BACKGROUND OF THE INVENTION f Field of the In yen tion

The presets* invention generally relates to methods and systems for accelerated training of a machine teaming based model for semiconductor applications.

2. Description of the; Related. Art

The following description and examples are not admitted to be prior art by ^■ virtue of their inclusion in this section .

Fabricating semiconductor devices such as logic and memory devices, typically includes processing a substrate such as a semiconductor wafer using large number of semiconductor fabri.cati.on processes to form various features and multiple levels of the semiconductor devices. For example, lithography is a semiconductor fabrication process that involves- transferring a pattern from a. reticle to a. resi st arranged on a .sem c nduc o wafer. Additional examples of semiconductor fabrication processes include,, but are not limited to, chemical-mechanical polishing (CMP), etch, deposition, and ion implantation. Multiple semiconductor devices may ^' be fabricated in an arrangement on a single- semiconductor wafer and then separated into individual semiconductor devices. inspectio processes are used at various steps during- a semiconductor

manufachuing process to detect defects on wafers to drive higher yield in the

manufacturing process and thus higher profits, inspection; has -always been an important part offahrieating -semiconductor devices. However, as the dimensions of semiconductor devices decrease, inspection becomes even more important to the successful manufacture of acceptable semiconductor devices because smaller defects can cause the devices to .fail Defect review typically involves i¾Hieteeimg defects detected as such by an inspection process and generating additional information about the defect at a higher resolution using either a high magnification optical system or a scanning electron microscope (SEM), Defect review is therefore performed at discrete locations on the wafer where defects have been detected by inspection. The higher resolution data for the defects generated by defect review is more -suitable for determining attributes of the defects such as profile, roughness, more accurate size information, etc.

Metrology processes are also used at various steps during semiconductor manuiaeturing process to monitor and control the process. Metrology processes are different than inspection, processes in" that _* unlike inspection processes in which detects are detected on a wafer, metrolog processes are used to measure one or more

characteristics of the water that cannot be determined using currently used inspection tools. For example, metrology processes are used to measure one or more characteristics of a water such as a dimension (e.g,, line width, thickness, etc.) of features formed on the wafer daring .a process suclif that the performance of the process can b determined from the one or more characteristics, in addition , if the one or more _. , characteristics of the wa fer are -unacceptable (e«g„ oat of a predetermined range for the charaeieristic(s)> ₅ the measurements of the one or more characteristics of the wa ter may ^' be used to after One or more parameters of the process such that additi onal wa fers manufactured by the -process have acceptable

Metrology processes are also different than defect review processes in tha , unlike defect review processes m which defects that are detected by inspection are re-visited in defect review, metrology processes may be performed at locations at which no defect has been detected. In other words, unlike defect review, the locations at which a metrology process is performed on a wafer may be independent of the results of an inspection process performed on the wafer. In particular, the locations at which a metrology process is performed may be selected independently of inspection results. Current algorithms thai ar used in most semiconductor applications like

inspection, tend to require hand crafted features. There is teaming involved in the system, but it tends to happen in an ad hoc manner and. at a very high ('lossy") level where there is a loop typically involving a SEM review where the results of the inspection are deemed as true defects, false alarms,- nuisance, etc. B "lossy," we mean that the relationships between process-related causes and our ability to observe the effects are so complex that conventional methods do not lend themselves to converging on. recipes that take Ml ad vantage o f the raw capability of the metrology and inspection equipm ent used to control th e semiconductor manufacturing process .

It should also be noted that rigorous process simulation design of experiments (DOEs) through process window can also be employed. However, they are much slower and require post-optical proximity correction (OPC) data (input to mask writer), which is often not available in the fab. For example, a product named Design Sean, which was developed by KLA-Teoeor, Milpitas, Calif,, endeavored to predict the pattern variations that occur through the lithography process window.

Th field of neural network S eaming i s Uttered wi th decades of unsuccessful attempts at unsupervised learning. However, in the last "two years, we disco vered tha t some in the research community have had a fair amount of success in natural image scene classification in leveraging techniques related to transfer learning and serni-supervi-sed. Seaming for character . recognition, Deep learning models work well under the assumption that there are plenty of samples for the deep neural network to work on. When there are very few sample examples to work with, success with these methods tends to be quite limited. Accordingly, it would be advantageous, to develop systems and Methods- for training a machine .learning based model that do not have one or more of the

disadvantages described above.

SUMMAEV OF THE INVENTION

The following description of various emhodiraents is not to be construed m any way as limiting the subject matter of the appended c laims.

One embodiment relates to a system configured, to train a. machine learning based model. The system includes one or more computer subsystems and one or more

component executed, by the one or more computer subsystems. The one or more components include a machine learning based model configured for performing one or more simulations for specimens. The machine learning based model is trained with only information for nominal instances of one or more of the specimens. The one or more computer subsystems are configured for acquiring information for non-nominal instances of one or more of the specimens on which a proces is performed. The one or more com uter subsystems are also configured for re-training the machine learning based model with the information lor the no -nominal instances of the one or more of the specimens thereb performing transfer learning of the ^' information for the non-nominal instances of the one or more of the specimens to the machine learning based model The ^• system may be further configured as described herein.

Another embodiment relates to a compnter-implemented method for training a machine learning based model. The method includes acquiring information for non- nominal instances of one or more specimens on which a process is performed. A

machine learning based model is configured for performing one or more simulations for the specimens. The machine learning based model is trained with only information for nominal instances of one or more additional specimens. The method also includes retraining the machine learning based model with the- ormation. for the non-nominal instances of the one or more specimens thereby per orming transfer learning of the information for the non-nominal instances of file one or more specimens to the machine learning based model. The ac uiring and egaining are pertbtnied by one or more computer systems. One or more components are executed by the one or more computer systems. The one or more components include the machine learning based model.

Each of the steps of the method described above may fee further performed as described further herein, hi addition, the -embodiment of the method described above may include an other step(s) of any other method(s) described herein. Furthermore, the method descri bed abo ve may be performed by any of the systems described herein.

Another embodiment r Sates to a non-transitory computer-readable medium storing program instructions executabie on one or more computer systems for performing a computer-implemented method for training machine learning based model The computer-implemented method includes the steps of the method described above. The computer-readable medium may be farther configured as described herein. The steps of the comptiter-impiemented method may be performed a described .further herein, in addition, the computer-implemented method for which the program instructions are executable may include any other step(s) of any other roethod(s) described herein .

BRIEF DESCRIPTION OF THE DRAWINGS

Further advantages of the present invention will become apparen to those skilled in the art with the benefit of the following detailed description of the preferred embodiments and upon reference to the accompanying drawings in which;

Figs. 1 and l a are schematic diagrams, illustrating side views of embodiment of a Fig. 2 is a flow chart illustrating steps that may be performed by the embodiments described herein; and

Fig. 3 is a block diagram illustrating one embodiment of a non-transitory computer-readable -medium storing program instructions for causing -computer system(s) to perioral a computer-implemented method- described herein.

While the invention is susceptible to various modifications and alternative forms, speci fic embodi ments th ereof are shown by way of example i n the drawi ngs and are herein described in detai l. The drawings may not b to scale. It should be understood, however, that the drawings and. detailed description thereto are not intended to. limit the invention, t the particular form disclosed, but. ft the contrary, the intention is to cover all. modifications, equi valents and alternatives falling within the spirit and scope of the present invention as defined by the appended claims.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

The terms "design," "design data,' ^' and "design information" as used

interchangeably herein generally refer to the physical design (layout) of a IC and data, deri ved from the physical design through complex simulation or simple geometric and Boolean operati ons, In addition, a image of a reticle acquired b a reticle inspection system and/or derivatives thereof can be. used as a "proxy'' o proxies" for the design. Such a retieie image or a derivative thereof can serve as a substitute for the design layout in an embodiments described herein, that use a design. The design may include any other design data or design data proxies described in commonly owned U.S. Patent Nos.

7,570,796 issued o August 4, 2009 to Zarar et al. and 7,676,077 issued on. March 9, 2010 to JKulkarnl et al, both of which are incorpor ted ^' .by reference as if ful ly set forth herein. addition, the design data can be standard cell library data. Integrated layou data, design data for one o -more layers, deri vatives of the design data, and full or ^' partial chip design data. in addition, the . "design," "design data," and "design information' ^* described herei refers to information and data thai is generated by semiconductor device designers in a design process and is iherefore available for use in the embodiments described herein well in advance of printing of the design on any physical specimens such as reticles and wafers.

Turning now to the drawings, it is noted that the figures are not drawn to scale. In. particular, the scale of some of the elements, of the figures is greatly exaggerated to emphasize characteristics of the elements. It is also noted that the figures are not drawn, to the same scale. Elements shown in more than one figure that may be similarly configured have been indicated using ^' the same reference numerals, Unless otherwise noted herein, any of the elements described and shown may include any suitable commercially available elements.

One embodiment relates to a system configured to train a machine learning based model. The embodimen ts described herein provi de systems and methods for accelerating learning based systems when training samples are relatively limited. The embodiments described herein, may be utilized for a variety .of semiconductor manuiacturmg related applications including, but not limi ted to, inspection, metrology, defect review, and simulation.

One embodiment of such a. system is shown in Fig. 1. The system includes one or more computer subsystems {e.g., computer subsystems 36 and 102) and one or more components. 100 executed by the one or more computer subsystems, l te one or more components include machine learning based model 104 configured for performing one or more simulations for specimens. As described further herein, the machine learning based model is trained with only information for nominal instances of one or more of the specimens. in some embodiments, th system, includes optical based imaging subsystem 10, In general, the optical based imaging subsystem is ^' configured to generate optical images of the specimens, and the one or more computer subsystem s are configured for receiving the optical images from the optical based imaging subsystem. The optical based imaging subsystem may also be configured to generate the optical images with multiple modes, in one embodiment, the specimens include wafers. The wafers may include any wafers known in die art. In another embodiment, the specimens include reticles- The reticles may Include any reticles- known in the art.

The optical based imaging subsystem, ma generate the optical images by directing light to or scanning light, over the specimens and detecting light from the specimens. In one such example, in the embodiment of the- .system- shown i Fig. L optical based imaging subsystem 10 includes an illumination subsystem configured to direct light to specimen 1.4. The illumination subsystem includes at least one light source. For example, as shown in Fig, 1 , the illumination subsystem includes ligh -source 16, In one embodiment, the illumination subsystem is configured to direct the light to the specimen at one or more angles of incidence, which, may include one or more oblique angles and/or one or more normal angles. For example, as shown in. Fig. 1, light from light source 16 is directed through optical element 18 and theft lens 20 to specimen 14 at an oblique an gle of incidence. The oblique angle of incidence may include ^' any Suitable oblique angle of incidence, which may vary depending on, for instance, characteristics of the specimen.

The optical based imaging subsystem may be configured to direct the Sight to the specimen at different angles of incidence at different times. For example, the optical based imaging subsystem may be configured to alter one or more characteristics of one or more elements of the illumination subsystem such that the light can be direc ted to the specimen at an angle of incidence that is different than that shown in Fig, 1, hi one such example, the optical based imaging subsystem may be configured to move light source 16, optical element 1.8, and lefts 20 .such that the light is directed to the specimen at different oblique angle .©.f incidence or a norma! (or .near ^' norma I.) angle of incidence. in some instances, the optical based imaging subsystem may be configured to 5 direct light to the. specimen at more than one angle of incidence at the same time. For example, the illumination■subsystem may include more than one illumination channel one .of the illumination channels may include light source 16, optical element I S, and leas 20 as shown in Fig. 1 and another of the i umtnation channels (not shown) may include similar elements, which may be configured differently or the same, or may include at least i d a light source and possibly one. or more other components such as those described .further herein, if such light is directed to the speci men at the same time as the other light one or more characteristics (e.g., wavelength, polarisation, etc,) of the light directed t the specimen at different angles of incidence may be different, such thai light resulting from illumination of the specimen at the different angles of incidence can be discriminated

15 from each other at the detector(s).

In another instance, the illumination subsystem may include only one light source (e.g., -source -16 shown i Fig. 1) and light from, the light source may be separated into different optical paths (e.g., based o wavelength, polarization, etc,) by one or more 0 optical elements (not shown) of the illumination, subsystem. Light in each of the different optical paths may then be directed to the specimen.. Multiple illumination channels may be configured t direct light to. the specimen at the same time or at different times (e.g., when different illumination channels, are used to sequentially illuminate the specimen), in another instance, the same illumination channel may be configured to direct light to the 5 specimen with different characteristics at different times. For example, in some

instances, optical element 18 may be configured as a spectral filter and the properties of the spectral filter can be changed in a variet of difteren ways (e.g., by swapping out the spectral filter) such that, different wavelengths of light can be directed to the specimen at difterent times. The .illumination subsystem may have any other suitable configuration k wn in the art for directing the- light having different or th same characteristics to the specimen at different or the same angles of incidence, sequentially or simultane usl . in one embodiment, light source 16 may include a broadband plasma (BBP) light source, in this manner, the light generated by the light sourc and directed to the specimen may include broadband light. However, the light source may include any other suitable light source: such as a laser. The laser may include any suitable laser known in the art and may be configured to generate light at any suitable wavelength or wavelengths known, in the art. In addition, the laser may be .configured to generate light that is monochromatic or nearly-monochromatic. In this manner, the laser may be a narrowband laser. The light source may also include a polychromatic light source that generates light at multiple discrete wavelengths ^' . or wavebands.

Light from optical element 18 maybe focused onto specimen 14 by lens 20. Although lens 20 is shown in Fig. 1 as a single refractive optical element, it is to be understood that, in practice, lens 20 may include a number of refractive and/or. reflective optical elements that m combination focus the light from the optical element to the specimen. The illumination subsystem shown in Fig. .1 and described herein may include an other suitable optical elements (not shown). Examples of such optical elements include, hut are not limited io, polarizing components), spectral filter ^' s), spatial ii.lter(s), reflective optical elements), apodizerfs), beam splitterCs), aperture(s), and. the like, which may include an such suitable optical elements known in the art, in addition, the optical based imaging subsystem may be configured to alter one or more of the elements of the illumination subsystem, based on the type of illumination to- be used for imaging.

The optieal based imaging subsystem may also include a scanning subsystem configure to cause the light to be scanned over the specimen. For example, the optical based imaging subsystem may include stage 22 on which specimen 14 is disposed during imaging. The scanning subsystem may include any suitable mechanical and/or robotic assembly (that includes stage 22) that can be configured, to move the. specimen such that the light an ^' be- scanned over- the specimen. In addit n, or alternatively, the optical based imaging subsystem may be configured such that one or more optical ^' elements of the optical based imaging subsystem perform- some scanning of the light over the specimen. The light may b scanned over the specimen m any suitable fashion such as in a serpentine-like path or in a spiral path.

The optica! based imaging- subsystem further includes one or more detection channels. At least one of the one or more detec tion -channels includes ¾ det ec tor configured to detect light from the s cime due to illumination of the specimen by the system a«d to generate output responsive to the detected light. For example, the optical based imaging subsystem shown in Fig. 1 includes two detection channels, one formed by collector 24, element 26, and detector 28 and -another ^' formed by collector 3CL element. 32, and detector 34, As- shown in Fig, 1, the two detection channels are configured to collect and detect light at different angles of collection. In some instances, both detection channels are configured to detect scattered light, and the detection channels are configured to detect light that is scattered at different angles from, the specimen.

Howe ver, one or more of the detection, channels- may be configured to detect another type of light from the specimen (e.g., .reflected light).

As further shown in Fig. 1 , both detection channels are shown positioned in the plane of the paper and the illumination subsystem is also shown positioned in the plane of the paper. Therefore, in this embodiment, both detection channels are positioned in (e.g., centered in) the plane of incidence. However, one or more of the detection channels may be positioned out of the plane of incidence. For example, the detection channel formed by collector .30, element 32, and detector 34 may be configured to collect, and detect light thai is scattered out of the plane of incidence, Theref re, such detection channel may be commonly referred to as a "side" channel, and such a side channel may be centered in a plane that is substantially perpendicular to the plane of incidence. Although Fig. 1. shows -an -emhodimeatofthe optical based imaging subsystem that includes two detection channels*- the optical based imaging subsystem may include a different, number of detection channels (e.g., -only one detection channel or two or more detection channels), h one such instance, the detection channel formed by collector 30, element 32, and detector 34 may form one side channel s described above, and th optical based imaging subsystem may include, an additional detection channel. (not shown) formed as another side channel that is positioned -on the opposite side of the plane of incidence. Therefore, the optical based imaging subsystem may include the detection channel that includes collector 24, element 26, and detector 28 and that is centered in the plane of inci dence an d configured to coll ect and detect light at scatteri ng angle(s) that are at or close to normal to the specime surface. This detection channel may therefore be commonly referred to as a "top" channel, and the optica! based imaging subsystem, ma also include two or more side channels configured as described above. As such, the optical based imaging subsystem may include at least three channels (i.e., one top channel and two side channels), and each of the at least three channels has its own collector, each of which is configured to collect light at different scattering angles than each of the other collectors.

As described further above, each of the detection channels included in the optical based imaging .subsystem, may be configured to detect scattered light. Therefore, the optical based imaging subsystem shown in Fig, 1 may be- Configured for dark field (DP) imaging of specimens. However, the optical based imaging subsystem may also or alternatively include detection charmel(s) that are configured for bright field (BP) imaging of specimens. In other words, the optical based imaging subsystem may include at least one detection channel that is configured to detect light specularly reflected from the specimen. Therefore, the optica! based imaging subsystems described herein ma be configure for only OF, only B P, or both DF and BP imaging. Al though each of the collectors are shown in Fig. 1 as single retractive- optical elements, it is to be understood that each of the collectors may include one or more refractive optical element(s) and/or one or more reflective optical elemeut(s). The. one or more detectio channels ma include any suitable detectors known in the art. For example, the detectors may include phoio-imu!iip!ier tubes (P Ts), charge coupled devices (CCDs), time delay migration (TDI) cameras, and any other suitable detectors known m the art. The detectors may also include non-imaging detectors or imaging detectors , In this manner, if the detectors are ^' non-imaging detectors, each of the detectors may be configured to detect, certain characteristics of the scattered light. such as intensity but may not be configured to detect such ciiaracterisiics as function of position withi n, the Imaging plane. As such, the output that is generated b each of the detec tors included in each of the detection channel s of the optical based im aging subsystem may be signals or data, but not image signals or image data. In such instances, a -computer ^' subsystem such as computer subsystem 36 may be configured to generate im ages of the specimen from the non-imaging output of the detectors. However, in other instances, the detectors may be configured as imaging detectors that are configured to generate imaging signals or image data. Therefore, the optical based imaging subsystem may be configured to generate the optica! images described herein- i n a number of ways.

I is noted that Fig- 1 is provided herein to generally illustrate a configuration of an optical based imaging subsystem that may be Included in the system embodiments described herein or that ma generate images that are used by the system embodiments described herein. Obviously, the optical based imaging subsystem configuration

described herein may be altered to optimize the performance, of the optical based imaging subsystem as is normally performed when designing a commercial imaging- system, in addition, the systems described herein may be implemented using an existing system (e.g.. b adding functionality described herein to an existing system) such as the

29xx/2Sxx series of tools that are commercially available iron.) KLA-Tencor,. Milpltas, Calif. For some such systems, the embodiments described herein may be provided as optional functionality of the system (e.g., in addition to other inrtctionality of the system), Alternatively, the optical based imaging subsystem described herein may be designed "from scratch" to provide a completely new optical based imaging subsystem. Computer subsystem 36 coupled to the optical based imaging: subsystem, may be coupled to the detectors of ^* the optical based imaging subsystem in any suitable maimer (e.g., via one or more transmission media, which may include "wired" and/or "wireless" transmission media) such that the computer subsystem can receive the output generated by the detectors for the specimen. Computer subsystem 36 may be configured t perform a number of functions described further herein using the output of the defectors.

The computer subsystems shown in Fig. I (as well as other computer subsystems described herein) may also be referred to herei as computer system(s). Each of the computer subsystem(s) or system(s) described herein may take various forms, including a personal computer s stem, image computer, mainframe computer system, workstation, network appliance, Internet appliance, or other device. In general, the terra "computer system' ^' may be broadly defined to encompass any device having one or more processors, which executes instructions from a memory medium. The computer subsysiera(s) or system(s) may also include any suitable processor known in the art such a a parallel processor. In addition, the computer subsystera(s) or system(s) may include a computer platform, with high speed processing and software, either as a _: standalone or a networked tool. if the system, includes ^' more than one computer subsystem, then the different computer subsystems may be coupled to each other such that images, data,, information, instructions,, etc. can be sent between the .computer subsystems as described further herein. For example, computer -subsystem 36 may be coupled to computer subsystem^) 102 as shown by the dashed line in Fig. 1 by any suitable transmission media, which may include any -suitable wired and/or wireless transmission media, known in the art. Two or .more of ^* such computer subsystems may also be effectively coupled by a shared computer- readable storage medium (not shown). Although the imaging -subsystem is described above as being an optical or Sight- based imaging system, the system may also ^' or alternatively include an electron beam, based imaging subsystem configured to generat electron beam images of the specimens, and the one or more computer subsystems are configured for receiving the electron beam images from the electron beam based imaging subsystem, in one such embodiment th electron beam based imaging subsystem may be configured to direct electrons to or scan electrons over the. -specimen and to detect electrons from the specimen. In one such embodiment shown in Fig. la, the electron beam based imaging subsystem includes electron column 122 coupled to computer subsystera .124,

As also shown in Fig. l a, the electron column includes election beam source 126 configured to generate electrons that are focused to specimen 128 by one or more elements 130, The electron beam source may include* for example, a cathode source or emitter tip, and one or more elements 130 may include, for example, a gun lens, an anode, a beam limiting aperture, a gate valve, .a beam current selection aperture, an objective lens, and a scanning subsystem, ail of which may inc lude any such suitable elements known in the art.

Electrons returned from the specimen (e.g., secondary electrons) may be focused by one or more elements 1.32 t detector 134. Que or more elements 1:32 ma include, for example, a scanning Subsystem, which may be the same scanning subsystem included in elenient(s) 130,

The electron colum may include any other suitable elements known in the art. In addition, the election column may be further configured as described in U.S. Patent Nos. .8,664,594 issued April 4, 2014 to Jiang _, et al, 8,692,204 issued April 8, 2014 to Kojima et a.l„ 8,698,093 issued April. 15, 2014 to Gubbens et -al, and 8,736,662. issued May 6, 2014 to aeDonak! et al., which are incorporated by reference as if fully set forth herein. Although the electron column is s wn in Pig. la as being configured such thai the electrons are directed to the Specimen at an oblique angle of incidence and are scattered ^' from the specime at another oblique angle, it is to be understood that the electron beam ma be directed to and scattered from the specimen at any suitable angles. In addition, the electron beam based imaging subsystem may be configured to use multiple modes to generate image of the specimen as described further herein (e,g., with different illumination angles, collection angles, etc.). The multiple modes of the electron beam based imaging subsystem may be different in any image generation parameters of the electron beam based imaging subsystem.

Computer subsystem 124 may be coupled, to detector 134 as described above. The detector may detect electrons returned from the surface of the specimen thereby forming electron beam images of the specimen. The electron beam images may include any suitable electron beam images. Computer subsystem 124 may be configured to- erform one or more functions described further herein for the .specimen using output geiiemted by detector 1.34, Computer subsystem 124 ma be configured to perform any additional step(s) described herein. A system that includes the electron beam based imaging subsystem shown in Fig. l a ma be further configured as described herein.

It is noted that Fig. fa is provided, herein t generally illustrate a configuration of an electron beam based imaging subsystem that ma be included in the embodiments described herein. As with the optical based imaging subsystem described above, the electron beam based imaging subsystem configuration described herein may be altered to optimize the performance of the imaging subsystem as is normally performed when designing a commercial imaging subsystem. In addition, the systems described herein .may be implemented using an existing system (e.g., by adding functionalit described herein to an existing system) such as the eSxxx. and eD -xxxx serie of tools thai are commercially available from KLA-Teocor, For some such systems, the embodiments described herein may be provided as optional iimctionaSity of the system (e.g., in addition to other iimetiona!hy of the system) . A lternati vely, the system described herein may be designed "from scratch" to provid a completely new system.

Although the imaging subsystem is described above as being a light-based or electron beam-based imaging subsystem,, the imaging subsystem may be an ion beam- based imaging subsystem. Such an imaging subsystem may be configured as shown in Fig, la except that the electron beam source may be replaced with an suitable ion beam source known in the art, hi addition, the imaging subsystem may be an other suitable ion beam-based imaging subsystem such as those included in commercially available .focused ion beam (FIB) systems, helium ion microscopy (HIM) systems, and. secondary ion mass spectroscop (SIMS) systems.

As noted above, the optical and electro beam based imaging subsystems may be configured for directing energy (e.g.. Sight, electrons) to and/or scanning energ over a physical version, of the specimen thereby generating actual images for the physical version of the specimen. In this maimer, the optical and electron beam based imaging subsystems may be configured as. "actual " imaging systems, rather than "virtual" systems. For example, a storage medium (not shown) and computer subsystern(s) 102 shown in Fig. 1 may be configured as a 'Virtual' ^* system. In particular, the storage medium and the computer subsystem(s) are not part of imaging subsystem 10 and do not have any capability for handling the physical version of the specimen, in other words, in systems configured as virtual systems, the output of its one or more "'detectors" may be output that was previously genera ted by one or more detectors of an actual sy stem and that is stored in the virtual system, and during the 'imaging and/or scanning," the virtual system may replay the stored output as though the specimen is being Imaged and/or scanned. In this manner. Imaging and/or scarmi rig the specimen with a vi rtual system may appear to be the same as though physical specimen, is being imaged and/or scanned with an actual system, while, in reality, the "imaging and/or scanning" involves simply replaying output for the specimen in the same manner as the specimen may be imaged and/or scanned. Systems and methods configured as 'Virtual ^' ' inspection systems are described in commonly assigned U.S. Patent Mos. 8,126,255 issued ©» February 28, 2012 to Bhaskar et al. and 9,222,895 issued on December 29, 15 to Daffy et at., both of which are incorporated by reference as if fully set forth herein. The embodiments described herein ma be further configured as described in these patents. For example, th one or more com ute ' subsystems described herein may be further configured as described in these patents.

As further noted above, the imaging subsystem may be configured to generate images of the specimen with .multiple modes. In general, a "mode" can be defined by the values of parameters of the Imaging subsystem used for generating images of a specimen or the output used to generate images ^' of the specimen. Therefore., modes that are different may be different in the value for at least one of the imaging parameters of the imaging subsystem. For example, in one embodiment of an optical based imaging subsystem, at least one of the multiple modes uses a t least one wavelength of the light for illumination that is different from at least one wavelength of the light for illumination used for at least one other of the multiple modes, The modes may be different in the illumination wavelength as described further herein (e.g., by using different light sources, different spectral filters, etc.) for different modes. In another embodiment, at least one of the multiple modes uses an illumination channel of the Imagi ng subsystem that is different from an il l umination channel of the imaging subsys tem used for at least one other of the multiple modes. For example, as noted above, the imaging subsystem may include more than one illumination channel. As such, different illumination channels may be used for different modes.

In one embodiment, the system includes an inspection subsystem configured to generate output for the specimens, and the one or more computer subsystems are configured for receiving the output from the inspection subsyste and detecting defects on the specimens based on the output. Fo example, the optical and electron bea imaging subsystems described herein may be configured as inspection subsystems. The computer suhsystem(s) may be configured to receive the output from the inspection subsystem as described .above (e.g., from deteetorfs) of the imaging subsystem) and may be configured to defect defects on the specimens based on the output in any suitable maimer. For example, the computer subsystem(s) may be configured to compare the ouipui to one or more thresholds, any output thai is above the one or more thresholds may be identified by the computer subsystemCs) as. defects or potential defects, and any output, that is not above the one or more thresholds may be identified by the computer suhsystero(s) as not. defects or not potential defects. However, _, the computer subsystemCs) may be configured to use any suitable algorithmfs) and/or method(s) to detect defects on the specimens based on the output.

In another embodiment, the system includes a. defect review .subsystem configured to generate output for defects detected on the specimen, and the computer subsysteniis) are configured for receiving the output from the defect review subsystem and determining properties of the defects detected o the specimens based on the output. For example, the optical and electron beam imaging subsystems described herein may be configured as defect review subsystems. The computer subsystemf ^' s) may be configured to receive the output from the defect review subsystem as described above (e.g., from detector(s) of the imaging subsystem} and may be configured to determine properties of the defects on the specimens based on the output in any suitable manner. For example, the computer subsystem(s) may be configured to use the output to determine one or more propertie such as size and shape of the defects using any suitable algoriihm(s) and/or method(s) known in the ait.

In a further embodiment, the system includes a metrology subsystem configured to generate output fo the specimens, and the one or more compute subsystems are

configured for receiving the output from the metrology subsystem and determining properties of the specimens based on the output. For example, the optical and electron beam imaging subsystems described herein may be configured as metrology subsystems. The computer sobsysteni(s) may be configured to receive the output from the metrology •subsystem as described above (e.g.. from delector{s) of the imaging subsystem) and may be configured to determine properties of the specimens based on the ^' output in any suitable manner. For example, the computer subsystem(s) may be configured to us the output to determine one or more properties such as si e and shape of the patterned features formed on the specimens using any suitable algorithm(s) and/or meihod(s) known in the art.

The embodimen ts of the imaging subsystems described herein and shown in f igs. I and 1 a ma be modified in one or more parameters to provide di ferent imaging capability depending on the application for which they will be used. In one suc example, the imaging subsystem shown in Pig. 1 may be configured to have a higher resolution if it is to be used for defect review or metrology rather than fo inspection. In other words, the embodiments of the imaging subsystem shown in Figs. 1 and l a describe some general and various configurations for an imaging subsystem that can be tailored in a number of manners that will be obvious to one skilled in the ar to produce imaging subsystems having different imaging capabilities that are more or less suitable for different applications.

The inspection subsystem, defect review subsystem, and metrology subsystem, may also be configured for inspection, defect review, and metrology of specimens such as wafers and reticles. For example, the embodiments described herein may be configured For training a machine learning based model thai performs one or more simulations for the purposes of mask inspection, wafer inspection, and wafer metrology. In particular, the embodiments described herein may be installed on a computer node or computer cluster that is a component, of or coupled to an imaging subsystem such as a broadband plasma inspector, an electron beam inspector or defect review tool, a mask Inspector, a virtual inspector, etc. in this manner, the embodiments described herein may perform simulations that can. be used for variety of applications that include, but re not limited to, wafer inspection, mask inspection, electron beam inspection and review, metrology, etc. The characteristics of the imaging subsystems shown in Figs. 1 and la can be modified as described above based on the specimen for which it will generate actual linages. in another embodiment, the system includes a semiconductor fabrication

5 subsystem configured to perform one or more. fabrication processes on the specimen. For example, as shown in Fig, 1, the system may include semiconductor fabrication subsystem 106, which may be coupled to computer subsystem^) 102 and/or -any other elements of the system, described herein. The semiconductor fabrication subsystem may include any semiconductor fabrication tool and/or chamber known in the art such as a i d lithography track, an etch chamber, a chemical mechanical polishing (CMP) tool, a

deposition chamber, a stripping- or cleaning chamber, and the like. Ex amples of suitable semiconductor fabrication tools that may- be included in the embodiments described, herein are described in U,S, Patent No, 6,891,627 to Levy et ah issued on Ma 10, 2005, which is incorporated by reference as if fully set forth herein. The embodiments

15 described herein may be further configured as described in this patent.

As described above, therefore, the one or more computer subsystems described herein, may he included in a system with one or more other subsystems with actual wafer handling and/or processing capability (e.g., imaging subsystems, inspection, subsystems, 0 defect review subsystems, metrology subsystems, semiconductor fabrication proces

subsystems) _* In this manner, the embodiments described herein may be configured as predictive systems including data in situ inside any semiconductor platform such as a metrology tool, an inspection tool, an etch chamber, etc. thai has detectors) and a computational platform to learn a model of its world (e.g., defects on a wafer in the case 5 of a semiconductor inspector ). However, the one or more computer subsystems described herein may be included in a system that does not include one or more subsystems that have actual wafer handling and o processing capability, in this manner, th

embodiments described, herein may be- configured ^' as a predictive syste that incliides data ex situ where the data is exported to a persistent storage mechanism. For example, 0 the system may be configured -as a virtual system as described farther herein that may or may not acquire m&rrnadon from actual physical wafers generated by other systems and/or methods, in. particular, another system, or method may generate information for nominal, and/or non-nominal ins tances of one or more specimens using actual physical specimens and may store that information in a storage medium. The one or more computer subsystems described herein may then ^■■ acquire such information from the storage medium.

The component^), e.g., components) 100 shown in Fig. 1, executed by the computer sub$ysiem(s), e.g., computer ^' Subsystems 36 and/or J.02, include machine learning based, model 104. The machine learning based, model is configured for performing one or more simulations for specimens, and the machine learning based mode! is trained with-oniy information for nominal instances of one or more of the specimens. The training with only the nominal instances of the one or more specimens may be considered as a kind of initial trainin that is followed b the re-training described further herein. For example, prior to being capable of giving a prediction, such a machine learning based model typically will need to be trained after it has been given a set of examples to learn from. Training the machine learning based model with, only the nominal instances may be performed by the one or more computer subsystems described herein. However, training the machine learning based model with only the nominal instances may be performed by another system or method, and the trained machine learning based model may be acqui ed from the other system o method (o a storage medium in which the trained machine learning based model has been stored by the other system or method) and then re-trained by the computer subs stem(s) described herein.

Machine learning can be generally defined as a type of artificial intelligence (ΑΪ thai provides computers with the ability to leant without being explicitly programmed. Machine learning focuses on the development of computer programs that can teach themselves to grow and change when exposed to new data. In other words, machine learning can be defined as the suhfiek! of computer science that "gives computers the ability to learn without being explicitly programmed; ^' Machine, learning explores the study and construction of algorithms that can. team, from and make predictions on data - such algorithm overcome following strictly static program .inst uctions by making data drive predictions or decisions, through building a model from sample inputs.

The machine learning based model described herein may be further configured as described in "Introduction to Statistical Machine Learning," by Sttgiyama, Morgan Kaufmann, 2016, 534 pages; "Discriminative, Generative, and Imitative Learning," Jebara. MIT Thesis, 2002, 232 pages; and ^' "Principles of Data Mining (Adaptive

Computation and Machine ^' Learning)" Hand et al, MIT Press, 200 , 578 pages; which are incorporated by reference as if fully set forth herein. The embodiments described herein maybe further configured as described in these- references.

The one or more nominal instances of the specimens and the- training performed using the one or more nominal instances may vary depending on the simulations that will be perforated by the machine learning based model and the machine learning based model itself. For example, if the machine learning based model will be used to perform simulations of how the specimens will appear in images of the specimens generated by an imaging system., then the nominal instances of the specimen(s may include design- data for the specimens) as well as actual images generated, of the speeimen(s) on which the design data is printed. The training may then involve providing the nominal instances to the machine learning based model such that the machine learning based model can find (i.e., learn or identify) one or more parameters of the .machine learning based model (e.g., weights) thai cause the machine learning based model to produce simulated images -from the corresponding design information that are substantially the same as the .corresponding- actual images. Othe ^' types of information may be included in the information tor the nominal instances used for training the machine learning based model depending on othe simulations described herein (e.g., metrology system output and/or measurements with corresponding specimen desig information), fa addition, the characteristics of the

model, itself in that different types of machine learning based models (such as those described further herein ma require- -different amounts of information, for training (e.g>, different, numbers of nominal instances). The characteristics of the hifonnation for the nominal, instances needed for adequate training of any particular machine learning based model can be determined in an suitable manner known in the art,

"Nominal instances" as that term is used herein is generally defined as "instances" of specimens on which no defects are known to be- present. For example, the specirnen(s) mat are nominal may be processed with one o more parameter of one or more processes used to form a design on the specimen(s) thai are known to generate. non-defective specimens, in other words, the processes) used to generate the specimens) ma he known good, proeessfes). la addition, the "nominal instances" of the specinien(s) may include- instances (e.g., locations o specimens, whole specimens, etc-.) -at which no defects have been determined to be present. For example, the nominal instances of the specimenCs) may include locations and/or specimens at which an inspection or other qualit control related process (e.g., defect, review, metrology, etc.) have not detected any defects, in contrast, the "non-nominal instances" of the specimen{s) as that term is used herein is generally defined as "instances" of the specimen(s) that are- "defective" in some manner, either intentionally (via programmed or synthetic defects as described further herein) or unintentionally (via defects discovered on one or more -specimens by inspection., defect review, metrology, ^' etc.). hi addition, it is noted that a specimen may be "defective'* ^' even if no "defects" per se are detected on the specimen. For example, a specimen may be considered "defective" if one or more measurements indicate that one or more characteristics of a design formed on the specimen are outside of a desired range of values ^' for those one or more characteri stics.

In one embodiment, performing the one or more simulations include generating one or more simulated, images for one of the specimens, and the one or more simulated, images illustrate how the one of the specimens appears i one or more actual images of the one or more specimens generated by an imaging system. For example, the one o more simul tions performed by the machine learnin based model may generate simulated itnage(s) illustrating how a specimen appears in actual image(s) generated by one of the imaging systems- described herein. In. this manner, the simulated image(s) may represent images that may be generated of the specimens by an inspection s stem, a defect review system, or other imaging system described herein. In one such example, the inpu to the one or more simulations performed by the machine learning based model may include design information for specimen (e.g., any of the design data -described farther herein) and the output of the one or more simulations may include one or more simulated optical or electron beam images illustrating how the specimen on which the design information has been formed will appear in those images.

In one such embodiment, the imaging system is an optical based imaging system. in another such embodiment, the imaging system is -a¾ electron beam based imaging, system. These imaging systems may be configured as described further herein. In some such embodiments, the one or more simulations performed by the machine learning based model may be performed t generate one or more reference images that is/are used for inspection of spe ^' cimen(s). In this manner, the embodiments described herein can enable advanced inspection ^' algorithms ^' such as die-to-database defect detection -methods and/or algorithms in a manner that are extraordinarily efficient to apply in research and development and -manufacturing scenarios. Enabling such defect detection is particularly valuable for electron beam, based inspection where throughput is constrained, by the rate of area coverage. By limiting image acquisition to "test" images only , throughput can be doubled or tripled versus currently used die-to-die inspection methods. The embodiments described herein may be further configured as described in tLS. Patent Application Serial No, 15/353,210 by Bhaskar etaL filed November 16, 2016, which is incorporated by reference as if folly set forth herein. For example, the embodiments described herein may be configured to perform single image detection as described in this patent application.

In another embodiment, performing the one or more simulations includes generating one or more simulated measurements for one of the specimens, and the one or more simulated measurements represent output generated for the one of the specimens b a metrology system. For example _* , the one or mer simulations performed by the machine learning based model may generate simulated measuremeni(s) (e.g., image(s), output, data, etc.) representing output genera ted by one of the systems described herein for speeinien(s). m this manner, the simulated measurements) may represent measurements, images, output, da a, etc, that may be generated for the specimens by a metrology system described herein. In one such example, the input to the one or more simulations performed by the machine learning based model may include design information fo a specimen (e.g., any of the design data described further herein) and the output of the one o more simulations may include on or more simulated optical, or electron beam measurements .representing output generated for the specimen by those systems.

The computer suhsystern(s) are configured for acquiring Information for non- nominal instances of one or more of the specimens on which a process is performed. As will be described ^' further herein, the o tation for the non-nominal instances will be used for retraining of the machine learning based mode! thereby performing transfer learning of the non-nominal instances to the machine learning based model . Therefore, acquiring the Information for the «.o.n-nominal instances may essentially he transfer learning training input generation. Transfer learning training input generation can b performed in a number of ways described further herein including: a) empirical simulation of .real defect events on wafers and masks using process design of experiments (DOEs); b) introduction ^' of ^' virtual defect events in design/simulation space by using synthetic approaches; and c) hybrid approaches using empirical pins synthetic method in conceit In this manner , the embodiments described herein may use process DOEs, simulation and programmed defects to generate ^' training sets for neural networks and other machine learning based models described herein.

Acquiring the information for t he non-nominal instances of the specimen! s) may include generatin the information. For example, the systems described herein may be configured to generate the information for the non-nominal instances by perfbmting one or more processes on actual specimens (e.g., performing one or more fabrication processes on an actual, physical specimen arid then, performing one or more measurement and/or imaging processes on the actual physical specimen). I ierefore, generating the information for the non-nominal instances may be performed, using the computer subsystem(s) described herein with one or more other subsystems described herein (e.g., one or more semiconductor imbrication subsystems, one or more imaging subsys tems, one or more metrology subsystems, etc.). h another example, the systems described herein may be configured to generate the information for the notwiominal instances by performing one or more simulations (with, or without using actual, physical specimens). In one such example, as described further herein, the one or more computer subsystems may be configured to use another model, to generate the information for the non-nominal instances of the speei nien(s). Alternatively, the embodiments described herein may be configured to acquire the information for the non-nominal instances of the specimens from a storage medium in which the information has been stored by one or more other methods and/of systems that generated the information. hi one embodiment the non-nominal instances _, includ instances of defects on the one or more specimens. In this manner, the information for the non-nominal instances may include infbmiation for one or more detects on one or more specimens. Such, informatio may be generated in a variety of ways described further herein. in another embodiment, the fton-hominal instances include instances of defects on the one or more specimens, and the one or more specimens include one or more actual specimens on which the process is performed with two or more different values of one or more parameters of the process. For example, for transfer learning training input generation that includes empirical simulation of real defect events on wafers and reticles using DOEs, real world wafers can be leveraged for defects as described further herein.

In one such embodiment, the process is performed with the two or more different values of the one or more parameters of the process in a process window qualification (PWQ) method. For example, one strategy for generating the transfer learning training input is to use DOEs such as P WQ as a generator of systematic defects. The one or more parameters of the process that are varied in the PWQ method may include focus and exposure (e.g., as in a focus-exposure PWQ process). PWQ methods may also he performed as described in U.S. Patent Nos. 6,902,855 to Peterson et al. issued on June 7, 2005, 7,418 J 24 to Peterson et al. issued on August 26, 008, 7,729,529 to Wu et al. issued on June 1, 2010, 7,769,225 to Kekare et at. issued on August 3, 2010, 8,04 IJ 06 to Pak et al. issued on October 18, 201 1, , ϊ ϊ ^' ϊ ,900 to Wu et al issued on February 7, 2012, and 8,213,704 to Peterson et al issued on July 3, 201.2, which are incorporated b reference as if folly set forth herein. The embodiments described herein may include an step(s) of any metkxi(s) describe in these patents and may be farther configured as described in these patents, A PWQ wafer may be printed as described in these patents.

In another such embodiment, the process is performed with the two or more different values of the one or more parameters of the process in a P WQ method designed for overlay margin determination. For example, one strategy for generating the transfer learnin training input is to use DOEs such as overlay-PWQ wafers as a generator of systematic defects. Overlay margin determination may be performed in any suitable manner including as. described in the above-referenced patents. Therefore, an overlay- PWQ method may be used to print such dies on a specimen, and the noti-nomiiial instances ma include instances of any defects detected on such a Specimen. hi an additional such embodiment, the process is performed with the two or more different values of the one or more parameters of the process in a focus exposure matrix (FBM) method. For example, one strategy for generating the transfer learning training input is to use DOEs such as FBM methods and/or wafers a a generator of systematic defects. FBM methods generally involve printing a number of dies on a wafer at different combina tions of focus and exposure parameter val ues of a lithography process. The different dies can then be inspected in any suitable manner to detect defects in the different dies. That mformation is then typically used to determine a process window for the focus and exposure of the lithography process. Therefore, a FEM method may be used to print such dies on a specimen, and the- non-nominal instances may include instances of any defects detected on such "a specimen.

As described above, therefore, the transfer teaming training input generation may 5 use tme or more DOEs such as one or more of PWQ, Ff; M, and overfay-PWQ wafers as generators of systematic defects, in other words, the machine, learning based mode! may be trained by using information generated from a PWQ or F EM wafer that can act as a generator of systematic defects, in general, PWQ is a technique invented b KLA-Teneor in the early 2000s for lithography focus and exposure process window characterization i d and is widely adopted in one .form or another. The basis tor PWQ is to create an

inspector compatible wafer where there are nominal dice and modulated dice next to each other in a systematic fashion to maximize signal for the inspector. Similar wafers can be made for determining overlay margins. While PWQ and their "cousin" FEM wafers are primarily used for determining process margin today, they can he repurposed for training

15 deep neural networks (or any other .machine learning based model described herein) with real defects since they wilt occur in. abundance on -a given wafer. These wafers and the information generated from them can then he used as a set of training samples for the retraining described .further herein. However, such samples may not provide a complete set of possible defects since there i no -guarantee to see all defect types on such wafers, 0 Therefore, the information generated from such wafers may be complemented with ther information generated by synthetic defect generation, which ma be performed in a number of different manners as described further herein.

In some embodiments, the acquired, information is -generated from synthetic

5 design data for the one or more specimen s produced by an electron, design automation (EDA) too!., ^' in. this manner, the embodiments described herein may- have an added dimension of leveraging synthetic- data generated with EDA computer aided design.

(CAD) tools. The EDA tool may include any suitable commercially available EDA tool , In some such embodiments, one or more of the computer subsystems described herein (e.g., computer subsystem(s) 102) maybe -configured as an EDA tool or may be a computer subsystem included in. an EDA. tool. in a further embodiment the non-nominal instances- include instances of defects on the one or more specimens, and the defects include one or more synthetic defects generated by altering a design. for the one or more specimens to create the synthetic defects in the design, "Synthetic" defects as that term is used herein can be generally defined as one or more defects purposefully caused on a specimen, e.g., by manipulation of the design information for the specimen. Therefore, "synthetic" defects may also be referred to as "hypothetical' ^* defects or "programmed'* defects, in one such embodiment, for transfer learning training input generation via introduction of virtual defect events in design/simulation space by using synthetic approaches. CAD design can be used to synthetically generaie- defects (e.g., opens, shorts, protrusions, line ends, metrology markers, etc.) and then be treated by a deep generative or other model as described further herein (to create realistic defects on training images) and/or be used to print the synthetically generated defects on one or more specimens, which ca then be used to generaie images of the synthetically generated defects on the specimens). The CAD work can be automated, with a programmable graphical EDA editor, which may include any suitable EDA software, hardware, system, or method, hi one such embodiment, the one or more components include an inception module configured for altering the design to create the synthetic defects in the design. For example, the machine learning _, based models described herein may be trained by a defect hallucination system such as those suggested by GoogEeNet inception for natural scene images. A traditional neural network that is pre-trained o defects can then play these backwards to create new defect types on other geometry structures. Examples of systems and methods, for performing GoogLeNet inception, can be found in. "Going Deeper with Convolutions," Szegedy et aL, 2015 IEEE Conference on Computer Vision and Pattern Recognition (CV.P ), June 2015, 9 pages, which is incorporated by reference as if fully set forth herein. The embodiments- described herein may be ikrthef configured as described in this reference. in some embodiments, the nominal instances of the one or more specimens include natural scene images, For example, a model created using natural scene images can be used as a starting point to frae-tune the mode! that is trained on semiconductor data (optica! or SEM or any other non-nominal instance information described herein). This is an example of transfer learning where low level features learned from training on relatively large amounts of natural images aid in classifying semiconductor delect data. We have demonstrated a I Ox reduction in training time when leveraging transfer learning from natural to semiconductor data, in this manner, natural images can he used to accelerate training time on semiconductor data. This reduction in training time ^' is achieved by using the model weights from natural images as the initialization point fo the training session that uses semiconductor data. This sharing of weights between natural and semiconductor data can be achieved via any of the transfer learning described herein. As an added benefit, we achieve a 7% increase in classification accuracy when training a model with weights initiali ed from natural Images, The natural scene images used for the nominal instances may include any suitable natural scene images known in the art, including standard natural scene images. In this manner, the nominal instances of the specimens may include non-semiconductor specimens (i.e., not wafers or reticles), while the no -nominal instances of the specimens may include semiconductor specimens (i.e., wafers and reticles). Therefore, the training and the re-training may he -performed, using different types of specimens. Furthermore, the nominal instances of the specimens may include images of different type of natural scenes (e.g., animals, humans, objects, etc.) in another embodiment, the nomi nal instances of the one or more specimens comprise more than one type of data, The more than one type of data may be generated for the sam type of specimens (e.g., only wafers or only reticles). In addition, the more than one type of data ma he generated using different imaging types (e.g., optical and electron beam), in this tnanner, a variety of semiconductor data may be used to accelerate the training ti me for a target datasei. For example, In a varian ^' of the .above embodiment, data from various semiconductor data (optical and SEM) can be used to train a model which can then be used as a starting point to train a model on a targeted, semiconductor datasei. This approach has a similar- impact of learning low level features from natural images and fine-tuning them to -semiconductor data.- Training data. from, a variety of semiconductor data helps the model or network to learn low level representations of .semiconductor data that can then be transferred t a targeted datasei. litis manifestation of transfer learnin also accelerates the training time significantly.

In some embodiments, the . non-nominal instances include instances of defects on the one or more specimens, the defects include one or more synthetic defects generated by altering a design for the one or more specimens to create the synthetic defects in the design, and the information for the non-nominal instances includes output generated by an imaging or metrology system for the one or more .specimens on which the ^' synthetic defects are printed. For example, an example of a transfer learning training input generation method, which uses a hybrid approach in vol ving empirical -and synthetic methods in combination, ma include generation of synthetic layout defects as described above in design space and determining the empirical impact of those synthetic layout defects on wafers by making masks with the modified design and processing wafers with those masks, in this manner, the machine learning based model can. be trained by actual images of programmed defects.

In one embodiment, the non-nominal instances include instances of defects on the one or more -.specimens, the defects include one or more synthetic defects generated by altering a design for the one or more specimens to create the synthetic defects in the design, and the information for the non-nominal instances includes the altered design. For example, synthetic defects in design can be- consumed in isolation without feeding them to a generative mode. The detects created in design do not necessarily have to be legal struc tures as they act as examples of structures where a nominal pattern is broken for the network to learn. These synthetic defects reduce the dependency on the need for real pattern defects thereby reducing the data acquisition period which in turn accelerates model training time, in this manner, the embodiments described herein may use synthetic design defects directly (in the absence of a generative model) to reduce training time).

The design process for transfer learning training input generation me hods b) and c) described above may leverage process, lithography, and tool si mulation capabilities. For example, in another embodiment, the non-nominal instances include instances of defects on the one or more specimens, the defects include one or more synthetic defects generated by altering a design for the one or more specimens to create the synthetic defects in the design, the information for the. non-nominal instances includes output of another model, and the output of the other model represents the one or more specimens on which the synthetic defects are printed, One example of an empirically trained process model includes SEMulator 3D, whic is commercially available from Coventor, Inc., Cary, NC An example of a rigorous lithography simulation model is Prelith, whic is commercially available from KLA»Tencor, which can be used in conceit with the SEMulator 3D product. However, the other model that is used in this embodiment may include any suitable model of any of the proeess(es) involved ^' in. producing actual specimens from the design data. In this manner, the altered desig (altered to include one or more synthetic defects) may be used to simulate what a specimen on which the altered design has been formed will look like in specimen space (not necessarily what such a specimen would look like to a system such as an imaging or metrolog system).

Therefore, the output of the other model may represent what the specimens would look like in 2D or 3D space of the specimens.

In an additional embodiment, the non-nominal instances include instances of defects on the one or more specimens, the defects include one or more synthetic defects generated by altering a design for the one or more specimens to create the synthetic: defects in the design, the information for the non-nominal instances includes output of another model, and the output of the other model illustrates how the one or more specimens on which the synthetic defects are printed appear in one or more actual images of the ^■ specimen generated b an imaging system. For example, an additional capability that can. be applied is simulation of the inspector whose recipe is being developed. An example of such -a model is WINsini, which is commercially a vailable from KLA-Tenco.r _» and which can rigorously model the response of an inspector using an electromagnetic (EM) wave solver. In this manner, defect behavior for programmed defects can ^' be learned in one model and applied in. another. Such simulations maybe performed for an other imaging subsystems or systems described herein. In addition, such simulations may be performed using any other suitable software, algorithm(s), method(s), or sysiem(s) known in the art.

In a Further embodiment, the non-nominal instances include instances of defects on the one or more specimens,, the detects include one or more synthetic defects generated by altering a design for the one or more specimens t create the synthetic defects in the design, the information for the non-nominal instances includes output of another model, and the output of the other model represents output generated by a metrology system for the one or more specimens on which the synthetic defects are prin ted. For example, an additional capability that can be applied is simulation of the metrology tool whose recipe is being developed. In this manner, defect behavior for programmed defects can be learned in one mode! and applied in another. The output of the metrology system for the one or more specimens on which the synthetic defects are printed may be generated using any ^' .suitable model of the metrology system known in the art, hi some embodiments, the non-nominal instances include instances of defects on the one or more specimens, the defects include one or more synthetic defects generated by altering a design for the one or more specimens to create the synthetic defects in the design, the information, for the non-nominal instances- includes output of another model, the output of the other model represents output generated by another system for the one or more specimens on which the synthetic defects are printed, and the other model is a deep generative model. In one such embodiment, for transfer learning training input generation via introduction of virtual defect events in design simulation space by using synthetic approaches, CAD design can be used to s nthetically - generate -defect^. (e.g _{< *} opens., shorts, protrusions, line ends, metrology markers, etc) and then be treated by a deep generative mode! as described further herein (to create realistic defects on training images). In this manner, defect behavior for programmed defects can be learned in one .model and applied in another. The generative model may be used to generate simulated output that represents the output generated by any of the sys tems described herein (eg., a semiconductor fabrication system, an inspection system, a defect review system, a metrology system, etc.).

A "generative" model can be generally defined as a model that is probabilistic in nature,. In other words, a "generative" model is not one that performs forward simulation or rule-based approaches and,. as such, model of the physics of the processes involved in generating an actual image or output (for which a simulated image or output is being generated) is not necessary, instead, as described further herein, the generative model can be learned (i that its parameters can be learned) based oil a suitable training set of data. As described further herein, such generative models have a number of advantages for the embodiments described herein, m addition, the generative model may be configured to have a deep learning architecture in that the generative model may include, multiple layers, which perform a: number of algorithms or transformations. Th number of layers included in the generative model ma he use case dependent For practical purposes, a suitable range of layers is from.2 layers to a few tens of layers.

In one embodiment, the one or more components include a deep generative model configured to create the information for the nominal instances of the one or more specimens. For example, deep generative models that learn the joint probability distribution (mean, and variance) between the SEM (image of actual, wafer) and design (e.g., CAD or a vector representation of intended layout) can be used to generate the nominal instances that are used to train the machine learning based model, A generative model may also he used to generate other simulation results described herein for son- nominal, instances of the specimen. Once the machine learning based model, is trained for nominal (as i.ntended.½on-deieeiive) samples, as described further herein, a transfer learning training input dataset _» which includes defective images or other .non-nominal instances described herein, can be used, to re-train the machine learning based model, in addition, a machine learning based model can be pre-irained by using synthetic data thai is generated by modifying design data (e.g., CAD or EDA data) used to make

semiconductor wafers. Defect artifacts such as opens, shorts, protrusions, intrusions, etc. along with metrology markers such as line end pull backs could be inserted into the CAD and then fed into a generative model trained by a network, described in the above- referenced patent application by Zhang et al to creat realistic defects.

As further described above, a deep generative model may be used to generate the information for the non-nominal instances of the one or more specimens, The same deep generative model may be used to generate the information for the nominal instances of the one or more specimens, h this manner, a deep generative model may be used for generating nominal (reference) and defective (test) images or ther information described herein.

In one embodiments the machine learning based model is a discriminative model, in this manner, the embodiments may be configured to train a discriminative model. In addition, the discriminative model may be part of an inspection system (e.g., a light based narrow band or broad band inspector, electron beam based inspector, etc.) or other system (e.g., a metrolog system, defect review system, etc.) described herein. As such, the embodiments described herein may be configured to train various inspection and/or other systems for discriminative learning. The discriminative model may have any suitable architect tire and/or configuration, known in. the art.

Discriminative models, also called conditional models, are a class of models used in mach ine teaming for model ing the dependence of an unobserved variable y o an observed variabl x. Within a probabilistic framework, _, this is done by modeling the -conditional probability distribution P(yjx), which can be used for ^' predicting y from x. Discriminative .models, as opposed to generative models, do not allow one to generate samples from the joint distribution of x and. y. However, for tasks such as classification and regression that do not require the joint distribution, discriminative models can yield superior performance. On the other hand, generative models are typicall more flexible than discriminative models in expressing dependencies in complex learning tasks, in addition, most discriminative models are inherently supervised and cannot easily foe extended to unsupervised learning. Application specific details ultimately dictate the suitability of selecting a discriminative versus generative model. in another embodiment, the machine learning based, model is a neural network. For example,, the machine learning b sed model may be a deep neural network with a set of weights that model the world according to the data that it has been fed to train it. Neural networks can be generally defined as a computational approach which is based on a relatively large collection of neural, units loosely modeling the way a biological ^' brain solves problems with relatively-large .clusters ^' of biological neurons connected by axons. Each neural unit is connected with many others, and links can fee enforcing, or inhibitory in their effect op the .activation stale of connected neural units, These systems are self- learning and trained rather tha explicitly programmed and excel in areas where the solution or .feature detection is difficult to express in a traditional computer program.

Neural networks typically consist of multiple layers, and the signal path traverses from front to back. The goal of the neural network is: to solve problems in the same way thai the human brain would, although several neural networks are much more abstract. Modern neural network projects typically work with a few thousand to a few million, neural units and mi llions of connections. The neural network may have any suitable architecture anchor configuration known in the art.

In a forther embodiment, the machine learning based model is a convolution^ and deconvolution neural network. For example, the. embodiments described herein can ^' take advantage of deep Learning concepts such m ^' convolution and deconvolution neural network to sol ve the normall intractable representation conversion problem (e>g. _s rendering). The machine learning based model may ha ve any convolution and

deconvolution neural network configuration or architecture known in the art.

The computer subsystemC ^' s) arc also configured for retraining the machine Seaming based model with information for the no.n-iiomi.nal instances of the one or more of the specimens thereby performing transfer learning of the in formation for the non- nominal instances of the one or more of the specimens to the machine learning based model For example, general purpose of the embodiments described herein is to provide ^' systems and meihods for simulated and/or empirical inspection and/or metrology of semiconductor devices using efficiently trainable machine learning based models with a l imited training set To this end, a series of transfer learning methods can be used to enable and accelerate the efficient training of machine learning based models in a principled manner. These transfer learning methods and machine learning based models can be used for numerous inspection and metrology applications _*

Transfer learning can be generally defined as the improvement of learning in a new task (or a target task) through the transfer of knowledge trom. a related task that has alread been learned (one or more source tasks). In ^' the embodiments described herein, therefore, training the machine learning based, model with only the nominal instances involves learning the one or more source tasks, and re-training the machine learning based model with the non-nominal instances transfers the knowledge from the source tasks (the nominal instances) to the target task (the non-nominal, instances), in transfer learning, the agent knows nothing about a target task (or even that there will be a target task) while it is learning a source task. For instance, in the embodiments described herein, the machin lear ing based model knows nothing about the non-nominal instances while it is being trained with the nominal instances. in general however, the transfer learning described, herein may be performed in an suitable manner known in the art. For example, in an inductive learning task, the objective is to induce a predictive model from a set of training examples. Transfer in inductive learning works by allowing source-task knowledge to affect the target task's inductive bias. I an inductive transfer method, the target-task inductive bias is chosen or adjusted based o the source-task knowledge. The way this is done varies depending on which inductive learning algorithm is used to learn the source and target tasks.

Inductive transfer can be viewed as not only a way to improve learning in a standard supervisedTearMng task, but also as a way to- offset the- difficulties posed by tasks that involve relatively small dataseis. That is, if there are relatively small amounts of dat or class labels for task, treating it as a target task ami performing, inductive transfer from a related source task can lead to more accurate models. These approaches therefore use source-task data to enhance target-task data, despite the fact that the two dataseis are assumed to come from different probability distributions.

Transfer learning as described herein can be further performed as described in "Transfer lear ing," Torrey et ah, Handbook f Research on Machine Learning

Applications, published by 101 Gl bal _* edited by E. Sorja, I. Martin, R. Magdalena, M. Martinez and A. Serrano, 2009, 22 pages, and. "How transferable ^' are features in a deep neural net ork? ^** Yosihsfci et al , MIPS 2014, November 6. 2 14, 14 pages, which are incorporated by reference as if fully set forth herein. The embodiments described herein may be further configured as described in these references.

The re~ training (and training) architecture used by the embodiments described herein is preferabl designed to converge to the ground truth (for validation samples) with the minimum number of samples, in one embodiment, the one o more components include one or more additional components, the re-training is performed using the one or more additional components, and the one or more additional components include a common mother network for all layers on the specimens, a grand common mother network for all layers on the specimens, an adversarial network, a deep adversari l generative network, an adversarial auioeneoder, a Bayesian Neural Network* a component configured for a variational Bayesian method,, a ladder network, or some combination thereof. For example, the transfer learning methods ma he applied to the training set(s) ^• generated using any of the transfer learning training input generation method s described, herein to train the machine learning based models. There are many potential concepts that apply here. The ones, that have a substantially- high probability for working with semiconductor wafers include; using a common mother network for front end of line ( BOL), middle of line (MOL), and back end of line (BEOL) layers each; using grand common mother network for all layers (will likely work on SEM); using an adversarial network- to accelerate training; using a Bayesian Neural Network (Variational Bayes), which requires -far fewer layers; and using the concept of the ladder ne work for training. For example, the embodiments described herein are configured for semiconductor wafer and mask inspection and other applications and -accelerating training by "legally amplifying" samples. These methods are also known as semi-supervised (a few examples are available, but the vast majority are not labeled, by humans ^' or ground truth). m one such example, the -computer subs stem's) may leverage a pre-trained system of weights using the concept of a mother network. The computer subsystern(s) can also use methods s ^' uch as semi-supervised methods that combine Bayesian generative modeling to achieve their results in a minimum number of samples. Examples of such methods are described in U.S. Patent Application Serial No, 15/176,139 to Zhang et al. fi led June 7, 2016. and "Semi-supervised Learning with Deep Generative Models," ingma et al, NIPS 2014, October 31 , 2014, pp. 1 -9, which are incorporated by reference as if fully set forth herein.. The embodiments described herein may be further configured as described in. these references. In addition, the computer subsystem(s) may leverage ladder networks where supervised, and unsupervised learning are combined in, dee neural networks such as the ones proposed i "Semi-Supervised ^' Learning with Ladder

Networks," Rasmus et. al, NIPS 2015, November 24, 2015, pp. 1-19. which is

incorporated by reference as if fully set forth herein. The embodiments- described herei may be further configured as described in this reference. The computer subsystem(s) described herein may further be configured to tram the machine learning based model using a deep ad versarial generative network of the type described in "Generative

Adversarial Nets" GoodfeUow et at, June 10, 20.14, pp. 1-9, which is incorporated b reference as if fully set forth herein. The embodiments described herein may he- further configured as described in this reference. In addition or alternati vely, the computer suhsysiero(s) described herein ma be configured to train the machine learning based model using an. adversarial antoeneoder (a method that combines a variational

aatoencoder (VAE) and a deep generative adversarial network (DGAN)) such as that described in "Adversarial Autoenceders,'* Maktani et al., arXiv:1511.05644v2, May 25,. 2016, 16 pages, which is incorporated by reference as if fully set forth herein. The embodiments described herein may he farther configured as described, in this reference. In some instances, the computer subsystem^) may be configured to perform Bayesian Learning as described in "Bayesian Learning for Neural Networks," Heal, Springer- Verfag New York, 1996, 204 pages, which is incorporated by reference as if fully set forth herein. The embodiments described herei may be further configured as described iti this reference. The -computer subsystem(s) may also be configured to perform the variational Bayes method as described i "The Variational Bayes Method in Signal

Processing,' ^* Smidl, Springer- Verlag Berlin Heidelberg, 2006, 228 pages, which is incorporated by reference as if fully set forth herein. The embodiments described herein may be further configured as described in this reference, hi another embodiment, the re-training includes transferring all weights of eonvolutkmai layers of the trained machine- learning based method and ti ne tuning weights of folly connected layers of the trained machine learning based method. For example, the re-training may include- transferring all the weight of the convolution layers (e.g., layers 1-3) and then fine tuning the fully connected layers only. In another example, a machine learning based, niodei such as those described herein may be pre-trabe b a method known as transfer learning thai copies the weights from the mother network for the early layers and fine tones the latter layers where the -classification specialization occurs. Re-training may, however, Include al tering any one or more tram-able parameters- of- the- machine learning based model, i addition, the one or more parameters of the machine learning based model thai are trained by the embodiments described herein may include one or more weights for any layer of the machine learning based model that has trainable weights. In one such example, the weights may include weights for the convolution layers but not pooling layers.

Fig. 2 illustrates one embodiment of machine learning based model restraining that may be performed by the embodiments described herein. For example, one or more deep generative methods (DGMs) that can. learn nominal (designer intent) structures and generate simulated optical and SEM images may be used by the embodiments described herein to generate information for the nominal instances that is used to train the machine learning based model . In one such example, generative model 200 may be used to generate the information for the nominal instances. The generative model may learn the joint probability distribution (e.g., mean and variance) between an actual image of an actual wafer (e.g.. SEM image) and design (e.g., CAD or a vector representation of intended layout). The generative model may be configured to perform any other simulations described herein and may have any configuration described herein .

The infonn tion for the nominal instances may be used to train machine learning based model 206, which ma be performed as described further ^' herein. Once a neural network is trained for nominal (as intended.½on-def¾ciive) samples, the defect induced design may be used to generate synthetic data that looks substantially similar to the real world dat (such as the PWQ data and/or defect proxies designed into the structures printed on the wafers). For example, as shown in Fig. 2, design data 202 may be altered as described herein to generate synthetic defects in the design data. That altered design data may be input to generative model 200 to _: generate information ^' for the non-nominal instances (i.e., the synthetic- defects) of the one or more specimens based on the altered design data. The information for the non-nominal instances generated in this manner may also be provided to machine learning based model 206 for re-training of the model, which may be performed as described further herein. The design data and/or di altered design data may also be provided to the machine learning based model so that the design data and/or the altered design data ca be used for one or more steps performed by the machine learning based model

As further shown in Fig. 2, actual wafer data 204 may be. generated as described further herein (e.g., using PWQ and/or FEM training wafers). That actual wafer data may be provided to generative model 200, so that the actual wafer data may be used to train and/or update the generative model. That actual wafer data ma also be provided to machine learning based model 206 so that re-training of the machine learning based model may be performed using the actual wafer data, which may be performed as described further herein.

The machine learning based model described herein may be generated for specific specimens (e.g., specific wafers or reticles), processes, imaging parameters, etc. in other words, the machine learnin based models described herein may be specimen specific, process specific, imaging parameter specific:, etc. In this manner, different machine learning based models mav be generated for different wafer lavers. In addition, different machine learning based models tnay.be generated for different sets of imaging parameters (e.g., different imaging modes). Each, of the different models may be generated with different training sets of data. Each of the different training sets of data ma be generated as described further herein.

The embodiments described herein have a number of advantages described further herein. In addition, the embodiments described herein, can enable 100 ~ l OO

acceleration of achieving a trained image generation model usable for inspection and metrology applications.

Each of the embodiments of each of the systems described above may be

combined together into one single embodiment. Another embodiment relates to a computer-implemented method for training a machine learning based model. The method includes acquiring information for non- nominal instances of one or more specimens on which a process is performed, A

5 machine learning ^' based model is configured for performing one or more simulations for the specimens, and the machine learning based model is trained with only information for nominal instances of one or more additional specimens. The method also includes retraining the machine learning based model with, the information for the non-nominal instances of the one or more specimens thereby performing transfer learning of the i d information for the non-nominal instances of the one or more specimens to the machine learning based model. The acquiring and ega ning are performe by one or more computer systems. One or more components are executed by the one or mo ^' re computer systems, and the one or more components include the machine learning based model,

15 Each of the steps of the method may be performed as described further herein.

The method may also include any other step(s) that can be performed by the system, computer system(s), and/or machine learning based models described herein. The computer system(s may be c figured according to any of the embodiments described herein, e.g., computer subsystem(s) 102, in addition, the method described above ma be0 performed by any of the system embodiments described herein .

An additional embodiment relates to a .non-transitory computer-readable medium storing program instructions executable on one or more computer systems for performing a comptiter-irapleniented method for training a machine learning based model. One such5 embodiment is shown in Fig. 3. in. particular, as shown in Fig. 3, non-transitory

co.mputer-reada.bie medium 300 includes program instructions 302 executable on computer system(s) 3 , The computer-implemented method may include any siep(s) o any raethodfs) described, herein. Program instructions 302 implementing .methods such as those described herein may be stored on: computer-readable medium 300. The computer-readable medium m be a storage medium such as a magnetic or optical disk, a magnetic tape, or any other suitable non-transitory computer-readable medium known in the art.

The program instructions may be implemented in any of various ways, including, procedure-based techniques, component-based techniques, and/or object-oriented techniques, among others. For example, the program instructions may be implemented using ActiveX controls, O-f objects, JavaBeans, Microsoft Foundation Classes ("MFC"), S ^' SE (Streaming S1MD Extension) or other technologies or methodologies, as desired.

Computer systetn(s) 304 may be configured accordin ^' to: any of the embodiments described herein.

Further modifications and alternative embodiments ^' of various aspects of the invention will be apparent to those skilled in the art in view of this description. For example, methods and systems for accelerated training of a machine l earning based model for semiconductor -applications are provided. Accordingly, this description is to be construed as illustrative only and is for the purpose of teaching those skilled in the art. the general manner of carrying out the invention. It is to be understood t hat, the forms of the invention shown and described herein are to be taken as the presenrty _. prefe ^' rred embodiments. Elements and materials may be substituted for those illustrated and described herein, parts and processes may he reversed, and certain features of the invention may be utilized independently, all as would be apparent to one skilled in the art after having the benefit of this description of the invention . Changes may be made in the elements described herein without departing from the spirit and scope of the invention as described in the following claims.