Method for maintaining light source module in semiconductor lithography

Abstract

A computer-implemented method for maintaining a light source includes, using a computer, learning, for each of M machine learning models, where each model is trained to classify a state of a module of the light source as requiring or not requiring maintenance, to (1) form a pair of performance parameters in a set of N performance parameters, (2) scoring each of the pairs to produce a pair score, and (3) summing the scores to produce a model score; the model with the highest score is also implemented, and the state of a particular module of a particular light source is repeatedly classified over time as maintenance is required or not required using the implemented model; and performing maintenance of the particular module when indicated by the model, where M is an integer greater than 1, and N is an integer greater than 2.

Description

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims priority to U.S. application No. 63 / 482,094, filed on January 30, 2023, entitled “PROCESSES FOR MAIN-TENANCE OF MODULES OF LIGHT SOURCES IN SEMICONDUCTOR PHOTO-LITHOGRAPHY,” which is hereby incorporated by reference in its entirety. Technical Field

[0003] The disclosed subject matter relates to the maintenance of light sources, such as those used in integrated circuit lithography manufacturing processes. Background Art

[0004] The light used for semiconductor lithography, which can be laser radiation, is typically provided by a system called a light source. These light sources generate radiation as a series of pulses at a specified repetition rate (e.g., in the range of about 500 Hz to about 6 kHz). Additionally, such light sources typically have an expected lifetime, measured in terms of the number of pulses they can produce before requiring repair or replacement, typically expressed in billions of pulses.

[0005] A system for generating light, such as laser radiation, at a frequency useful for semiconductor lithography (such as at deep ultraviolet (DUV) wavelengths) involves the use of 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 light source as a whole can be considered a whole of modules. Each module typically has a lifetime that is shorter than the lifetime of the entire light source. Therefore, during the life of the light source, the health of each module is evaluated to determine whether the module should be repaired or replaced, and the modules are repaired or replaced based on this evaluation. Summary of the Invention

[0006] In some general aspects, a computer-implemented method for maintaining a light source includes using a computer, for each of M machine learning models, wherein each model is trained to classify a status of a module of the light source as requiring maintenance or not requiring maintenance, and wherein each model has a corresponding set of N performance parameters representing the performance of the model, to (1) form pairs of performance parameters in the set of N performance parameters, (2) score each pair to produce a pair score, and (3) sum the pair scores to produce a model score; also implementing the model with the highest score and using the implemented model to repeatedly classify the status of a particular module of a particular light source as requiring maintenance or not requiring maintenance over time; and automatically or manually performing maintenance on the particular module in response to the implemented model classifying the particular module as requiring maintenance, wherein M is an integer greater than 1 and N is an integer greater than 2.

[0007] Implementation of the computer-implemented method may include one or more of the following. The method may include first generating and training M models using a data set. The method may include, at some time after performing the method, repeating the method using a new data set. The new data set may include new data from a particular module. The new data set may include new data from multiple modules and multiple light sources. The method may include at least one of the N performance parameters being a performance parameter that is not used to generate and train the M models. Forming pairs of performance parameters for the set of N performance parameters may include forming pairs of performance parameters from performance parameters that are equal to or adjacent to each other in a priority ranking of the set of N performance parameters. Forming pairs of performance parameters for the set of N performance parameters may include forming ordered pairs ranked according to relative priorities of the paired respective performance parameters. Forming pairs of performance parameters for the set of N performance parameters may include forming N-1 pairs. Forming pairs of performance parameters for the set of N performance parameters may include forming ordered pairs ranked according to relative priorities of the paired respective performance parameters.

[0008] Scoring each pair may include, for each parameter, normalizing each parameter relative to the standard range and relative to the target of the corresponding parameter by adjusting the range of each parameter and / or shifting the range of each parameter, if necessary, to match the standard range and by subtracting the target of the parameter from each parameter; and, for each pair, taking the square root of the sum of the squares of the normalized parameters for the pair. Normalizing each parameter may include, for parameters in percentage form, subtracting the parameter from the target percentage of the parameter, the result of the subtraction being the normalized parameter, and for parameters not in percentage form, compressing, expanding, and / or shifting the range of the parameter to a value within the range of 0 to 1 and subtracting the parameter from the target value of the parameter. Scoring each pair may further include, for each pair, multiplying the square root of each pair by C1 when both parameters are equal to or better than their respective targets, multiplying the square root of each pair by negative C2 (-C2) when both parameters are worse than their respective targets, multiplying the square root of each pair by negative C3 (-C3) only when the higher priority parameter of the pair is worse than its corresponding target, and multiplying the square root of each pair by negative C4 (-C4) only when the lower priority parameter of the pair is worse than its corresponding target, where C2>C3>C4>0 and C1>0. In this method, the constants may have values ​​C1=C4=1, C2=0.75 and C3=0.5. Summing the pair scores to generate the score for the corresponding model may include a weighted summation.

[0009] The details of one or more implementations are set forth in the accompanying drawings and the description below. Other features will be apparent from the description and drawings, and from the claims. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 are schematic cross-sectional views of various aspects of a light source.

[0011] Figure 2 Such as Figure 1 Schematic diagram of a light source together with a photolithography exposure device.

[0012] Figure 3 is a flow chart of various aspects of the model selection process.

[0013] Figure 4 is a flow chart of various aspects of the automatic model selection process.

[0014] Figure 5A 、 Figure 5B 、 Figure 5C and Figure 5D It involves such Figure 4 A diagram of one implementation of the scoring process within the scoring process.

[0015] Figure 6 It involves such Figure 4A diagram of one implementation of the scoring process within the scoring process.

[0016] Figure 7 Such as Figure 4 A flowchart of one implementation of the scoring process within the scoring process.

[0017] Figure 8 Such as Figure 7 A flowchart of one implementation of a portion of the scoring process. DETAILED DESCRIPTION

[0018] refer to Figure 1 The light source 100, which may be a deep UV (DUV) light source 100, may be in the form of a dual-stage pulsed light source that produces a pulsed amplified light beam 105 as a light beam 105. The light source 100 includes a solid-state or gas discharge master oscillator (MO) system 160, a power amplification (PA) system such as a power ring amplifier (PRA) system 165, relay optics 170, and a light output subsystem 175.

[0019] The MO system 160 may include, for example, a MO chamber module 161, wherein a discharge between electrodes (not shown) may induce a laser gas discharge in a laser gas to produce an inverse population of high-energy molecules, such as argon, krypton, or xenon, to produce relatively broadband radiation that is line-narrowed to a relatively very narrow bandwidth and a center wavelength selected in a line narrowing module ('LNM') 162. The MO system 160 may also include a MO output coupler (MO OC) 162, which may include a partially reflecting mirror that, together with a reflective grating (not shown) in the LNM 162, forms an oscillator cavity in which the MO system 160 oscillates to form a seed output pulse, thereby forming a master oscillator. The MO system 160 may also include a line center analysis module (LAM) 163. As described above, the LAM 180 may include, for example, an etalon spectrometer for fine wavelength measurement and a grating spectrometer for coarser resolution.

[0020] The relay optics 170 may include a MO wavefront engineering box (WEB) 171 for redirecting the output of the MO system 160 toward the PA system 165 and may include beam expansion, for example using a multi-prism beam expander (not shown) and coherence destruction, for example in the form of an optical delay path (not shown).

[0021] The PA system 165 includes a PRA chamber module 166, which is also an oscillator formed, for example, by injecting the output beam from the MO system 160, and output coupling optics (not shown). The output coupling optics can be incorporated into a PRAWEB 167 and can be redirected back through the gain medium in the chamber 166 via a beam reverser 168. The PRA WEB 167 can incorporate a partially reflective input / output coupler (not shown) and a mirror for maximum reflection at the nominal operating wavelength (which can be approximately 193 nm for an ArF system) and one or more prisms. The PA system 165 optically amplifies the output beam from the MO system 160.

[0022] The optical output subsystem 175 can include a bandwidth analysis module (BAM) 176 at the output of the PA system 165 that receives the pulsed output beam from the PA system 165 and picks off a portion of the beam for metrology purposes, such as measuring the output bandwidth and pulse energy. The pulsed output beam then passes through an optical pulse stretcher module (OPuS) 177 and an output combined automatic shutter metering module (CASMM) 178, which can also be the location of a pulse energy meter. One purpose of the OPuS 177 can be to convert a single output pulse into a pulse train. The secondary pulses generated from the original single output pulse can be delayed relative to each other. By distributing the original laser pulse energy into a series of secondary pulses, the effective pulse length of the beam can be extended while reducing the peak pulse intensity.

[0023] The light source 100 is composed of modules. Each component of the light source 100 (such as the MO chamber 161, the LNM 162, the MO WEB 171, the PRA chamber 166, the PRA WEB 167, the OPuS 177, the BAM 176) is a module. The overall availability of the light source 100 is a direct result of the respective availability of these individual modules that make up the light source 100. In other words, the light source 100 is not available unless all of these modules that make up the light source 100 are available. The sensing and evaluation device 120 monitors these modules so that they can be refreshed or replaced before they fail in order to maintain the operation of the light source 100 and optimize and improve the productivity of the output device 210. The sensing and evaluation device 120 provides maintenance alerts that can be used to perform automatic and / or manual maintenance tasks for one or more specific modules, including replacement tasks when necessary to prevent failure of the light source 100.

[0024] Reference Figure 2 , the amplified light beam 105( Figure 1) can be used as a light beam 205 used by a lithographic exposure apparatus 210 to pattern features on a substrate or wafer 211. Wafer 211 is placed on a wafer stage 212, which is configured to hold wafer 211 and is connected to a positioner configured to precisely position wafer 211 according to certain parameters. Light beam 205 can have a wavelength in the deep ultraviolet (DUV) range, which can include, for example, wavelengths from about 100 nanometers (nm) to about 400 nm. For example, a light source 100 that generates such light beam 105, 205 can be a gas discharge light source, such as an excimer light source, or an excimer laser that uses a combination of one or more rare gases and a reactive gas as a gain medium, the rare gas including argon, krypton, or xenon, and the reactive gas including fluorine or chlorine. Light source 100 can be an excimer light source. Thus, for example, the gain medium can include argon fluoride (ArF), krypton fluoride (KrF), or xenon chloride (XeCl). If the gain medium comprises argon fluoride, the wavelength of amplified light beam 205 is approximately 193 nm, and if the gain medium comprises krypton fluoride, the wavelength of amplified light beam 205 is approximately 248 nm. The size of the microelectronic features patterned on wafer 211 depends on the wavelength of light beam 205, with lower wavelengths resulting in smaller minimum feature sizes. When the wavelength of light beam 205 is 248 nm or 193 nm, the minimum size of the microelectronic features can be, for example, 50 nm or less. The bandwidth of light beam 205 can be the actual instantaneous bandwidth of its optical spectrum (or emission spectrum), which contains information about how the optical energy of light beam 205 is distributed across different wavelengths.

[0025] The lithographic exposure apparatus 210 includes an optical arrangement including, for example, one or more converging lenses, a mask, and an objective lens arrangement. The mask can be moved in one or more directions, such as along the optical axis of the light beam 205 or in a plane perpendicular to the optical axis. The objective lens arrangement includes a projection lens and enables the image to be transferred from the mask to the photoresist on the wafer 211. The lithographic exposure apparatus 210 also includes an illumination system that adjusts the angular range of the light beam 205 impinging on the mask. The illumination system also homogenizes (makes uniform) the intensity distribution of the light beam 205 across the mask.

[0026] Among other features, the lithography exposure apparatus 210 may include a lithography controller 213 that controls how layers are printed on the wafer 211. The lithography controller 213 includes a memory that stores information such as a process recipe. The process or recipe determines the exposure length on the wafer 211, the mask used, and other factors that affect the exposure. During lithography, multiple pulses of the light beam 205 illuminate the same area of ​​the wafer 211 to collectively constitute an exposure dose.

[0027] The quality of features produced on wafer 211 by photolithography exposure apparatus 210 is directly dependent on the quality and reliability of the light pulses from light source 100. Pulses with less than required power can result in underexposure of areas of wafer 211. Missing pulses can also result in underexposure. Shifts in wavelength or bandwidth distribution can result in shifts in image position and changes in the pattern produced on wafer 211.

[0028] Information from various sources in the light source 100 can be used to assess the need for maintenance or replacement of modules of the light source 100, such as the MO chamber module 161 or the PRA chamber module 166. These requirements may require, for example, replacement of gases, or even replacement of the entire module if module performance (and therefore light source performance) begins to degrade.

[0029] It is important to perform maintenance, such as changing gases or replacing modules, at the right time. Maintenance done too early reduces the portion of time the light source can be used for productive operation. Maintenance done too late can result in incorrectly processed products - products that may have very high sunk production costs and may have to be scrapped.

[0030] Machine learning can be applied to determine an appropriate time to perform maintenance. Data from various parts of the light source, and potentially even data from associated lithographic equipment, can be used by the machine learning model to determine a reasonable ideal time to maintain the light source module. Machine learning and the execution of the machine learning model can be performed by various computing hardware. In various implementations, a computing module configured to perform machine learning and / or execute a machine learning module may include an electronic processor, an electronic storage device, and an I / O interface. The electronic processor includes one or more processors suitable for executing computer programs, such as general-purpose or special-purpose microprocessors, and any one or more processors of any type of digital computer. Typically, the electronic processor receives instructions and data from read-only memory, random access memory, or both. The electronic storage device can be volatile memory, such as RAM, or non-volatile memory. In some implementations, the electronic storage device includes non-volatile and volatile parts or components. The electronic storage device can store data and information used to learn or execute the model, components of the control system, and / or other data or instructions. This information can be stored, for example, in a lookup table or database.

[0031] A two-class machine learning classification model can be generated and trained on data generated from the operation of existing light sources used in production and / or from the operation of prototype light sources or new light source designs in a test program. Using standard machine learning techniques, the classification model can be trained to distinguish between two categories or states, a state requiring maintenance and a state not requiring maintenance. For example, a state requiring maintenance can be defined as a state from which a specified fault or a specified exceedance condition of a module occurs within a specific number of light pulses of the light source or within a specific amount of time (in the training data set). Similarly, a state not requiring maintenance can be defined as a state from which a specified fault or a specified exceedance condition of a module does not occur within a specified time (in the training data set). Optionally, additional instances of artificial faults or exceedances can be generated, for example, by extrapolating from existing data and adding to the training data set. This can help improve the performance of the trained model, as two-class classification models generally perform best when trained with an approximately equal number of instances of each class (maintenance required and maintenance not required).

[0032] Using machine learning tools for various types of models (such as classification and regression trees (CART), random forests, feed-forward neural networks, etc., including combinations), multiple models can be automatically generated and trained using a training data set. The multiple models can be ranked using any of a variety of statistics, such as, for example: (1) a true positive (TP) rate, which is the percentage of "positives" that are "true positives," where a "positive" is a classification (determination) made by the model that maintenance is required, and a "true positive" is a correct classification (determination) made by the model that maintenance is required (i.e., a classification [determination] that maintenance is required within a specified number of light pulses or within a specified amount of time within the training data set, where maintenance is actually required or will be required), and (2) a true negative (TN) rate, which is the percentage of "negatives" (classifications / determinations that "no maintenance is required") that are true (i.e., "correct") negatives. Thus, a method is needed to select which model to implement when performing maintenance on a module of a light source.

[0033] Additional complexities may arise when, based on business or engineering needs or other considerations, it is desirable to include in the model selection process factors ("data") that were not used in the model training process. Such data may be in the form of data relating to new or changing metrics that are or will be used in the management of the light source, or metrics that are not easily represented in the classification model, or a combination of such metrics. For example, the number of light pulses from a light source that will be "lost" (i.e., not produced) between the time of a "maintenance required" classification and the actual or expected failure of the light source, minus some safety interval, may be referred to as "lost pulses." The lost pulses metric may not be easily represented in the classification model, or may be a new or frequently changing metric, and therefore may not be a metric involved in model training.

[0034] Figure 3 and Figure 4 Flowcharts are shown representing aspects of a process that may be automatically performed to select a model from a plurality of models based on parameters included or inherent in the model training process, such as TP% and TN%, and parameters that may not be used in the model training process, such as missing pulses.

[0035] Figure 3 A preliminary process 320 is shown that includes at least two steps. The first step is to collect requirements in the form of parameters for evaluating the model (and the performance of the model for maintenance purposes) (321). For example, the parameters may be TP%, TN%, and lost pulses (LP). Collecting the requirements also includes collecting target values ​​for each parameter and collecting or assigning and applying utility functions or prioritization of the parameters. The goals and the relative priorities of the parameters and / or associated goals may be determined by the goals or requirements of the associated production process and / or the business associated therewith. For example, LP may be more important than TP%, which may be more important than TN%. The goals may be specified as a number of pulses (such as, for example, 20,000) or a time (e.g., 2 days or 7 days), with TP% and TN% specified as percentages (such as 60% and 95%, respectively). The second step is to select a pairing method and, optionally, a weighting scheme (322). The performance parameters may be paired by forming pairs of those performance parameters that are adjacent to each other in the prioritization. For example, if the performance parameters are TP%, TN%, and LP, and the order of priority from highest to lowest is LP, TP, TN, then the pairs may be formed as (LP, TP) and (TP, TN), resulting in N-1 pairs for the N parameters. Alternatively, an additional pair may be formed from the lowest and highest ranked performance parameters (e.g., (LP, TN)) and included in the total pairs, resulting in N pairs for the N performance parameters. Other suitable pairing methods may be used, including methods that produce other total numbers of pairs, if desired. The pairs may also be in the form of ordered pairs that are ranked according to the relative priority of the respective performance parameters being paired. By identifying and ranking the performance parameters and by selecting a pairing method, a method such as Figure 4 The automation process is shown in the flowchart of the automation process.

[0036] Figure 44 is a flow chart of a computer-implemented automated process 430 for maintaining a light source. After generating (431) a plurality of models (a total of M models) for evaluation, for each of the M models, the performance parameters (of the set of N performance parameters) are normalized and paired, and each pair is scored to produce a "pair score," which is then summed to produce a model score (432). Next, the model with the highest score is selected (433) and implemented (i.e., its solution is implemented) (434) for maintaining at least one particular module of a particular light source, becoming the "implemented module." The implemented model is then used to repeatedly classify the particular module over time. If or when the implemented module classifies the particular module as requiring maintenance (435, "yes" branch), the required maintenance action is performed (436). The action can be automatically performed. The process can then return to repeatedly classifying the particular module (435). If the classification is "not requiring maintenance" (435, "no" branch), the process can simply return to classify the particular module again (435). If desired, a time delay (not shown) may be implemented between repeated classification steps. After at least some of the repeated classification steps (435) and / or steps of performing required maintenance actions (436) (indicated by dashed arrows), or alternatively, after each instance of these steps (indicated by solid arrows without following dashed arrows), an evaluation (437) may be performed to determine whether a model update is appropriate. For example, after a certain number of classifications (435) have been performed, or after a certain number of maintenance actions (436) have been performed, or after a certain amount of time has elapsed, or after a certain amount of new data has been collected from the operation of a particular light source and module and / or additional similar light sources and modules, a model update may be appropriate so that a new data set (including or even consisting solely of the new data) can be used for model development and training, or based on a combination of these or other factors. If a model update is determined to be appropriate (437, "yes" branch), the process begins again, generating multiple models for evaluation (431).

[0037] For each performance parameter, when the performance parameter is numerically greater than a target, such as, for example, a bowling score, the performance parameter may be preferred over the target or objective for that parameter. Alternatively, when the performance parameter is numerically less than the target, such as, for example, a golf score, the performance parameter may be preferred over the target or objective for that parameter. For example, TP% and TN% are preferred when they are greater than their respective targets, whereas lost pulses (LP) should be minimized, so LP is preferred when it is less than its target.

[0038] For a pair of performance parameters P1, P2, where P1 is the higher priority parameter of the two and is represented on the vertical axis of a grid, and where P2 is represented on the horizontal axis of a grid, where the origin of the grid represents the targets for both parameters, Figures 5A-5D The result expressed in is produced as follows: if P1 and P2 are both better when they are both greater than their targets (this can be referred to as high P1 / high P2 or simply high / high), then Figure 5A As a result, the preferred quadrant when plotting the pair is the upper right, or Quadrant I. If P1 is better when it is above its target, and P2 is better when it is below its target (high / low), then Figure 5B The result shows the situation where the "good-good" quadrant is in the upper left (quadrant II). If P1 and P2 are both better than their corresponding targets when both are less than the target (low / low), then Figure 5C The result is that the lower left of the grid (Quadrant III) is the good / good quadrant because P1 and P2 are "good" in this quadrant, as shown. If P1 is better when it is below its target and P2 is better when it is above its target (low / high), then Figure 5D The results show the situation where the "good-good" quadrant is in the lower right (Quadrant IV). Figures 5A-5D The classification shown can be used to score pairs of performance parameters for a given model, as shown below. Figure 6 and Figure 7 Explained.

[0039] Figure 6 Shown with Figure 5A , which corresponds to the grid of , shows a portion of the process for scoring high / high pairs (pairs in which both parameters are better when greater than their targets). The scoring method for P1, P2 can include finding the radius R from the origin of the grid to the point (P2, P1) (or in other words, finding the square root of the sum of the squares of P1 and P2), and then multiplying by a factor that depends on the quadrant in which point P is located. For example, for P in Quadrant I (or on the axis), R can be multiplied by 1. For P in Quadrant III, R can be multiplied by -1 (because both parameters are worse than the targets there). For P in Quadrant II, R can be multiplied by -0.5, and for P in Quadrant IV, R can be multiplied by -0.75 because the high priority parameter (P1) being worse than its target (as shown in Quadrant III) can be considered a more serious situation than the situation in which the low priority parameter (P2) is worse than its target (as shown in Quadrant II). Figure 7 The flowchart shows the Figure 6 The procedure is extended to all combinations of performance parameters within a pair, along with additional details of the scoring procedure.

[0040] Figure 7is a flow chart of a process or subprocess 732 for scoring a model having a set of N performance parameters grouped into pairs of performance parameters. For example, process 732 may be used as Figure 4 Step or process 432 within process 430 of . First, for a given model, for a first pair of performance parameters P1, P2, the parameters are normalized and R is set to the square root of the sum of the squares of P1 and P2 (741). Normalizing each parameter may include, for parameters in percentage form, subtracting the parameter from a target percentage for the parameter, the result of the subtraction being the normalized parameter, and for parameters in non-percentage form, compressing or expanding and / or shifting the range of the parameter to a value in the range of 0 to 1 and subtracting the parameter from the target value for the parameter. Next, based on P1, P2 being classified as high / high (742 "yes" branch), high / low (743 "yes" branch), low / low (744 "yes" branch) or low / high (744 "no" branch), the quadrant of the point (P2, P1) listed in tables 745a, 745b, 7445c and 745d is selected based on the factors in the respective "quadrant" columns (see Figures 5A-5D ), multiplying the previously obtained value of R by the factors shown in the factor selection table 745a-745d in the respective "score Si" column. ("Si" represents the score of the i-th pair among the pairs being scored.) The various scores Si produced by this first part 748 of the process 740 can be summed directly (solid arrow to 752), or alternatively can be calculated by the various coefficients ωi-ω Ptot (shown in steps 751-1-751-Ptot) Scaling is performed for each of the pair scores Si (dashed arrows to 750). This process is repeated ("yes" branch of 753) until there are no more pairs (a total of "Ptot") to be scored ("no" branch of 753). The sum of the pair scores is then stored (754) as the score of the corresponding model. If there are more models to be scored ("yes" branch of 755), the above process is repeated until each of the M models has been scored. Then, it can be used as Figure 4 The process 732 of step 432 or process 432 may proceed to the next step, such as Figure 4 Step 433.

[0041] exist Figure 8A slightly generalized and simplified version of the first portion 748 of the process or subprocess 740 is shown in the formal process or subprocess 848g of FIG. The process or subprocess 848g uses constants C1, C2, C3, and C4 that satisfy the conditions (855) shown (i.e., C2>C3>C4>0 and C1>0) and classifies each pair of performance parameters (in 856a, 856b, and 856c) according to which performance parameter is better or worse than its corresponding target. If both parameters are better than their corresponding targets (856a "yes" branch), then R is multiplied by C1. If both parameters are worse than their corresponding targets (856b "yes" branch), then R is multiplied by negative C2 (-C2). If only the higher priority parameter of the pair is worse than its corresponding target (856c "yes" branch), then R is multiplied by negative C3 (-C3), and for the rest of the cases (only when the lower priority parameter of the pair is worse than its corresponding target), R is multiplied by negative C4 (-C4). Then it can be as follows Figure 7 The resulting pairwise scores are summed directly as in step 752 of process 740, or as Figure 7 Scaled as in the process or subprocess 750 and then summed. Figure 7 In the implementation of the coefficients, C1=C4=1, C2=0.75, C3=0.5.

[0042] Aspects and implementations of the present disclosure may be further described using the following terms:

[0043] 1. A computer-implemented method for maintaining a light source, the method comprising:

[0044] Using a computer:

[0045] For each of the M machine learning models, each model is trained to classify a status of a module of a light source as requiring maintenance or not requiring maintenance, and wherein each model has a corresponding set of N performance parameters representing performance of the model:

[0046] pairs of performance parameters forming the set of N performance parameters;

[0047] scoring each of the pairs to produce a pair score; and

[0048] summing the pair scores to produce a model score;

[0049] implementing the model with the highest score and using the implemented model to repeatedly classify the status of a particular module of a particular light source as requiring maintenance or not requiring maintenance over time; and

[0050] automatically or manually performing maintenance on the particular module in response to the implemented model classifying the particular module as requiring maintenance,

[0051] wherein M is an integer greater than 1, and N is an integer greater than 2.

[0052] 2. The computer-implemented method of claim 1 , further comprising: first generating and training the M models using a dataset.

[0053] 3 . The computer-implemented method of claim 2 , further comprising: at some time after executing the method, repeating the method using a new data set.

[0054] The computer-implemented method of claim 3 , wherein the new data set comprises new data from the particular module.

[0055] The computer-implemented method of claim 4 , wherein the new data set comprises new data from a plurality of modules and a plurality of light sources.

[0056] 6. The computer-implemented method of claim 2, wherein at least one of the N performance parameters is a performance parameter that is not used to generate and train the M models.

[0057] 7. The computer-implemented method of claim 1 , wherein forming pairs of performance parameters in the set of N performance parameters comprises forming pairs of performance parameters from performance parameters that are equal to or adjacent to each other in a priority ordering of the set of N performance parameters.

[0058] 8 . The computer-implemented method of claim 7 , wherein forming pairs of performance parameters in the set of N performance parameters comprises forming ordered pairs sorted according to relative priorities of the respective performance parameters being paired.

[0059] 9 . The computer-implemented method of claim 7 , wherein forming pairs of performance parameters in the set of N performance parameters comprises forming N−1 pairs.

[0060] 10. The computer-implemented method of claim 1, wherein forming pairs of performance parameters in the set of N performance parameters comprises forming ordered pairs sorted according to relative priorities of the respective performance parameters being paired.

[0061] 11. The computer-implemented method of claim 1 , wherein scoring each of the pairs comprises:

[0062] for each parameter, normalizing each parameter relative to the standard range and relative to the target for the corresponding parameter by adjusting the range of each parameter and / or shifting the range of each parameter if necessary to match the standard range and by subtracting the target for that parameter from each parameter; and,

[0063] For each pair, take the square root of the sum of the squares of the normalized parameters for the pair.

[0064] 12. The computer-implemented method of claim 11 , wherein normalizing each parameter comprises:

[0065] For a parameter in percentage form, subtracting the parameter from the target percentage of the parameter, the result of the subtraction being the normalized parameter; and

[0066] For parameters that are not in percentage form, the range of the parameter is compressed, expanded, and / or shifted to a value within the range of 0 to 1, and the parameter is subtracted from the target value of the parameter.

[0067] 13. The computer-implemented method of claim 11 , wherein scoring each or the pairs further comprises:

[0068] For each pair, multiply the square root of each pair by:

[0069] C1, when both parameters are equal to or better than their corresponding targets;

[0070] negative C2 (-C2), when both parameters are worse than their corresponding targets;

[0071] negative C3 (-C3), only if the higher priority parameter of the pair is worse than its corresponding target; and negative C4 (-C4), only if the lower priority parameter of the pair is worse than its corresponding target;

[0072] Among them, C2>C3>C4>0 and C1>0.

[0073] The computer-implemented method of claim 13 , wherein C1=C4=1, C2=0.75, and C3=0.5.

[0074] 15. The computer-implemented method of claim 1, wherein summing the pair scores to produce a corresponding score for the model comprises a weighted summation.

[0075] 16. A DUV light source configured to perform the computer-implemented method according to any of clauses 1-15.

[0076] 17. A module configured to be installed in a DUV light source and configured to perform the computer-implemented method according to any of clauses 1-15.

[0077] 18. A computing system configured to communicate with a DUV light source and configured to perform the computer-implemented method of any of clauses 1-15.

[0078] The above aspects and implementations and other implementations are within the scope of the following claims.

Claims

1. A computer-implemented method for maintaining a light source, the method comprising: Using a computer: For each of the M machine learning models, each model is trained to classify a status of a module of a light source as requiring maintenance or not requiring maintenance, and wherein each model has a corresponding set of N performance parameters representing performance of the model: pairs of performance parameters forming the set of N performance parameters; scoring each of the pairs to produce a pair score; as well as summing the pair scores to produce a model score; implementing the model with the highest score and repeatedly classifying the status of a particular module of a particular light source as requiring maintenance or not requiring maintenance over time using the implemented model; as well as automatically or manually performing maintenance on the particular module in response to the implemented model classifying the particular module as requiring maintenance, wherein M is an integer greater than 1, and N is an integer greater than 2.

2. The computer-implemented method of claim 1 , further comprising: First, the M models are generated and trained using the dataset.

3. The computer-implemented method of claim 2 , further comprising: At some time after executing the method, the method is repeated using a new data set. The computer-implemented method of claim 3 , wherein the new data set comprises new data from the particular module. The computer-implemented method of claim 4 , wherein the new data set comprises new data from a plurality of modules and a plurality of light sources.

6. The computer-implemented method of claim 2, wherein at least one of the N performance parameters is a performance parameter that is not used to generate and train the M models.

7. The computer-implemented method of claim 1 , wherein forming pairs of performance parameters in the set of N performance parameters comprises forming pairs of performance parameters from performance parameters that are equal to or adjacent to each other in a priority ordering of the set of N performance parameters. 8 . The computer-implemented method of claim 7 , wherein forming pairs of performance parameters in the set of N performance parameters comprises forming ordered pairs sorted according to relative priorities of the respective performance parameters being paired. 9 . The computer-implemented method of claim 7 , wherein forming pairs of performance parameters in the set of N performance parameters comprises forming N−1 pairs.

10. The computer-implemented method of claim 1, wherein forming pairs of performance parameters in the set of N performance parameters comprises forming ordered pairs sorted according to relative priorities of the respective performance parameters being paired.

11. The computer-implemented method of claim 1 , wherein scoring each of the pairs comprises: for each parameter, normalizing each parameter relative to the standard range and relative to the target for the corresponding parameter by adjusting the range of each parameter and / or shifting the range of each parameter if necessary to match the standard range and by subtracting the target for that parameter from each parameter; and, For each pair, take the square root of the sum of the squares of the normalized parameters for the pair.

12. The computer-implemented method of claim 11 , wherein normalizing each parameter comprises: For a parameter in percentage form, the parameter is subtracted from the target percentage of the parameter, and the result of the subtraction is the normalized parameter; as well as For parameters that are not in percentage form, the range of the parameter is compressed, expanded, and / or shifted to a value within the range of 0 to 1, and the parameter is subtracted from the target value of the parameter.

13. The computer-implemented method of claim 11 , wherein scoring each or the pairs further comprises: For each pair, multiply the square root of each pair by: C1, when both parameters are equal to or better than their corresponding targets; negative C2 (-C2), when both parameters are worse than their corresponding targets; Negative C3 (-C3), only if the higher priority parameter of the pair is worse than its corresponding target; and negative C4 (-C4), only if the lower priority parameter of the pair is worse than its corresponding target; Among them, C2>C3>C4>0 and C1>0. The computer-implemented method of claim 13 , wherein C1=C4=1, C2=0.75, and C3=0.

5.

15. The computer-implemented method of claim 1, wherein summing the pair scores to produce a corresponding score for the model comprises a weighted summation.

16. A deep ultraviolet (DUV) light source comprising: processor, and a memory comprising machine-executable instructions configured to, when executed by the processor, perform the following actions: For each of the M machine learning models, each model is trained to classify a status of a module of a light source as requiring maintenance or not requiring maintenance, and wherein each model has a corresponding set of N performance parameters representing performance of the model: pairs of performance parameters forming the set of N performance parameters; scoring each of the pairs to produce a pair score; as well as summing the pair scores to produce a model score; implementing the model with the highest score and using the implemented model to repeatedly classify the status of a particular module of a particular light source as requiring maintenance or not requiring maintenance over time; and in response to the implemented model classifying the particular module as requiring maintenance, automatically or manually performing maintenance on the particular module, wherein M is an integer greater than 1, and N is an integer greater than 2.

17. The DUV light source of claim 16, wherein the actions further comprise first generating and training the M models using a dataset.

18. The DUV light source of claim 17, wherein the steps further comprise: At some time after executing the method, the method is repeated using a new data set.

19. The DUV light source of claim 18, wherein the new data set comprises new data from the particular module.

20. The DUV light source of claim 19, wherein the new data set comprises new data from a plurality of modules and a plurality of light sources.

21. The DUV light source of claim 17, wherein at least one performance parameter of the N performance parameters is a performance parameter that is not used to generate and train the M models.

22. The DUV light source of claim 16, wherein forming pairs of performance parameters in the set of N performance parameters comprises forming pairs of performance parameters from performance parameters that are equal to or adjacent to each other in a priority order of the set of N performance parameters.

23. The DUV light source of claim 22, wherein forming pairs of performance parameters in the set of N performance parameters comprises forming ordered pairs sorted according to relative priorities of the respective performance parameters being paired.

24. The DUV light source of claim 22, wherein forming pairs of performance parameters in the set of N performance parameters comprises forming N-1 pairs.

25. The DUV light source of claim 16, wherein forming pairs of performance parameters in the set of N performance parameters comprises forming ordered pairs sorted according to relative priorities of the respective performance parameters being paired.

26. The DUV light source of claim 16, wherein scoring each of the pairs comprises: for each parameter, normalizing each parameter relative to the standard range and relative to the target for the corresponding parameter by adjusting the range of each parameter and / or shifting the range of each parameter if necessary to match the standard range and by subtracting the target for that parameter from each parameter; and, For each pair, take the square root of the sum of the squares of the normalized parameters for the pair.

27. The DUV light source of claim 26, wherein normalizing each parameter comprises: For a parameter in percentage form, the parameter is subtracted from the target percentage of the parameter, and the result of the subtraction is the normalized parameter; as well as For parameters that are not in percentage form, the range of the parameter is compressed, expanded, and / or shifted to a value within the range of 0 to 1, and the parameter is subtracted from the target value of the parameter.

28. The DUV light source of claim 26, wherein scoring each or the pairs further comprises: For each pair, multiply the square root of each pair by: C1, when both parameters are equal to or better than their corresponding targets; negative C2 (-C2), when both parameters are worse than their corresponding targets; Negative C3 (-C3), only if the higher priority parameter of the pair is worse than its corresponding target; and negative C4 (-C4), only if the lower priority parameter of the pair is worse than its corresponding target; Among them, C2>C3>C4>0 and C1>0.

29. The DUV light source of claim 28, wherein C1 = C4 = 1, C2 = 0.75, and C3 = 0.

5.

30. The DUV light source of claim 16, wherein summing the pair scores to produce corresponding scores for the model comprises a weighted summation.

Images