Tải bản đầy đủ - 0 (trang)
3 Challenge of IC Production—Prerequisites for Efficient Product Ramp-up

3 Challenge of IC Production—Prerequisites for Efficient Product Ramp-up

Tải bản đầy đủ - 0trang

9 Product Ramp-up for Semiconductor Manufacturing …


which are assigned to a particular IC product in order to ensure proper treatment. At

the same time the yield (i.e., the ratio of good ICs compared to all produced ICs)

has to be around 98 % for the production system to be profitable. To achieve this

goal, every single process step has to be executed in a flawless manner (Xiao 2012).

Therefore, in modern IC factories, almost every single process step is followed

by a metrology step which measures samples and production-IT which adjusts the

process setup in real time and triggers corrective actions immediately and automatically at the event of excessive process variations or error-related process

changes. Throughout the remaining chapter, an arbitrary sequence of process steps

followed by one or more metrology steps shall be named process segment.

Referring to the previous sections, we remember that a process segment represents a

single node within the hierarchical composition structure of the production process.

Over the years, the production-IT was enhanced by a variety of powerful control

components, which run generally fully automated. These control components

usually require a product-specific control setup as it was introduced with the simple

example of checking the color of the cake while baking it in the oven in the

introduction Sect. 9.2.1. The most common control components currently in use are

shown in Fig. 9.3 and provide the following functionality:

Fig. 9.3 Assignment of data collection and control components to process segments and

production resources (e.g., equipment)


R. Willmann and W. Kastner

• Statistical process control (SPC) aligns data of measured samples with historical

trends and control limits which are based on observed statistical data. Notifications to process engineers or immediate corrective actions are triggered

automatically if an instable behavior of the controlled process is recognized

(Dietrich and Schulze 2003).

• Run to run control (R2R) complements SPC. While SPC monitors metrics and

triggers corrective actions in case of deviations, R2R acts through readjustment

of processes before a deviation occurs. R2R uses pre- and post-process measurements of samples (individual wafers in case of IC production) which are

taken out of the controlled process, as well as the expected quality target (e.g.,

the thickness of a metallization layer which was deposited to the wafer) and

knowledge about the correlation between this quality target and a few setup

parameters of the controlled process (e.g., the duration of the deposition process

and the temperature in the process chamber during metal deposition).

• Feed forward control/feed backward control (FF/FB) considers dependencies

along the process sequence. For example, the layer thickness of the photo

resistor after the lithography process can affect the subsequent reactive ion etch

process (Ruegsegger et al. 1998). An FF model holds the information, how the

reactive ion etch process recipe has to be adjusted for compensation of photoresistant thickness. FB can be understood similarly to R2R with the only

difference that not the immediately last process step but the setup of another one

in the process history is adjusted.

• Data Collection—the previously introduced control components require production data to be recorded with high quality. Therefore, also production data

collection models need to be individualized for each product.

This structure of process segments and associated control components results in

hundreds of control setups which have to be reviewed and adapted during product

ramp-up, with the important purpose to ensure the proper monitoring and control of

the quality of occurrences of the new product during the later volume production.

In order to reduce complexity also during a product ramp-up, IC manufacturers

have introduced the concept of process technologies and silicon intellectual property (silicon IP). Silicon IP comprises off-the-shelf functions like A/D converters,

memory, and processors (Nenni and McLellan 2013, p. 19) which can be randomly

combined during the design of a new IC product. Process technologies (e.g.,

CMOS) are commonly used by IC manufacturers as well.

A process technology is defined by a specific stack of material layers which is

built by a particular sequence of process segments on a silicon wafer. Also the size

range of features on each layer is specific for process technologies. Each process

technology is therefore linked to a dedicated set of process plans which can be

performed to build the requested stack of layers. Therefore, process technologies

represent a link between the process plan and the design of specific IC products.

To some extent process technologies can be seen as reusable templates in order

to build individual IC products. Variations of individual products are achieved by

9 Product Ramp-up for Semiconductor Manufacturing …


modification of the layout of single layers, the thicknesses of layers, controlled

impurities of semiconductor material, or other means of parameterization.

Because of this branch-specific method of process technologies and modularization of products by the use of silicon IP, the semiconductor industry provides a

good pattern for customization of products using templates and unified underlying

production processes. Some other industries, like the production of printed circuit

boards (PCBs) apply this method as well (Macleod 2002).

Also the automotive manufacturing industry uses the concept of product templates. In this domain, product templates are called platforms. An outstanding

example of a platform approach is Toyota’s policy concerning its car models.

Toyota is currently launching new generations of its successful car models that

utilize more than 70 % of their forerunners’ components, and the platforms of these

cars have remained largely constant through successive car generations (Hüttenrauch and Baum 2008; Slamanig and Winkler 2012, p. 486).

The concepts of process technologies and of product platforms can be considered similar. It has a general validity in the manufacturing industry for reduction of

the complexity of management of product variants and therefore the complexity of

product ramp-up projects. Facing the emerge of Industrie 4.0, this method will be

rather likely adopted by other manufacturing industries as well.

Driven by the complexity and quality demands of IC production a comprehensive set of control components was developed as part of the production-IT.

Setting up those control components in order to meet product-specific needs is an

essential task during a product ramp-up project. In order to master the complexity of

a production process and therefore of product ramp-up projects, concepts of process

technologies or product platforms were introduced. Such concepts support the reuse

of existing production knowledge for new products.

From the perspective of K-RAMP, the approach of templates or platforms helps

to develop an algorithm which determines reusable subproducts or process segments of forerunner products while planning or performing ramp-up projects. In the

next sections, K-RAMP is introduced as a process for gathering such existing

knowledge and matching it with the needs of a new product.


The Process Perspective of K-RAMP

In the sequel of this section, an overview of K-RAMP is provided. This section is

followed by a discussion of the involved ontology models and the necessary

architecture of software components which envelopes a Semantic Web-based

knowledge base. An essential part is the mapping of the product design—in particular quality-relevant characteristics of the product design—to the configuration

settings of control software components. However, in order to determine reusable

configuration settings, appropriate and thus reusable process segments and subproducts have to be found in advance. The corresponding process is described in

the next sections.


R. Willmann and W. Kastner

Successful searching of reusable subproducts of forerunner can be realized by

matching them with the quality characteristics of the new product. Searching for

reusable process segments is supported by the mapping between the composition

structure of products and the composition structure of the production process as

introduced in Sect. 9.2.2.

The principle idea of this mapping was introduced as axiomatic design by Suh

(1990, 2001). It describes the technique of mapping (Fig. 9.4) between concepts of

the customer domain, the functional domain, the physical domain, and the process

domain. In particular, customer attributes (CA) of the customer domain are mapped

to functional requirements (FR) of the functional domain, design parameters

(DP) of the physical domain are mapped to FRs, and process variables (PV) of the

process domain are mapped to DPs. This mapping is performed on several levels of

a tree-like composition structure. The mappings between CAs and FRs, as well as

between FRs and DPs are closely related to quality function deployment

(QFD) (Breyfogle III 2003; Said El-Haik 2005). However, the mapping between

CAs and FRs is not covered within K-RAMP as it is assumed to be part of the

preceding product design phase.

The essential parts of the axiomatic design which have been considered with

K-RAMP are the mapping between the functional domain and the physical domain

in order to determine reusable subproducts, and the mapping between the physical

domain and the process domain in order to determine reusable process segments.

In the sequel of this chapter, particularly the mapping of the physical domain and

the process domain is in the focus. According to Suh, it is the association between

Fig. 9.4 Mapping between domains of axiomatic design based on Suh (1990)

9 Product Ramp-up for Semiconductor Manufacturing …


DPs and PVs. The concept of DPs of axiomatic design shall be equivalent with the

concept of products or subproducts, and the concept of PVs shall be equivalent with

the concept of process segments during the following discussion.

The focus of this chapter is the recommendation of control system setup as

outcome for the ramp-up of a new product. Automated recommendation of control

system setup for new products requires knowledge about the currently applied

setup. From Fig. 9.3, it can be seen that control system setup is associated with the

process segments which shall be controlled. K-RAMP assumes that every process

segment is capable to produce a particular category of subproducts and that the

applied control models are equivalent for all subproducts of this category, at least

on a certain level of generalization. Deducing associations between new subproducts and already existing process segments is therefore essential for an automated

recommendation of the setup of control software components (Question 2). Ahead

of this particular topic, the approach for gathering reusable process segments from a

new product’s design has to be discussed (Question 1). How K-RAMP solves

Question 1 is illustrated in this section.

Generally, in all three relevant knowledge domains of K-RAMP (i.e., the

functional domain, the physical domain and the process domain) each node of the

composition structure is specified by a set of characteristics. Moreover, nodes of the

composition trees of mapped knowledge domains correspond to each other. For

instance, a particular node of the process structure (a process segment) requires an

arbitrary number of nodes of the product structure in order to satisfy a particular

other node of the product structure on the next higher composition level (see

discussion in context of Fig. 9.2).

The V-model in Fig. 9.5 shows the overall domain of production knowledge as

it is contained within the K-RAMP knowledge base. It covers a design perspective

where elements of the functional domain, the physical domain, and the process

domain are mapped to each other, and it covers a control side where the connection

to the physical representatives of the production-IT is modeled. On the design side,

the nodes of the composition structures of each domain are specified by sets of

Fig. 9.5 V-Model of design and control in manufacturing


R. Willmann and W. Kastner

characteristics (specifications). By the use of these specifications, it is possible to

determine weights of the associations satisfies and requires as used in Fig. 9.2. It is

for instance possible to determine the coverage (satisfaction) of a subproduct’s

specification through a particular specification of a process segment.

Each domain of the design has a counterpart on the control side of the V-model.

The control side is implemented in the production system through the use of control

software components. These components ensure that the specifications of the design

side are fulfilled. The specifications of the design side are therefore related to the

setups of the control software components on the control side of the V-model.

The elements of the K-RAMP knowledge base do not provide any mappings to

the layout of the physical production system. These equipment engineering topics

are covered by other research papers, like Moser et al. (2010), Winkler and Biffl

(2010), Moser et al. (2011).

The specifications of the process domain interact with the setups of each process

segment. For instance, there might be a specific processing duration specified for a

heating process. The recipe adjustment as it is implemented in the production

system uses the specified processing duration in order to combine it with a couple

of other settings of an equipment recipe. All settings together are controlling an

automated procedure of the equipment or a machine.

The specifications of the physical domain are associated with the setups of

control software components in the inline metrology area. For instance, the specification of a subproduct may require a layer thickness within a particular range of

tolerance. The inline metrology of the production system measures this thickness.

The setup of the SPC software component or the R2R software component ensures

that appropriate actions are performed if the thickness is running out of range.

FRs of the functional domain are associated with the quality assurance of the

production system. For instance, the electrical capacitance of a layer is specified by

an FR in the functional domain. The measurement of this electrical capacitance is

performed during the quality assurance of the production system. Through

product-specific setup of the SPC software component and appropriate actions, the

electrical capacitance is kept within the limits during volume production.

At the control side of the V-model, production data of forerunner products are

collected, which comprise of data of handling instructions of resources (e.g., a

process duration been set up), inline metrology data (e.g., the thickness of a metallization layer, but also quality measurements of received material from suppliers),

and quality assurance data (e.g., the capacitance of a layer). These production data

are an important source of information for the following reasons:

• Across the production data it is possible to determine dependencies and correlations. For instance, the handling instruction of a resource—the process duration—correlates with some characteristic being measured during inline

metrology—the thickness of a deposited metallization layer—of this process

segment. On the other hand, the characteristic which is measured during inline

metrology correlates with a characteristic which is measured during quality

assurance—the electrical capacitance. To some extent, such correlations can be

9 Product Ramp-up for Semiconductor Manufacturing …


used to control the result of process segments or the overall production process

and thus the function of the final product.

• Between the control side of the V-model and the design side of the V-model it is

possible to determine whether the production resources, process segments, or

subproducts of the targeted production system are capable to produce instances

of the new product in accordance to its specification.

These dependencies are used by K-RAMP to answer Questions 1 and 2. In the

sequel of this section, the process perspective of K-RAMP is explained. This

process perspective is equivalent to the one which has to be followed by the

ramp-up teams manually today. The result of this discussion is a list of requirements

which has to be satisfied by a knowledge-based approach.

K-RAMP is performed recursively on every level of composition of the structure

of FRs, the (sub-)products, and the process segments. The first and most obvious

step is to determine existing (sub-)products which already match the characteristics

of a new (sub-)product, or which satisfy the FRs of the new (sub-)product.

Matching between existing (sub-)products and new (sub-)products is based on

the left side of the V-model only. However, Semantic Web’s open world

assumption (OWA) only uncovers not-matching (sub-)products, but it is not possible to determine a set of matching (sub-)products. Therefore, an external

matchmaking algorithm has to support the reasoning.

The coverage of new FRs by existing (sub-)products can be rather likely derived

by overlapping quality assurance data of existing (sub-)products (right side of the

V-model) and the specifications of new FRs (activity A in Fig. 9.6).

Fig. 9.6 Overview of K-RAMP


R. Willmann and W. Kastner

Before detailed studies on the relationship between process segments and

characteristics of (sub-)products, K-RAMP tries to find existing (sub-)products

whose characteristics overlap with new (sub-)products. If the specification bands of

all characteristics of an existing (sub-)product are enclosed by the specification

bands of the corresponding characteristics of a new (sub-)product it can be implied

that every occurrence of the existing (sub-)product is also a member of the

enclosing new (sub-)product. Consequently, the process segment which creates the

existing (sub-)product should be also capable to produce the new (sub-)product.

In Fig. 9.7, there are three (sub-)products PX, PY, and PZ, all being members of

the same product class P. Product classes, for instance, are chocolate or copper

metallization layer, while the (sub-)products represent specific brands of chocolate

vendors or specific interlayer dielectric layers (ILDs) of an IC’s process technology

(Xiao 2012, pp. 371−374). Each produced occurrence (or sample) is a member of a

(sub-)product. Each (sub-)product in Fig. 9.7 has characteristics C1 and C2 which

are in common with other (sub-)products—particularly if they are of the same

product class. However, for each (sub-)product these characteristics have different

specification ranges, meaning the value ranges which are allowed without harming

the quality of the respective (sub-)product. In Fig. 9.7, specification ranges are

written in a form where the subscripted letter refers to the (sub-)product and the

superscripted number refers to the characteristic.

R1Y is the specification range of characteristic C1 for (sub-)product PY.

In Fig. 9.7, there is also a subclass of (sub-)product PZ named PZ’ which shall be

discussed later as well.

Fig. 9.7 Examples of mutually enclosing specification ranges

9 Product Ramp-up for Semiconductor Manufacturing …


How is reusability in the product domain determined? It is not useful to try

matching between all existing and all new (sub-)products. Each new (sub-)product

is used to determine related existing (sub-)products instead. For this purpose, the

ontology model of K-RAMP provides a generalization structure of (sub-)product

classes. Based on the set of related (sub-)product classes their members have to be

found which are enclosed by the new (sub-)product. The list of characteristics and

the limits of their specification ranges are the defining elements of memberships in

(sub-)product classes. A (sub-)product class is more specific, the more characteristics are necessary to specify its members. A (sub-)product class C is a subclass of

another (sub-)product class D if the list of characteristics of C is a superset of the list

of characteristics of D and the specification range of each common characteristic in

D encloses the respective characteristic’s specification range in C.

According to Fig. 9.7, PY is enclosed by PX because R1X fully encloses R1Y , and

R2X fully encloses R2Y . Another already existing subproduct PZ is not enclosed. The

specification ranges R1X and R1Z are only overlapping each other. This is the most

trivial decision-making to determine reusable (sub-)products and, consequently,

associated process segments being capable to produce the (sub-)products.

Having, for instance, a look to PY and PX, the decision is obvious. Occurrences

of the forerunner product PY are already produced with large volume and sufficient

high yield. For this reason, the majority of measured characteristics C1 and C2

which are taken from samples in a process segment while producing the subproduct

PY are within the specification ranges R1Y and R2Y . So it is rather likely, that if PX is

produced by the same process segment the majority of measured C1 and C2 of

samples of PX will be within the even wider specification ranges R1X and R2X .

However, even (sub-)products whose specification ranges are not fully enclosed

by the equivalent specification ranges of a new (sub-)product can be potential

candidates for reuse. This situation is demonstrated by (sub-)product PZ in Fig. 9.7.

Occurrences of the (sub-)product PZ—which are represented by measured samples

of the related process segment—might be within the specification range R1Z of (sub-)

product PZ and within the specification range R1X of the new subproduct PX as well.

It is now the challenge of K-RAMP (activities C and D in Fig. 9.6) to find a

systematic cause why parts of the samples also fit to the new specification range of

PX while others do not. If no such cause is found the ramp-up team has to be aware

about this knowledge gap and has to plan the development of a new process

segment (activity I in Fig. 9.6) which covers the needs of PX independently of

existing process segments.

However, if such a systematic cause is determined (e.g., a specific range of the

process duration or the applied heating temperature) it can be used as supplemental

constraint of a new handling instruction which is derived from the handling

instruction of the existing process segment. This new handling instruction can be

the baseline for the new process segment. The occurrences (samples) of subproduct

PZ which are produced due to the systematic cause become members of (sub-)

product PZ’ which is a subclass of PZ. Both are still members of the same product

class P.


R. Willmann and W. Kastner

The coverage of a FR by (sub-)products shall be above an appropriate threshold

in order to determine them as appropriate reusable (sub-)products with some certainty. With respect to this coverage, Suh recommends to calculate the entropy as

introduced by Shannon. In K-RAMP, the theoretical possible first pass yield (FPY)

with a range between 0 and 100 % is used to determine how much the members of

PZ’ and therefore the derived handling instruction of the new process segment

would cover the specification range R1X . If the theoretical FPY is above a certain

threshold the existing subproduct can be considered for reuse.

The FPY is calculated as the portion of defect-free parts of a specific (sub-)

product which are passing a particular process segment of the overall production

process (Wappis and Jung 2010, pp. 179−180) at the first pass (without rework).

Defect-free parts are within the respective specification ranges of all given characteristics of a (sub-)product. During production of the forerunner products, this

condition is evaluated against the specification ranges of the forerunner’s (sub-)

products. However, K-RAMP uses the historical measurements of forerunner

products and evaluates them against the specification ranges of the new (sub-)


It may happen that there is no (sub-)product available above the threshold and

the search result is empty. However, if there is enough information available

concerning the correlation of data from inline metrology of individual characteristics and the characteristics of functional requirements (from quality assurance)

then it may be possible to determine a regression function. If this regression

function is part of the knowledge base as well, it can be used to calculate the ideal

adjustment of the existing (sub-)product’s characteristics in order to maximize the

coverage of the FR of a new (sub-)product. A similar approach is also possible by

considering the correlation between existing (sub-)product’s characteristics and

setup parameters from the handling instruction of the related process segment.

Reuse of an existing (sub-)product or a process segment by adjustment is considered in activity F of Fig. 9.6. If, for instance, PZ, does not perfectly cover the

specification ranges of PX it can be tried to evaluate potential shifting of setup

parameters by the use of correlation models in order to improve the coverage.

For instance, as demonstrated in Fig. 9.8, by keeping a dedicated process temperature and particular deposition material—both are parts of the handling

instruction of a process segment—it is possible to alter the Deposition time for the

purpose of layer thickness (axis d) adjustment. As a part of the continuously

ongoing quality monitoring, the layer thickness measurements of samples of forerunner (sub-)products are used to determine the shape of the distribution function of

this process segment and to calculate the average and the standard deviation. Based

on these indicators and the specification range of the layer thickness (upper specification limit USL1 and lower specification limit LSL1) it is possible to calculate

the FPY of the forerunner (sub-)product with respect to the layer thickness.

The regression function of the deposition time and the layer thickness, as well as

the shape of the distribution function and its standard deviation is now used by

K-RAMP. An adjusted deposition time for the layer thickness of the new (sub-)

9 Product Ramp-up for Semiconductor Manufacturing …


Fig. 9.8 Impact of shifting process segment settings on shifting of product design based on

(Smietana et al. [2014])

product (specification range LSL2 to USL2) is determined. Knowing the specification limits of the new subproduct’s layer thickness it is possible to calculate

(1) the adjusted deposition time by applying the regression function on the new

layer thickness and (2) the attempted FPY by shifting the distribution function in a

way that the average is equal to the target T2 of the new subproduct’s layer

thickness. It is important to mention that there might be more than one characteristic, like the layer thickness in our example, and the regression function could be

multivariate accordingly.

By performing activity F of Fig. 9.6, it is therefore even possible to determine

appropriate process segments and their handling instructions if there is no immediate coverage visible. If adjustment is no option because the resulting quality

(determined FPY for new subproduct) is below a certain threshold then a gap was

considered and the ramp-up team must be notified (activity I in Fig. 9.6).

Based on available information about existing subproducts and process segments, it is thus possible to perform matchmaking with the new subproduct.

A similar process of decision-making needs to be performed by ramp-up teams

during the ramp-up of every new product in every branch of the manufacturing

industry. In addition, all new or adjusted handling instructions of process segments

also require appropriately modified setup of aforementioned control software procedures. An increasing number of process segments and (sub-)products increase the

efforts with respect to information which needs to be collected, and it increases the

risk of failures which are caused by the lack of information and knowledge

exchange. K-RAMP comprises the activities of ramp-up teams in a semiautomated

process as it was described in the previous sections.

Keeping the recursive approach of K-RAMP in mind, the final situation is a

structured sequence of device-specific process segments including handling

Tài liệu bạn tìm kiếm đã sẵn sàng tải về

3 Challenge of IC Production—Prerequisites for Efficient Product Ramp-up

Tải bản đầy đủ ngay(0 tr)