forming the light beam 105) passes.

[0056]

[0066] As another example, the bandwidth and energy data may be obtained from a bandwidth analysis module ("BAM") positioned to receive a portion of the amplified light beam 105 (or a precursor light beam that forms light beam 105). The BAM is configured to measure the bandwidth and energy of pulses of light beam 105 and may include a grating spectrometer. In a grating spectrometer, light beam 105 (or a precursor thereof) is directed onto an echelle grating, which separates or disperses the light beam according to its wavelength, and the light beam reflected from the grating is directed onto a camera, such as a charged coupled device camera, which can resolve the wavelength distribution of the light beam.

[0057]

[0067] As another example, data regarding the energy of the pulses of the amplified light beam 105 can be measured using an energy monitor that measures the pulse energy of the light beam 105 before it enters the photolithography apparatus 110. The energy monitor can include a photodiode module. With reference to FIG. 4, the energy of the output light beam from the PRA chamber module 466 can be measured using a metrology unit positioned at the output side of the PRA chamber module 466, and the energy of the output light beam from the (OpuS) module 477 can be measured using a metrology unit positioned at the output side of the OpuS module 477. If the energy of the pulses of the output light beam from the PRA chamber module 466 is 10 millijoules (mJ) and the energy of the pulses of the output light beam from the OpuS module 477 is only 2 mJ, then the OpuS module 477 is not operating within an acceptable range. To track the operation of the OpuS module 477, one ML model can be trained for the entire light source 400 and one ML model can be trained for only the PRA chamber module 466, and the aggregation unit 124 references the results from both of these models to determine whether the OpuS module 477 requires repair or replacement.

[0058]

[0068] The training data obtained for the partial feature training set 241-sf is obtained from removed modules of a previous light source 100P. For example, there is a particular population of modules that were removed as part of a comprehensive maintenance run. A subset of these modules would have failed. A complementary subset of these modules will be replaced not because they failed (though they would not have failed), but because another module is being replaced. Each of these removed modules has an array of stored historical data associated with it in the form of metrics collected at various pulse counts. These metrics can be extracted and sorted into the partial feature training set 241-sf. The model ML-sf is then trained by evaluating how well its predictions align with actual historical field results. Ideally, the model ML-sf should predict failure within the prediction increment and not if it is not within the prediction increment. For example, each of the approximately 1000 removed modules can be evaluated initially at 10 Bp, then at 0.1 Bp increments up to the pulse count at the time of removal. This means that if the average removal pulse count is 30 Bp, then each module will yield an average of 200 evaluations ((30 Bp-10 Bp) / 0.1), yielding 200,000 evaluations across the group of removed modules to provide raw data 251A. As will be described below, the build-up (or usage) of increments can be expressed in terms of time. When considering usage in terms of time, the increments per day to pulse count can be considered at removal. Each of the 1000 removed modules can be removed in an average of 180 days, by the end of the expected life of the 10 Bp to 30 Bp.

[0059]

[0069] Metrics can also include additional parameters such as laser attributes including model, area deployed, account associated with the module, customer type, power levels employed, temperature and blower speed, bandwidth and wavelength stability, bearing hours, neon decay settings and failures, efficiency metrics, etc. The above data is captured in moving lookback windows or as a ratio, for example, average wavelength sigma in the previous 1 Bp and previous 5 Bp, or average wavelength sigma in the previous 1 Bp to the first 5 Bp of life. The metrics can also include invariant features such as the total number of individual failure signatures in the previous 100 Mp, 1 Bp, and 2 Bp, or data from individual sensors such as voltage and MO energy.

[0060]

[0070] Assuming at least 4 metrics for each of the 200,000 evaluation outcomes, we get a total of roughly 800,0000 metrics.

[0061]

[0071] These metrics are fed to the respective model ML-sf in block 254A as partial feature training set 241-sf (block 253A). In training, the model ML-sf makes predictions, compares them to field results, then adjusts its application of metrics, and compares again to converge to a prediction logic that is consistent with the field results. For example, the prediction logic can be run to derive approximately 200,000 predictions. These 200,000 predictions can then be compared to a prediction target to predict failures within 2 Bp or less of the removal pulse count for modules that have actually failed, and not otherwise. In one embodiment, the training results in a binary module failure prediction in the prediction increment, for example, predicting module failures within the prediction increment prior to removal due to failure, while not predicting any module failure prior to removal of the module for modules that do not cause technical issues. This system can work for any module and is designed and tested using historical data.

[0062]

[0072] A set of machine learning models ML-sf (block 254A) is stored in memory 128 in maintenance apparatus 120. Machine learning models ML-sf are accessed by prediction unit 122 in operational pipeline 250B during operation of light source 100 (FIG. 2B). Operational pipeline 250B includes data 251B acquired during data acquisition by metrology apparatus 202, which includes a number of metrology units, each configured to acquire information about light source 100 during operation of light source 100. Examples of metrology units in metrology apparatus 202 are described with reference to metrology apparatus 202P.

[0063]

[0073] The raw data 251B can be pre-processed to be in a form from which the situation indicators can be extracted. The pipeline unit 126 receives the raw data 251B (including the partial feature data sets 140-sf) and identifies, extracts and classifies these partial feature data sets 140-sf from the data 251B. The partial feature data sets 140-sf reflect the partial feature training sets 241-sf. Thus, the partial feature data set 140-1 corresponds to real-time data received in relation to the specific performance criteria of the light source 100 during standard operation, while the partial feature training set 241-1 corresponds to training data associated with the specific performance criteria of the previous light source 100P. Thus, in the above example, if the partial feature training set 241-1 includes a set of metrics related to the long-term bandwidth of the amplified light beam 105 generated by the previous light source 100P, then the partial feature data set 140-1 includes a set of metrics related to the long-term bandwidth of the amplified light beam 105 generated by the light source 100. As another example, if the partial feature training sets 241-8, 241-9, and 241-10 include metrics related to the long-term wavelength (of the amplified light beam 105 output from the previous light source 100P), the short-term wavelength (of the amplified light beam 105 output from the previous light source 100P), and data regarding the number of error events (BQ) associated with the wavelength of the light beam 105 output from the previous light source 100P, then the partial feature data sets 140-8, 140-9, and 140-10 include sets of metrics related to data regarding the long-term wavelength, short-term wavelength, and number of error events (BQ) associated with the wavelength of the light beam 105 output from the light source 100, respectively.

[0064]

[0074] Both the pre-processing of the raw data 251B and the feature identification and extraction of the partial feature datasets 140-sf from the data 251B are performed by the pipeline unit 126. In particular, the pipeline unit 126 receives this data 251B including all of the partial feature datasets 140-sf and classifies each partial feature dataset 140-sf into feature categories based on its specific performance criteria. These identified and classified partial feature datasets 140-sf and the assigned categories are provided to the prediction unit 122. Furthermore, the pipeline unit 126 maintains a pipeline between each partial feature dataset 140-sf and the respective machine learning model ML-sf trained based on the partial feature training set 241-sf corresponding to this partial feature dataset 140-sf, so that the prediction unit 122 applies a particular partial feature dataset 140-sf to the machine learning model ML-sf (FIG. 2A) trained based on the assigned category of the particular partial feature training set 241-sf. Each machine learning model ML-sf is applied to a respective partial feature data set 140-sf, and a predicted increment estimate Ev-sf is output from each machine learning model ML-sf. In particular, each machine learning model ML-sf predicts an outcome based on the model prepared and trained on past or historical input data (from data 251A and partial feature training set 241-sf). Like its respective training set 241-sf, each partial feature data set 140-sf targets a particular aspect of the performance of the light source 100 (a unique performance criterion), and each partial feature data set 140-sf is curated to model a particular aspect of the light source 100. The individual separate estimates Ev-sf are provided to the aggregation unit 124.

[0065]

[0075] Referring to FIG. 3A, a usage diagram 355 illustrates a time period used herein to describe the operation of the maintenance device 120 based on the usage of the light source 100. The diagram 355 unfolds along an increasing axis of usage 356 of the light source 100 (represented by an arrow pointing to the right). The diagram 355 is not necessarily drawn to scale in order to present the concept within a single page. The usage 356 of the light source 100 can be measured in any suitable configuration. For example, the usage 356 of the light source 100 can be measured in terms of the number of pulses of the amplified light beam 105 that are generated. Thus, in the diagram 355, the number of pulses of the amplified light beam 105 increases in the direction of the arrow. In other implementations, it may be appropriate to measure the usage 356 of the light source 100 in terms of time, for example, in days, hours, minutes, etc.

[0066]

[0076] The maintenance device 120 performs an evaluation at each evaluation instant or evaluation point Ei (where i=0, 1, 2, 3, etc.) along the axis of usage 356. Evaluation points E0, E1, E2, E3 are shown in FIG. 3, but more evaluation points can follow E3, although they are not shown in FIG. 3. The first evaluation E0 is performed after usage U0. An evaluation increment EIj is the usage interval between adjacent evaluations Ei and Ei+1, where j is indicated by 0, 1, 2, etc. Only three evaluation increments EIj are shown in FIG. 3A, namely EI0, ​​EI1, EI2, where EI0 is the usage interval between E0 and E1, EI1 is the usage interval between E1 and E2, and EI2 is the usage interval between E2 and E3. Prediction increments PIi are shown extending from each evaluation point Ei in the direction of increasing usage. Thus, the prediction increment PI0 extends from evaluation point E0, and the prediction increment PI1 extends from evaluation point E1.

[0067]

[0077] At each evaluation point Ei, there is at least one lookback increment LBIi that looks back from that evaluation point Ei. As shown in the implementation or example of FIG. 3A, there are three lookback increments LBIi associated with each evaluation point Ei. In other implementations, there may be fewer or more than three lookback increments LBIi associated with each evaluation point Ei. In this example, a lookback increment LBI0 is associated with evaluation point E0, a lookback increment LBI1 is associated with evaluation point E1, a lookback increment LBI2 is associated with evaluation point E2, and a lookback increment LBI3 is associated with evaluation point E3.

[0068]

[0078] Specifically, in operation (such as when performing the operational pipeline 250B), at each evaluation point Ei, the maintenance apparatus 120 evaluates for each partial feature data set 140-sf whether a failure mode is detected in the light source 100 (within the prediction increment Pi) and determines which performance criteria of the light source 100 are associated with the failure mode based on the associated partial feature data set 140-sf of the failure mode. The maintenance apparatus 120 may determine that a failure mode exists for more than one partial feature data set 140-sf. The maintenance apparatus 120 may further determine that there are multiple performance criteria associated with a particular failure mode. At each evaluation point Ei, the maintenance apparatus 120 looks forward at the prediction increment Pi and determines whether a failure mode occurs within that prediction increment Pi. On the other hand, at each evaluation point Ei, the maintenance apparatus 120 considers the aggregation of each partial feature data set 140-sf in use over the look-back increment LBIi preceding the evaluation point Ei. Thus, for example, at evaluation point E0, data for a particular partial feature dataset 140-sf is taken from an aggregation of use of data taken at three different lookback increments LBI0_1, LBI0_2, LBI0_3. Lookback increment LBIi can include a previous usage step back from evaluation point Ei, in which case the aggregation simply includes use of the previous partial feature dataset 140-sf in that previous usage step.

[0069]

[0079] The aggregation occurs while maintaining the pipeline: multiple partial feature data sets 140-1 in lookback increment LBIi are aggregated with each other (but not with other partial feature data sets 140-2, 140-3, etc.), multiple partial feature data sets 140-2 in lookback increment LBIi are aggregated with each other (but not with other partial feature data sets 140-1, 140-3, etc.), and multiple partial feature data sets 140-3 in lookback increment LBIi are aggregated with each other (but not with other partial feature data sets 140-1, 140-2, etc.).

[0070]

[0080] 3B, an example of three lookback increments LBI0 at evaluation point E0 is shown for three partial feature data sets 140-1, 140-2, 140-3. In this example, the third lookback increment LBI0 (indicated by the arrow 3) includes one previous use step, and the data evaluated at evaluation point E0 for the third lookback increment LBI0 (referred to as 140-1(AG3), 140-2(AG3), 140-3(AG3)) are simply the respective values ​​of these three partial feature data sets 140-1, 140-2, 140-3 at the time they existed in the previous use step. The first lookback increment LBI1 (indicated by the arrow 1) includes several back use steps. For example, the first lookback increment LBI1 may correspond to 100 million pulses back. Thus, for the first lookback increment LBI1, the data of the partial feature data set 140-1 is aggregated back by 100 million pulses to form the aggregated partial feature data set 140-1 (AG1), the data of the partial feature data set 140-2 is aggregated back by 100 million pulses to form the aggregated partial feature data set 140-2 (AG1), and the data of the partial feature data set 140-3 is aggregated back by 100 million pulses to form the aggregated partial feature data set 140-3 (AG1). The second lookback increment LBI1 (indicated by the two arrows) includes several back usage steps and is larger than the first lookback increment LBI1 and the third lookback increment LBI1. For example, the second lookback increment LBI1 may correspond to a billion pulses back. Thus, for the second lookback increment LBI1, the data of partial feature dataset 140-1 is aggregated going back 1 billion pulses to form aggregated partial feature dataset 140-1(AG2), the data of partial feature dataset 140-2 is aggregated going back 1 billion pulses to form aggregated partial feature dataset 140-2(AG2), and the data of partial feature dataset 140-3 is aggregated going back 1 billion pulses to form aggregated partial feature dataset 140-3(AG2).

[0071]

[0081] The data may be aggregated using any suitable mathematical construct. For example, aggregating the data of the partial feature data set 140-sf may correspond to simply combining values ​​according to an aggregation principle, such as using summation, or finding the median value, or computing the average of the values.

[0072]

[0082] In some implementations, each look-back increment LBIi is selected from a set of look-back increments that are assumed to be relevant to how the performance criteria changes as usage progresses. In some implementations, the set of look-back increments LBIi includes a first look-back increment (indicated by arrow 1) of approximately 2 billion pulses, a second look-back increment (indicated by arrow 2) of approximately 1 billion pulses, and a third look-back increment (indicated by arrow 3) of approximately 100 million pulses.

[0073]

[0083] Referring again to FIG. 3A, one particular implementation is described below. The usage U0 is 10 billion pulses, which means that the maintenance device 120 performs the first evaluation E0 after the light source 100 generates 10 billion pulses. The evaluation increment EIj is 100 million pulses. Thus, the maintenance device 120 performs an evaluation at each evaluation point Ei every 100 million pulses. There are three lookback increments LBIi (hence three lookback arrows as shown in FIGS. 3A and 3B), where the lookback increments LBIi are 100 million pulses, 1 billion pulses, and 2 billion pulses. In this particular implementation, for module 115-m in light source 100 operating over 40 billion pulses, evaluation point E0 is at the 10 billionth pulse, evaluation point E1 is at the 10.1 billionth pulse, evaluation point E2 is at the 10.2 billionth pulse, evaluation point E3 is at the 10.3 billionth pulse, and so on through the 40 billionth pulse. Further, in this particular implementation, at evaluation point E0 (the 10 billionth pulse), a first lookback increment LBI0 of 100 million pulses includes data from 9.9 billion pulses to 10 billion pulses, a second lookback increment LBI0 of 1 billion includes data from 9 billion pulses to 10 billion pulses, and a third lookback increment LBI0 of 2 billion pulses includes data from 8 billion pulses to 10 billion pulses. At evaluation point E1 (the 10.1 billionth pulse), the first lookback increment LBI1 of 100 million pulses contains data from 10 billion pulses to 10.1 billion pulses, the second lookback increment LBI1 of 1 billion pulses contains data from 9.1 billion pulses to 10.1 billion pulses, and the third lookback increment LBI1 of 2 billion pulses contains data from 8.1 billion pulses to 10.1 billion pulses.

[0074]

[0084] The size of the evaluation increment EIj can be adjustable and configurable, for example, the evaluation increment can be 50 million pulses or 1.5 billion pulses.

[0075]

[0085] In some implementations, at each evaluation point Ei, the maintenance device 120 looks ahead in short-term prediction increments Pi, such as 100 million pulses ahead to one day ahead. In other implementations, at each evaluation point Ei, the maintenance device 120 looks ahead in long-term prediction increments Pi, such as 2 billion pulses ahead to 20 days ahead.

[0076]

[0086] As described above, a portion of the partial feature data set 140-sf (such as short-term bandwidth, short-term wavelength, and short-term energy) may be associated with a performance criterion tracked with a relatively short lookback increment LBIi. Using the above example, the short-term lookback increment LBIi may be 100 million pulses. A portion of the partial feature data set 140-sf (such as long-term bandwidth, long-term wavelength, long-term energy) may be associated with a performance criterion tracked with a relatively long lookback increment LBIi. Using the above example, the long-term lookback increment LBIi may be 1 billion pulses. Furthermore, in an implementation where there is a long-term lookback increment LBIi, the pipeline unit 126 is configured to aggregate the partial feature data set 140-sf in use based on the same lookback increment, and the prediction unit 122 is configured to receive a plurality of partial feature data sets 140-sf by receiving an aggregate of the partial feature data sets 140-sf tracked using the same lookback increment LBIi.

[0077]

[0087] 4, implementation 400 of light source 100 is a dual stage pulsed light source that generates a pulsed amplified beam 405 as light beam 105. Light source 400 includes a solid state or gas discharge master oscillator (MO) system 460, a power amplifier (PA) system such as a power ring amplifier (PRA) system 465, relay optics 470, and an optical output subsystem 475.

[0078]

[0088] The MO system 460 may include, for example, an MO chamber module 461 in which a laser gas discharge between electrodes (not shown) creates a population inversion of energetic molecules, including argon, krypton, or xenon, in a laser gas to generate a relatively broadband emission that is line narrowed to a relatively very narrow bandwidth and center wavelength selected in a line narrowing module ("LMN") 462. The MO system 460 may also include an MO output coupler (MO OC) 462, which may include a partially reflective mirror that, together with a reflective grating (not shown) in the LNM 462, forms an oscillator cavity in which the MO system 460 oscillates to form a seed output pulse, thereby forming a master oscillator. The MO system 460 may also include a line center analysis module (LAM) 463. As discussed above, the LAM 180 can include, for example, an etalon spectrometer for precision wavelength measurements and a coarser resolution grating spectrometer.

[0079]

[0089] The relay optics 470 may include an MO wavefront engineering box (WEB) 471 that serves to redirect the output of the MO system 460 towards the PA system 465 and may include beam expansion, for example by a multi-prism beam expander (not shown), and coherence busting, for example in the form of an optical delay path (not shown).

[0080]

[0090] The PA system 465 includes, for example, a PRA chamber module 466, which is an oscillator formed by the input of the output light beam from the MO system 460, and output coupling optics (not shown) that may be incorporated into the PRA WEB 467 and redirected through a gain medium in the chamber 466 via a beam reservoir 468. The PRA WEB 467 may incorporate a partially reflective input / output coupler (not shown), a maximum reflecting mirror for the nominal operating wavelength (which may be about 193 nm for an ArF system), and one or more prisms. The PRA system 465 optically amplifies the output light beam from the MO system 460.

[0081]

[0091] The optical output subsystem 475 may include a bandwidth analysis module (BAM) 476 at the output of the PA system 465, which receives the pulses of the output light beam from the PA system 465 and extracts a portion of the light beam for metrology purposes, for example to measure the output bandwidth and pulse energy. The pulses of the output light beam then pass through an optical pulse stretcher (OpuS) module 477 and an output coupling autoshutter metrology module (CASMM) 478, which may also be the location of a pulse energy meter. One of the purposes of the OpuS 477 may be to convert the single output pulse into a pulse train. The secondary pulses generated from the original single output pulse may be delayed with respect to each other. By dispersing the original laser pulse energy into the secondary pulse train, the effective pulse length of the light beam may be stretched and the peak pulse intensity may be reduced at the same time.

[0082]

[0092] The light source 400 is composed of modules 115-m. Each of the components of the light source 400 (MO chamber 461, LNM 462, MO WEB 471, PRA chamber 466, PRA WEB 467, OPuS 477, BAM 476, etc.) is a module 115-m. The overall availability of the light source 400 is a direct result of the availability of each of these individual modules 115-m that compose the light source 400. In other words, the light source 400 cannot be available unless all of these modules that compose the light source 400 are available. The maintenance device 120 makes it possible to monitor these modules and replace them before they fail, thus maintaining the operation of the light source 400 and optimizing and increasing the productivity of the output device 110. The maintenance device 120 provides active failure alerts that can be used to pinpoint which module 115-m is about to fail.

[0083]

[0093] Referring to FIG. 5, in some implementations, the output device 110 is a photolithography exposure device 510. The photolithography exposure device 510 uses an amplified light beam 105 to pattern microelectronic features on a substrate or wafer 511. The wafer 511 is placed on a wafer table 512 that is constructed to hold the wafer 511 and is connected to a positioner configured to precisely position the wafer 511 according to predetermined parameters. The photolithography exposure device 510 can use a light beam 505 (output from the light source 100) having a wavelength in the deep ultraviolet (DUV) range, which can include wavelengths from about 100 nanometers (nm) to about 400 nm. For example, the light source 100 generating such a light beam 505 can be a gas discharge source, such as an excimer source, or an excimer laser that uses a combination of one or more noble gases (which can include argon, krypton, or xenon) and a reactive gas (which can include fluorine or chlorine) as a gain medium. The light source 400 may be an excimer light source. Thus, for example, the gain medium may include argon fluoride (ArF), krypton fluoride (KrF), or xenon chloride (XeCl). When the gain medium includes argon fluoride, the wavelength of the amplified light beam 105 is about 193 nm, and when the gain medium includes krypton fluoride, the wavelength of the amplified light beam 105 is about 248 nm. The size of the microelectronic features patterned on the wafer 511 depends on the wavelength of the light beam 505, with shorter wavelengths resulting in smaller minimum feature sizes. When the wavelength of the light beam 505 is 248 nm or 193 nm, the minimum size of the microelectronic features may be, for example, 50 nm or less. The bandwidth of the light beam 505 may be the actual instantaneous bandwidth of its optical spectrum (or emission spectrum), which contains information about how the optical energy of the light beam 505 is distributed across different wavelengths.

[0084]

[0094] The photolithography exposure apparatus 510 includes an optical arrangement having, for example, one or more focusing lenses, a mask, and an objective arrangement. The mask is movable along one or more directions, such as along the optical axis of the light beam 505 or in a plane perpendicular to the optical axis. The objective arrangement includes a projection lens and enables image transfer from the mask to the photoresist on the wafer 511. The photolithography exposure apparatus 510 also includes an illumination system that adjusts the range of angles at which the light beam 505 impinges on the mask. The illumination system also equalizes (uniforms) the intensity distribution of the light beam 505 across the mask.

[0085]

[0095] The photolithography exposure apparatus 510 may also include, among other features, a lithography controller 513, air conditioning devices, and power supplies for various electrical components. The lithography controller 513 controls how layers are printed on the wafer 511. The lithography controller 513 includes memory that stores information such as process recipes. The process program or recipe determines the length of exposure on the wafer 511, the mask used, and other factors that affect the exposure. During lithography, multiple pulses of the light beam 505 illuminate the same area of ​​the wafer 511 and together make up an illumination dose.

[0086]

[0096] With reference to FIG. 6, once the ML model ML-sf has been trained (as shown in the training pipeline 250A of FIG. 2A) and stored in the memory 128, the maintenance device 120 executes a procedure 680 for maintaining the light source 100 by evaluating whether and which modules 115-m should be repaired or replaced. First, the maintenance device 120 receives (681) a partial feature data set 140 in the form of operational data 251B (see also FIG. 2). Such data 251B is obtained from various metrology units of the metrology device 202. With further reference to FIG. 3A, the maintenance device 120 receives the operational data 251B just before an evaluation point Ei while the light source 100 is in use (and is increasingly being used). After the first evaluation E0, the maintenance device 120 receives each set of data 251B following an evaluation increment EIj. Additionally, the data 251B may be aggregated according to one or more look-back increments LBIi. Thus, the data 251B may be stored in the memory 128, and when the evaluation point Ei is reached, the historical data 251B may be aggregated according to the look-back increments LBIi at the evaluation point Ei.

[0087]

[0097] As described above, the operational data 251B is parsed into a number of partial feature data sets 140-sf, with each partial feature data set 140-sf associated with a unique performance criterion of the light source 100. An example of nine partial feature data sets 140-sf (sf=1, 2...9), each representing one of three performance criteria: bandwidth, wavelength, and energy, is shown in the table of Fig. 7A. The partial feature data sets 140-1, 140-2, 140-3 are all associated with the bandwidth of the amplified light beam 105 output from the light source 100, the partial feature data sets 140-4, 140-5, 140-6 are all associated with the wavelength of the amplified light beam 105 output from the light source 100, and the partial feature data sets 140-7, 140-8, 140-9 are all associated with the energy of the amplified light beam 105 output from the light source 100. Thus, ML models ML-1, ML-2, ML-3 were trained on the long-term bandwidth training set, the short-term bandwidth training set, and the bandwidth error training set, respectively, ML models ML-4, ML-5, ML-6 were trained on the long-term wavelength training set, the short-term wavelength training set, and the wavelength error training set, respectively, and ML models ML-7, ML-8, ML-9 were trained on the long-term energy training set, the short-term energy training set, and the energy error training set, respectively. As also mentioned above, the long-term training set is a data set aggregated over the longer lookback increment LBIi, while the short-term training set is a data set aggregated over the shorter lookback increment LBIi.

[0088]

[0098] Referring to FIG. 6, the maintenance device 120 performs an evaluation based on the received partial feature data set 140-sf (682). Specifically, in step 682, the maintenance device 120 evaluates 682-sf for each partial feature data set 140-sf whether a failure mode is detected in the light source 100. In particular, the prediction unit 122 performs each evaluation 682-sf using the ML model ML-sf associated with the partial feature data set 140-sf according to the pipeline described above. The prediction unit 122 performs the evaluation at an evaluation point Ei (FIG. 3A). In evaluating 682-sf whether a failure mode is detected for a particular partial feature data set 140-sf, the prediction unit 122 is evaluating whether a failure will occur in the light source 100 in future use (i.e., at a previous use prediction increment PIi (FIG. 3A)).

[0089]

[0099] The evaluation 682-sf is a score based on the results of the respective ML model ML-sf. The score is related to the probability that a failure occurs in the module 115-m of the light source 100 at the prediction increment PIi. The score can be a certainty score, a binary output, or a function ranging from 0 to 1. For example, if the ML model ML-sf is a random forest, the score is based on majority voting since a random forest is a collection of uniformed decision trees. As another example, for an ML model ML-sf that is a single decision tree with two classes, the score can correspond to the end node. For example, if the tree ends with an end node (leaf) that has 100 samples, 80 of which are class 1 and 20 of which are class 2, the certainty score can be defined as 80 / 100=0.8 for class 1 and 20 / 100=0.2 for class 2. Different end nodes and tree paths give rise to other certainty scores based on the data. The tree is motivated to generate pure end nodes. As a further example, the ANN ML model ML-sf may use an activation function such as a sigmoid function or a logistic function or some other function ranging from 0 to 1 to transform the results into a two-class output. Thus, the scores from the evaluation 682-sf may vary depending on which ML model ML-sf is used for the particular evaluation 682-sf.

[0090]

[0100] Referring to the table of FIG. 7B, a set of ratings 682-1, 682-2, 682-3, 682-4, 682-5, 682-6, 682-7, 682-8, 682-9 is shown along with output scores for each of the nine partial feature data sets 140-1, 140-2, 140-3, 140-4, 140-5, 140-6, 140-7, 140-8, 140-9 of FIG. 7A, with each partial feature data set 140-sf representing one of three performance criteria categories: bandwidth, wavelength, and energy. As described above with reference to FIG. 7A, ML models ML-1, ML-2, ML-3 were trained on the long-term bandwidth training set, the short-term bandwidth training set, and the bandwidth error training set, respectively; ML models ML-4, ML-5, ML-6 were trained on the long-term wavelength training set, the short-term wavelength training set, and the wavelength error training set, respectively; and ML models ML-7, ML-8, ML-9 were trained on the long-term energy training set, the short-term energy training set, and the energy error training set, respectively. Evaluations 682-sf are shown for evaluation points E1, E2, and E3. More evaluation points Ei than those shown in FIG. 7B may follow. For example, at evaluation point E1, the output of evaluation 682-1 (corresponding to evaluation of partial feature data set 140-1 using ML model ML-1) is 0.16, and the output of evaluation 682-8 (corresponding to evaluation of partial feature data set 140-8 using ML model ML-8) is 0.89. As a further example, at evaluation point E2, the output of evaluation 682-3 (corresponding to evaluation of partial feature dataset 140-3 with ML model ML-3) is 0.21, the output of evaluation 682-4 (corresponding to evaluation of partial feature dataset 140-4 with ML model ML-4) is 0.92, the output of evaluation 682-5 (corresponding to evaluation of partial feature dataset 140-4 with ML model ML-4) is 0.85, and the output of evaluation 682-8 (corresponding to evaluation of partial feature dataset 140-8 with ML model ML-8) is 0.28. The prediction unit 122 can determine that a failure mode has been detected if a particular score is greater than a minimum value.For example, the prediction unit 122 may determine that a failure mode has been detected if the score is greater than 0.8, such a score being shown in bold in FIG.

[0091]

[0101] Next, the maintenance device 120 determines (683) which one or more performance criteria are associated with any failure mode detected in step 682. In particular, in step 683, the aggregation unit 124 first determines whether a failure mode was detected (683-sf) by any of the evaluations 682-sf (in step 682), and if there is a detected failure mode (683-sf), the aggregation unit 124 determines (684-sf) which one or more performance criteria are associated with that failure mode 683-sf. The determination in 684-sf is based on the partial feature data set 140-sf associated with the failure mode 683-sf.

[0092]

[0102] The aggregation unit 124 can aggregate the assessments from the prediction unit 122 and determine which one or more performance criteria are associated with the failure mode based on the associated partial feature data set by prioritizing the reasons for failure associated with each detected failure mode (685).

[0093]

[0103] In this example, the aggregation unit 124 is configured to classify a single rating of a particular performance criterion as a "hard failure" if the rating is greater than 0.95, and as a "soft failure" if two separate ratings from two separate partial feature data sets within the particular performance criterion are between 0.85 and 0.95, and otherwise classify as no failure mode. Based on this configuration, the following evaluation 682-sf, i.e., evaluation 682-8 performed at evaluation point E1 (for partial feature data set 140-8, which is short-term energy), results in a hard failure associated with short-term energy. Furthermore, the following evaluations, i.e., evaluation 682-4 performed at evaluation point E2 (for partial feature data set 140-4, which is long-term wavelength), and evaluation 682-6 performed at evaluation point E2 (for partial feature data set 140-6, which is wavelength error), result in a soft failure. In this particular example, in step 683, the aggregation unit 124 determines that the failure mode detected during evaluation 682-8 at evaluation point E1 is related to the energy of the light beam 105 output from the light source 100, the failure mode detected during evaluation 682-4 at evaluation point E2 is related to the long-term wavelength of the light beam 105 output from the light source 100, and the failure mode detected during evaluation 682-6 at evaluation point E2 is related to a wavelength error of the light beam 105 output from the light source 100. A hard failure may indicate that a complete replacement of a particular module 115-m of the light source 100 is required. A soft failure may indicate that a particular module 115-m of the light source 100 requires observation or repair, or that the performance criteria require labeling for further evaluation. In other examples, other hard and soft failure values ​​may be used.

[0094]

[0104] As described above, some modules 115-m of the light source 100 are much more likely to affect a given performance criterion or criteria than other modules 115-m of the light source 100. For example, the operation of the MO chamber module 461 has a greater impact on the bandwidth, wavelength, and energy of the amplified light beam 405 from the light source 400. Therefore, for this reason, a hard failure (682-8 in E1) detected for short-term energy may indicate that the MO chamber module 451 will fail during the prediction increment PIi. The aggregation unit 124 may analyze those scores 682-sf associated with a particular assigned performance criterion or category when determining whether a module 115-m requires replacement or monitoring. For example, the aggregation unit 124 may analyze the scores 682-7, 682-8, 682-9, all of which are associated with the energy of the amplified light beam 105, to make a determination as to whether a particular module 115-m (such as the MO chamber module 461) requires monitoring or replacement. The maintenance device 120 can pinpoint the problem within the light source 100 (and more specifically within a particular module 115-m) based on which partial feature data set 140-sf resulted in the failure mode.

[0095]

[0105] In one example, the aggregation unit 124 is configured to label a performance criterion (e.g., energy) for further evaluation if at least one failure mode associated with that performance criterion is detected, regardless of whether other failure modes associated with other performance criteria (e.g., bandwidth and wavelength) are also detected. Indeed, in FIG. 7B, the aggregation unit 124 labels energy for further evaluation at evaluation point E1, even though no failure modes are detected for either bandwidth or wavelength at evaluation point E1.

[0096]

[0106] The aggregation unit 124 can aggregate the assessments 682-sf from the prediction unit 122 and determine which one or more performance criteria are associated with the failure mode based on the associated partial feature data set 140-sf by prioritizing the failure reasons associated with each detected failure mode.

[0097]

[0107] For example, if one of the bandwidth models ML-1, ML-2, ML-3 reports a score 682-sf of 0.8, but the wavelength and energy models ML-4, ML-5, ML-6, ML-7, ML-8, ML-9 of the MO chamber module 461 report scores 682-sf of 0.1 and 0.15, respectively, the aggregation unit 124 infers that the bandwidth of the light beam is out of range and will soon cause a failure in the MO chamber module 461. Furthermore, as described above, there are three partial feature data sets corresponding to the bandwidth: long-term bandwidth (associated with model ML-1), short-term bandwidth (associated with model ML-2), and bandwidth error (associated with model ML-3). If the scores 682-sf of models ML-1, ML-2, and ML-3 are 0.7, 0.1, and 0.15, respectively, and all other models ML-4 to ML-9 report scores less than 0.2, the aggregation unit 124 determines that a bandwidth error is the root cause and that an error has been generated, which leads to failure and replacement of the MO chamber module 461.

[0098]

[0108] As an example, the aggregation unit 124 may instruct the following actions based on each of the scores (ratings 682-sf) (where the scores may be any number between 0 and 1, and each score is associated with an ML model): For a score 682-sf between 0 and 0.25, the aggregation unit 124 may determine that there is nothing wrong and that it can be ignored and there is no need to report it; For a score 682-sf between 0.25 and 0.5, the aggregation unit 124 may determine that there is no problem but that it can be reported; For a score 682-sf between 0.5 and 0.75, the aggregation unit 124 may determine that a potential problem exists and that it can be reported and instructed to monitor; For a score 682-sf between 0.75 and 1, the aggregation unit 124 may determine that an imminent problem exists and that it can be reported and instructed to replace the module 115-m since a failure will occur soon.

[0099]

[0109] Prior to step 681, procedure 680 may further include receiving an overall feature set 140 including a plurality of partial feature datasets 140-sf and classifying each of these partial feature datasets 140-sf into feature categories based on its associated specific performance criterion. Specifically, pipeline unit 126 may perform these functions to prepare data for use by prediction unit 122 and to maintain a link between each partial feature dataset 140-sf and its associated ML model ML-sf. Additionally, pipeline unit 126 may aggregate data from each partial feature dataset 140-sf in use based on one or more lookback increments LBIi, as described above.

[0100]

[0110] The following clauses are used to further describe the embodiments. 1. An apparatus for maintaining a light source including one or more modules configured together to generate a light beam for semiconductor photolithography, comprising: A prediction unit, receiving a plurality of partial feature data sets, each partial feature data set being associated with a unique performance metric of the light source; and a prediction unit configured to evaluate, for each partial feature data set, whether a failure mode is detected in the light source in a prediction increment; An aggregate unit, receiving a plurality of estimates from the prediction unit; and and an aggregation unit configured to determine, for each failure mode, which one or more performance criteria are associated with the failure mode based on the associated partial feature data set of the failure mode. Device. 2. The apparatus of claim 1, wherein the prediction unit includes or has access to multiple models, each model generated by machine learning using a unique partial feature dataset. 3. The apparatus of claim 2, wherein each of the models in the prediction unit has an identical matching structure. 4. The apparatus of claim 2, wherein each of the models in the prediction unit has a different fitting structure. 5. The apparatus of clause 2, wherein at least one of the models in the prediction unit has a matching structure that is different from a matching structure of another model in the prediction unit. 6. The apparatus of claim 1, wherein the plurality of partial feature data sets include a partial feature data set relating to a bandwidth of the light beam, a partial feature data set relating to a wavelength of the light beam, and a partial feature data set relating to an energy of the light beam. 7. The apparatus of clause 1, wherein the plurality of partial feature datasets include a partial feature dataset for a long-term bandwidth of the light beam, a partial feature dataset for a short-term bandwidth of the light beam, a partial feature dataset for bandwidth error events, a partial feature dataset for a long-term wavelength of the light beam, a partial feature dataset for a short-term wavelength of the light beam, a partial feature dataset for wavelength error events, a partial feature dataset for a long-term energy of the light beam, a partial feature dataset for a short-term energy of the light beam, and a partial feature dataset for energy error events. 8. The apparatus of clause 1, wherein the aggregation unit is configured to determine which one or more performance criteria are associated with the failure mode based on the associated partial feature dataset by aggregating the evaluations from the prediction units and prioritizing the failure reasons associated with each detected failure mode. 9. The apparatus of clause 8, wherein the aggregation unit is configured to aggregate evaluations from the prediction units based on a performance criterion associated with each partial feature dataset. 10. The apparatus of clause 9, wherein the aggregation unit is configured to label a performance criterion for further evaluation if at least one failure mode associated with the performance criterion is detected. 11. The apparatus of clause 9, wherein the aggregation unit is configured to label a performance criterion for further evaluation if at least one failure mode associated with the performance criterion is detected, regardless of whether another failure mode associated with another performance criterion is detected. 12. The apparatus of clause 1, wherein the prediction unit is configured to evaluate whether a failure mode is detected in the light source in a prediction increment includes evaluating whether a failure mode is detected in a particular module of the light source based on a model trained for the particular module. 13. The apparatus according to claim 12, wherein the particular module of the light source includes a master oscillator module. 14. The apparatus of clause 1, wherein the prediction unit is configured to evaluate whether a failure mode is detected in the light source in a prediction increment, including evaluating whether a failure mode is detected in one or more specific modules of the light source based on a plurality of models each trained for a specific module. 15. The apparatus described in clause 14, wherein the one or more specific modules of the light source include one or more of a master oscillator module, a power amplifier module, a line narrowing module, a spectral feature analysis module, and a pulse stretcher module. 16. The prediction unit is configured to evaluate whether a failure mode is detected in the light source at a prediction increment; separately evaluating each partial feature data set according to a model generated by machine learning using the partial feature data set; and for each sub-feature dataset evaluation, outputting a score on the probability of failure. 2. The device according to clause 1. 17. The apparatus of claim 16, wherein the score is a certainty score, a binary output, or a function ranging from 0 to 1. 18. The apparatus of clause 1, wherein the prediction unit is configured to evaluate, for each partial feature data set, whether a failure mode is detected in the light source in a prediction increment by performing an evaluation after each evaluation increment. 19. The apparatus of claim 18, wherein the evaluation increment is 100 million pulses of the light beam. 20. An apparatus for maintaining a light source including one or more modules configured together to generate a light beam for semiconductor photolithography, comprising: A pipeline unit comprising: receiving a plurality of partial characteristic data sets, each partial characteristic data set associated with a unique performance criterion of the light source during operation; and a pipeline unit configured to classify each of the received partial feature data sets into a feature category based on an associated specific performance criterion; A prediction unit, receiving a plurality of partial feature data sets and an assigned category for each partial feature data set; a prediction unit configured to evaluate, at a prediction increment, whether a failure mode is detected in the light source; and an aggregation unit configured to determine which one or more performance criteria are associated with a failure mode based on an assigned category of each partial feature data set. Device. 21. The apparatus of clause 20, wherein the prediction unit is configured to evaluate whether a failure mode is detected in the light source in a prediction increment includes evaluating whether a failure mode is detected in a particular module of the light source based on a plurality of models trained for the particular module. 22. The apparatus of clause 21, wherein each model is grown by machine learning by supplying a respective partial feature dataset to train the model. 23. The apparatus of clause 21, further comprising an alert generation unit configured to prescribe a maintenance action for the particular module if the evaluation determines that a faulty module is detected in the particular module. 24. The prediction unit is configured to evaluate whether a failure mode is detected in the light source at a prediction increment; separately evaluating each partial feature data set according to a model generated by machine learning using the partial feature data set; and for each sub-feature dataset evaluation, outputting a score on the probability of failure. 20. Apparatus according to claim 20. 25. The apparatus of clause 24, wherein the aggregation unit is configured to determine which one or more performance criteria are associated with the failure mode based on the assigned category of each partial feature data set, comprising analyzing scores from partial feature data sets where the same assignment is associated with a category. 26. The apparatus of claim 24, wherein the score is a certainty score, a binary output, or a function ranging from 0 to 1. 27. The apparatus of clause 20, wherein the prediction unit is configured to evaluate whether a failure mode is detected in the light source at the prediction increment, including determining whether at least one module in the light source does not have at least a minimum probability of operating without failure at the prediction increment. 28. The apparatus of claim 20, wherein the prediction increment is measured as a number of pulses of the light beam. 29. The apparatus of clause 20, wherein prior to each evaluation, the pipeline unit is further configured to, in use, aggregate data of at least one partial feature dataset across one or more lookback increments preceding the evaluation. 30. The apparatus described in clause 29, wherein the prediction unit is configured to evaluate whether a failure mode is detected in the light source in a prediction increment, comprising: evaluating each partial feature dataset separately according to a model generated by machine learning using the partial feature dataset; and, in use, using the aggregated partial feature dataset when the partial feature datasets are aggregated across one or more lookback increments. 31. The apparatus of clause 30, wherein each look-back increment is selected from a set of look-back increments that are assumed to be related to how the performance criterion changes as use progresses. 32. The apparatus of clause 31, wherein the set of envisioned lookback increments includes a first lookback increment of approximately 2 billion pulses, a second lookback increment of approximately 1 billion pulses, and a third lookback increment of approximately 100 million pulses. 33. The apparatus described in clause 30, wherein one or more of the partial feature data sets are associated with performance criteria tracked at relatively short lookback increments and one or more of the partial feature data sets are associated with performance criteria tracked at relatively long lookback increments. 34. The apparatus described in clause 29, wherein, in use, the pipeline unit is configured to aggregate one or more partial feature data sets based on a same lookback increment, and the prediction unit is configured to receive multiple partial feature data sets by receiving an aggregation of partial feature data sets that are tracked using the same lookback increment. 35. A method for maintaining a light source including one or more modules configured together to generate a light beam for semiconductor photolithography, comprising: receiving a plurality of partial feature data sets, each partial feature data set being associated with a unique performance metric of the light source; For each partial feature data set, evaluating whether a failure mode is detected in the light source in the prediction increment; and for each detected failure mode, determining which one or more performance criteria are associated with the failure mode based on the associated partial feature data set of the failure mode. method. 36. After receiving the plurality of partial feature data sets, further comprising classifying each of the received partial feature data sets into a feature category based on an associated specific performance criterion; determining which one or more performance criteria are associated with the failure mode includes determining which one or more performance criteria are associated with the failure mode based on the assigned category of each partial feature data set; The method described in clause 35. 37. In a prediction increment, evaluating whether a failure mode is detected in the light source includes: separately evaluating each partial feature data set according to a model generated by machine learning using the partial feature data set; and for each sub-feature dataset evaluation, outputting a score on the probability of failure. The method described in clause 35. 38. The method of claim 35, wherein evaluating whether a failure mode is detected in the light source in a prediction increment includes performing the evaluation each time an evaluation increment elapses. 39. The method of clause 35, wherein determining which one or more performance criteria are associated with a failure mode based on the associated partial feature data set of the failure mode includes aggregating the evaluations and prioritizing the failure reasons associated with each detected failure mode. 40. The method of clause 39, wherein aggregating the evaluations includes aggregating the evaluations based on a performance criterion associated with each partial feature data set. 41. The method of clause 40, further comprising labeling a performance criterion for further evaluation if at least one failure mode associated with the performance criterion is detected. 42. The method of clause 40, further comprising, if at least one failure mode associated with a performance criterion is detected, labeling that performance criterion for further evaluation, regardless of whether another failure mode associated with another performance criterion is detected. 43. The method of claim 35, wherein in a prediction increment, evaluating whether a failure mode is detected in the light source includes evaluating whether a failure mode is detected in a particular module of the light source based on a model trained for the particular module. 44. The method of claim 35, wherein in a prediction increment, evaluating whether a failure mode is detected in the light source includes evaluating whether a failure mode is detected in one or more specific modules of the light source based on a plurality of models each trained for a specific module. 45. The method of claim 35, further comprising, prior to each evaluation, aggregating, in use, data of at least one partial feature dataset across one or more lookback increments preceding the evaluation. 46. ​​The method of clause 45, wherein in use aggregating data of at least one partial feature dataset across one or more lookback increments includes, in use, aggregating one or more partial feature datasets based on the same lookback increment, and receiving the multiple partial feature datasets includes receiving an aggregation of partial feature datasets tracked using the same lookback increment. 47. A method for maintaining a light source including one or more modules configured together to generate a light beam for semiconductor photolithography, comprising: receiving a plurality of partial feature data sets, each partial feature data set being associated with a unique performance metric of the light source; for each partial feature data set, evaluating whether a failure mode is detected in the light source in the prediction increment, comprising evaluating each partial feature data set separately according to a model generated by machine learning using the partial feature data set; determining, for each detected failure mode, which one or more performance criteria are associated with the failure mode based on the failure mode's associated partial feature data set, where determining includes aggregating the evaluations and prioritizing failure reasons associated with each detected failure mode; method. 48. The method of clause 47, wherein aggregating the evaluations includes aggregating the evaluations based on a performance criterion associated with each partial feature data set. 49. The method of clause 48, further comprising, if at least one failure mode associated with a performance criterion is detected, labeling that performance criterion for further evaluation, regardless of whether another failure mode associated with another performance criterion is detected.

[0101]

[0111] The above description includes examples of several embodiments. It is to be understood that for purposes of describing these embodiments, it is not possible to describe every conceivable combination of components or methods, but one skilled in the art will recognize that many more combinations and permutations of the various embodiments are possible. Accordingly, the described embodiments are intended to embrace all such modifications, improvements, and variations that fall within the spirit and scope of the appended claims. Furthermore, wherever used in the detailed description or claims, the term "including" is intended to be inclusive, in the same manner as the term "comprising" is interpreted when employed as a transitional term in the claims. Furthermore, although elements of the described aspects and / or embodiments may be described or claimed in the singular, the plural is also intended unless expressly stated to be singular. In addition, all or a portion of any aspect and / or embodiment may be utilized with all or a portion of any other aspect and / or embodiment, unless otherwise specified.

[0102]

[0112] Other implementations are within the scope of the claims.

Claims

1. 1. An apparatus for maintaining a light source including one or more modules configured together to generate a light beam for semiconductor photolithography, comprising: A prediction unit, receiving a plurality of partial characteristic data sets, each partial characteristic data set being associated with a unique performance metric of the light source; and a prediction unit configured to evaluate, for each partial feature data set, in a prediction increment, whether a failure mode is detected in the light source; An aggregate unit, receiving a plurality of estimates from the prediction unit; and and an aggregation unit configured to determine, for each failure mode, which one or more performance criteria are associated with said failure mode based on the associated partial feature data set of said failure mode. Device.

2. The apparatus of claim 1 , wherein the prediction unit includes or has access to multiple models, each model generated by machine learning using a unique partial feature dataset.

3. The apparatus of claim 2 , wherein each of the models in the prediction unit has an identical matching structure.

4. The apparatus of claim 2 , wherein each of the models in the prediction unit has a different fitting structure.

5. The apparatus of claim 2 , wherein at least one of the models in the prediction unit has a matching structure that is different from a matching structure of another model in the prediction unit.

6. The apparatus of claim 1 , wherein the plurality of partial feature data sets comprises a partial feature data set for a bandwidth of the light beam, a partial feature data set for a wavelength of the light beam, and a partial feature data set for an energy of the light beam.

7. 2. The apparatus of claim 1, wherein the plurality of partial feature data sets include a partial feature data set for a long-term bandwidth of the light beam, a partial feature data set for a short-term bandwidth of the light beam, a partial feature data set for bandwidth error events, a partial feature data set for a long-term wavelength of the light beam, a partial feature data set for a short-term wavelength of the light beam, a partial feature data set for wavelength error events, a partial feature data set for a long-term energy of the light beam, a partial feature data set for a short-term energy of the light beam, and a partial feature data set for energy error events.

8. 2. The apparatus of claim 1, wherein the aggregation unit is configured to aggregate the assessments from the prediction units and determine which one or more performance criteria are relevant to each detected failure mode based on the associated partial feature data set by prioritizing failure reasons associated with the failure mode.

9. The apparatus of claim 8 , wherein the aggregation unit is configured to aggregate evaluations from the prediction units based on the performance criterion associated with each partial feature dataset.

10. The apparatus of claim 9 , wherein the aggregation unit is configured to label a performance criterion for further evaluation if at least one failure mode associated with the performance criterion is detected.

11. 10. The apparatus of claim 9, wherein the aggregation unit is configured to label a performance criterion for further evaluation if at least one failure mode associated with the performance criterion is detected, regardless of whether another failure mode associated with another performance criterion is detected.

12. 2. The apparatus of claim 1, wherein the prediction unit is configured to evaluate whether the failure mode is detected in the light source at the prediction increment, comprising: evaluating whether the failure mode is detected in a particular module of the light source based on a model trained for the particular module.

13. The apparatus of claim 12 , wherein the particular module of the light source comprises a master oscillator module.

14. 2. The apparatus of claim 1, wherein the prediction unit is configured to evaluate whether the failure mode is detected in the light source at the prediction increment, comprising: evaluating whether the failure mode is detected in one or more specific modules of the light source based on a plurality of models each trained for a specific module.

15. 15. The apparatus of claim 14, wherein the one or more particular modules of the light source include one or more of a master oscillator module, a power amplifier module, a line narrowing module, a spectral feature analysis module, and a pulse stretcher module.

16. The prediction unit is configured to evaluate whether the failure mode is detected in the light source at the prediction increment, separately evaluating each partial feature data set according to a model generated by machine learning using the partial feature data set; and for each sub-feature dataset evaluation, outputting a score on the probability of failure.

2. The apparatus of claim 1.

17. The apparatus of claim 16, wherein the score is a certainty score, a binary output, or a function ranging from 0 to 1.

18. The apparatus of claim 1 , wherein the prediction unit is configured to evaluate, for each partial feature data set, whether a failure mode is detected in the light source at the prediction increment by performing an evaluation after each evaluation increment.

19. 20. The apparatus of claim 18, wherein the evaluation increment is 100 million pulses of the light beam.

20. 1. An apparatus for maintaining a light source including one or more modules configured together to generate a light beam for semiconductor photolithography, comprising: A pipeline unit comprising: receiving a plurality of partial characteristic data sets, each partial characteristic data set associated with a unique performance criterion of the light source during operation; and a pipeline unit configured to classify each of the received partial feature data sets into a feature category based on an associated specific performance criterion; A prediction unit, receiving the plurality of partial feature data sets and an assigned category for each partial feature data set; and a prediction unit configured to evaluate, at a prediction increment, whether a failure mode is detected in the light source; and an aggregation unit configured to determine which one or more performance criteria are associated with the failure mode based on the category assigned to each partial feature data set. Device.

21. 21. The apparatus of claim 20, wherein the prediction unit is configured to evaluate whether the failure mode is detected in the light source at the prediction increment comprises evaluating whether the failure mode is detected in a particular module of the light source based on a plurality of models trained for the particular module.

22. The apparatus of claim 21 , wherein each model is grown by machine learning by supplying a respective partial feature dataset to train the model.

23. 22. The apparatus of claim 21, further comprising an alert generation unit configured to prescribe a maintenance action for the particular module if the evaluation determines that the faulty module is detected in the particular module.

24. The prediction unit is configured to evaluate whether the failure mode is detected in the light source at the prediction increment, separately evaluating each partial feature data set according to a model generated by machine learning using the partial feature data set; and for each sub-feature dataset evaluation, outputting a score on the probability of failure.

21. The apparatus of claim 20.

25. 25. The apparatus of claim 24, wherein the aggregation unit is configured to determine which one or more performance criteria are associated with the failure mode based on the assigned category of each partial feature data set comprises analyzing the scores from the partial feature data sets where the same assignment is associated with a category.

26. The apparatus of claim 24, wherein the score is a certainty score, a binary output, or a function ranging from 0 to 1.

27. 21. The apparatus of claim 20, wherein the prediction unit is configured to evaluate whether the failure mode is detected in the light source at the prediction increment, comprising determining whether at least one module in the light source does not have at least a minimum probability of operating without failure at the prediction increment.

28. 21. The apparatus of claim 20, wherein the predicted increment is measured as a number of pulses of the light beam.

29. 21. The apparatus of claim 20, wherein prior to each evaluation, the pipeline unit is further configured to, in use, aggregate data of at least one partial feature dataset across one or more look-back increments preceding the evaluation.

30. 30. The apparatus of claim 29, wherein the prediction unit being configured to evaluate whether the failure mode is detected in the light source at the prediction increment comprises: evaluating each partial feature dataset separately according to a model generated by machine learning using the partial feature dataset; and, in use, if the partial feature datasets are aggregated across one or more lookback increments, using the aggregated partial feature dataset.

31. 31. The apparatus of claim 30, wherein each look-back increment is selected from a set of look-back increments assumed to be related to how the performance criterion changes as usage progresses.

32. 32. The apparatus of claim 31 , wherein the set of possible lookback increments includes a first lookback increment of approximately 2 billion pulses, a second lookback increment of approximately 1 billion pulses, and a third lookback increment of approximately 100 million pulses.

33. 31. The apparatus of claim 30, wherein one or more of the partial feature data sets are associated with performance criteria tracked at relatively short lookback increments, and wherein one or more of the partial feature data sets are associated with performance criteria tracked at relatively long lookback increments.

34. 30. The apparatus of claim 29, wherein the pipeline unit is configured, in use, to aggregate one or more partial feature datasets based on a same look-back increment, and the prediction unit is configured to receive the multiple partial feature datasets by receiving an aggregation of partial feature datasets that are tracked using the same look-back increment.

35. 1. A method for maintaining a light source including one or more modules configured together to generate a light beam for semiconductor photolithography, comprising: receiving a plurality of partial characteristic data sets, each partial characteristic data set being associated with a unique performance metric of the light source; For each partial feature data set, evaluating in a prediction increment whether a failure mode is detected in the light source; and for each detected failure mode, determining which one or more performance criteria are associated with the failure mode based on the associated partial feature data set of that failure mode. method.

36. and after receiving the plurality of partial feature data sets, classifying each of the received partial feature data sets into a feature category based on an associated specific performance criterion; determining which one or more performance criteria are associated with the failure mode includes determining which one or more performance criteria are associated with the failure mode based on the assigned category of each partial feature data set.

36. The method of claim 35.

37. Evaluating whether a failure mode is detected in the light source at a prediction increment includes: separately evaluating each partial feature data set according to a model generated by machine learning using the partial feature data set; and for each sub-feature dataset evaluation, outputting a score on the probability of failure.

36. The method of claim 35.

38. 36. The method of claim 35, wherein evaluating whether a failure mode is detected in the light source at the prediction increment comprises performing an evaluation after each evaluation increment.

39. 36. The method of claim 35, wherein determining which one or more performance criteria are associated with the failure mode based on the associated partial feature data set of the failure mode includes aggregating the evaluations and prioritizing failure reasons associated with each detected failure mode.

40. 40. The method of claim 39, wherein aggregating the ratings comprises aggregating the ratings based on the performance metric associated with each partial feature data set.

41. 41. The method of claim 40, further comprising labeling a performance criterion for further evaluation if at least one failure mode associated with the performance criterion is detected.

42. 41. The method of claim 40, further comprising, if at least one failure mode associated with a performance criterion is detected, labeling that performance criterion for further evaluation, regardless of whether another failure mode associated with another performance criterion is detected.

43. 36. The method of claim 35, wherein in the prediction increment, evaluating whether the failure mode is detected in the light source comprises evaluating whether the failure mode is detected in a particular module of the light source based on a model trained for the particular module.

44. 36. The method of claim 35, wherein in the prediction increment, evaluating whether the failure mode is detected in the light source comprises evaluating whether the failure mode is detected in one or more particular modules of the light source based on a plurality of models each trained for a particular module.

45. 36. The method of claim 35, further comprising, prior to each evaluation, aggregating, in use, data of at least one partial feature dataset across one or more look-back increments preceding the evaluation.

46. 46. ​​The method of claim 45, wherein in use aggregating data of at least one partial feature dataset across one or more look-back increments comprises, in use, aggregating one or more partial feature datasets based on a same look-back increment, and receiving the multiple partial feature datasets comprises receiving an aggregation of partial feature datasets tracked using the same look-back increment.

47. 1. A method for maintaining a light source including one or more modules configured together to generate a light beam for semiconductor photolithography, comprising: receiving a plurality of partial characteristic data sets, each partial characteristic data set being associated with a unique performance metric of the light source; evaluating, for each partial feature data set, whether a failure mode is detected in the light source in a prediction increment, comprising evaluating each partial feature data set separately according to a model generated by machine learning using the partial feature data set; determining, for each detected failure mode, which one or more performance criteria are associated with the failure mode based on the associated partial feature data set of the failure mode, said determining including aggregating the evaluations and prioritizing failure reasons associated with each detected failure mode; method.

48. 48. The method of claim 47, wherein aggregating the ratings includes aggregating the ratings based on the performance metric associated with each partial feature data set.

49. 49. The method of claim 48, further comprising, if at least one failure mode associated with a performance criterion is detected, labeling that performance criterion for further evaluation, regardless of whether another failure mode associated with another performance criterion is detected.

Images