Yield Management in LCD Fabs

In-process test and repair, especially when combined with careful data compilation, can markedly increase the manufacturing yield of LCD fabs.

by Tom Pye

YIELD MANAGEMENT is important in every commercial factory. In the liquid-crystal-display (LCD) industry, yield management is becoming a critical issue because of the increasing significance of the television market. TV adoption is primarily limited by price, which is driven by panel cost.

The conventional method of reducing the cost per panel by increasing the size of the motherglass has reached a point of diminishing returns. Because TV is driving larger displays that have more-complex pixel structures and more-stringent quality requirements, yield management is becoming an increasingly important factor in the LCD industry. Charles Annis, Vice President of FPD Manufacturing Research at DisplaySearch, says that "for a 32-in. LCD-TV module fabricated on a Gen 7 line, capital-equipment depreciation accounts for less than 10% of the total cost of production, while materials make up 75%. Decreasing costs by increasing motherglass size is becoming less and less effective, and scrapping heavily materials-burdened panels is extremely expensive. Yield improvement is a critical strategy for panel manufacturers; for example, a 1% increase in yield can save $10.5 million per year in wasted materials, while generating a substantial increase in potential revenue. Boosting yields by using more advanced and integrated yield-management systems is a critical mechanism in reducing the costs and growing the LCD market."

Fab Economics

Fab economics is driven by building, labor, and equipment costs. Balanced against these are the revenue and margins supplied by the produced panels.

Equipment Cost. Fab equipment cost has been increasing at roughly 30% per generation.1 As long as the increased glass area allows fab revenue to increase by more than this percentage, the fab's amortized equipment cost will decrease. The most recent proposed glass size, from 1870 x 2200 to 2160 x 2460 mm, increases the glass area by less than 30%. On a strictly area basis, the change in area does not pay for itself. In reality, there will be additional increases in revenue in these fabs because of a more-efficient product mix.

Materials Cost vs. Panel Size. In general, the cost of materials for a panel scales directly with area (Fig. 1). Even the backlight unit (BLU) costs are tied directly to panel area, although the linearity has discontinuities associated with changes in lamp count. The exception is glass, which has an anomalous rise for larger glass sizes until the required new factory costs have been fully amortized. When added together, the materials cost scales almost directly with area, maintaining a constant gross margin percentage for larger panels if the price of the panels also scales with area. In practice, the price of a panel scales with both area and pixel count.2 It is not known how much premium will be allowed for 1080p vs. 720p – with twice the pixel count – in HDTVs. Our analyses assume a 50% premium, which is less than monitor pixel-count premiums today.

Glass carries a larger penalty for larger sizes, so factories should utilize the smallest motherglass size that will support the production rate of the targeted panel sizes. To a lesser extent, the construction costs also scale to fab size, which is determined by glass size and throughput. If at all possible, previously established glass sizes should be used.

Photon Dynamics has developed a model that shows optimized revenue and margin for a given product mix vs. motherglass size (Fig. 2). The graph shows the output of a fab normalized to 32-in.-panel units for a range of glass sizes. The model indicates that a 2160 x 2460-mm fab is a good choice if the product mix is 26-, 32-, 45-, and 57-in. TVs. The next production plateau for this size is more than 30% higher than the previous one. The model scales revenue directly with area, and there is an additional 50% premium for larger panels at 1080p. Similar analyses for the product mixes of other manufacturers give results consistent with each manufacturer's announced next-generation fabs.

Fab Cycle Time and Scheduling. LCD-fab cycle times are measured in days, as opposed to semiconductor fabs where the cycle time is measured in months. At the same time, the fab throughput is three or more times greater than in typical semiconductor fabs. This places a large premium on optimized material flow, which favors in-line processing.

One example is the lithography cell, where the input track, exposure, development, and test functions are combined for the fastest cycle time. For array test and repair, the steps can also be done in-line, simplifying the automation with drastically lower overall fab costs. The automation cost advantage of in-line processing requires that the steps be at atmospheric pressure, so vacuum processes and vacuum test tools are excluded.

Yield Management and Modeling

General Poisson Model. The simplest yield model is the Poisson approximation of the binomial distribution.3 The equation is an exponential with defect density and defect- sensitive area as arguments. It is possible to create a model with yield presented as a function of typical fab defect densities and sensitive areas (Fig. 3).

The main point is that as each pixel becomes more complex, its sensitive area increases. But as panel sizes increase and fabs become larger, particulate levels also increase, having a profound effect on yield.

Systematic vs. Random Trends. The Poisson model is based upon the assumption that the defects are randomly distributed and independent of each other. These are reasonable assumptions for particulate contamination defects. Because they are random, these defects cannot be eliminated, except by redesigning to decrease the sensitive area or by cleaning up the fab. Identifying the defects is nonetheless very important because they can often be repaired, recouping yield loss.



Fig. 1: In general, the cost of materials for a panel, including the cost of the backlight unit, scales directly with area. The exception is glass, which has an anomalous rise for larger glass sizes until the required new factory costs have been fully amortized. (Data from Display Search Materials Costs, Panel Economic Model, March 2005.)


Systematic as opposed to random yield losses have root causes that can be eliminated. Systematic losses are characterized by signatures or patterns that are spatial, causal, or temporal (Table 1). Such losses are most prevalent at fab startup and ramp, but can also occur as excursions during mature production.

Revenue and Gross-Margin Savings. Fabs have a general lifetime that has a beginning period of startup and development followed by a ramp in both yield and production rate. This is followed by steady-state production with occasional design or capacity changes. An analysis of revenue and gross margin vs. yield shows two general cases. For new fabs that are building new-generation products, the average selling price (ASP) will start high and degrade over time (Fig. 4, left). This is the case for new TV fabs – Gen 6 and higher. In such cases, the speed of the startup and ramp is critical. The analogy for semiconductors is new-technology-node DRAM or microprocessor fabs, which make much of their profit in the first 12 months of operation.4

For a fab that is being built to address capacity in a mature market, the ASP is much more constant and only driven by general market conditions (Fig. 4, right). New Gen 5 fabs in China are examples. These fabs have faster ramps and a more constant stream of profit, but at a reduced percentage compared to the new-product fabs early in their production lives.

Fabs that utilize some level of yield management to enable the startup and production phases can be expected to show improved yields. The improvements are indicated by the "Cum GM from Yield Improvements" traces in the graphs shown in Fig. 4. This yield-management benefit is modeled as 1 month less of combined startup and ramp and 2% additional baseline yield.

An analysis of DisplaySearch data shows that ramp rates have been slowing down with every new generation.2 This has large effects on fab profitability for a new product market. The Gen 7 fabs will probably not reverse this trend.

Fab Yield Management. Fab product flow can be broken down into array, color filter, cell, and module assemblies (Fig. 5). The array typically consists of four or five mask steps for amorphous-silicon active-matrix LCDs (AMLCDs). Inspection is performed for particulates in incoming glass and during most of the mask steps. Automated optical inspection (AOI) is performed after the etch steps and, increasingly, after the lithography development. Repair is also often performed after etch. Functional tests, binning, and repair are performed when the array is finished.

Separate inspection, repair, and binning sequences are conducted during the color-filter-fabrication sequence. The cell assembly consists of placing the liquid crystal between the color filter and array. A test, repair, and binning sequence is then performed. When the driver circuitry is added to the cell, another test and repair sequence is conducted. Finally, when the backlight and additional films are added, final test, repair, and binning of the module are performed.

Process and Line Monitoring

Automated Optical Inspection. AOI is conducted at several steps in the array process flow. Its primary value is as a process diagnostic tool during the fab startup and ramp. Early feedback of process problems allows a faster ramp time. In production, the AOI data allow process excursions to be identified before subsequent steps are performed. Both steps are significant contributions to the overall benefits of yield management.



Fig. 2: Because glass carries a large penalty for larger sizes, a factory should utilize the smallest motherglass size that will support the production rate of the targeted panel sizes. Photon Dynamics has developed a model that shows optimized revenue and margin for a given product mix vs. motherglass size. The graph shows the output of the fab normalized to 32-in.-panel units for a range of glass sizes. The model indicates that a 2160 x 2460-mm fab is a good choice if the product mix is 26-, 32-, 45-, and 57-in. TVs. The next production plateau for this size is more than 30% higher than the previous one.


In the future, we should be able to achieve increased functionality by using AOI. Noam Cohen, Director of Flat Panel Marketing at Orbotech, says, "The more-sophisticated AOIs may also offer further analytical capabilities in addition to the optical detection of defects that will help the process engineer stabilize or tweak the process. These may be described as automated defect classification, whereby high-magnification images of the defects would be taken and automatically analyzed to determine the type/origin of the defect found and create process failure statistics." They may also be critical dimension and overlay measurements, or a "Digital Macro" feature that helps operators find large-sized mura defects.

Open/Short Testing. There are two methods for open/short testing. The first is basically an extension of the AOI data by analysis and binning. This can be perfromed after lithography and after etch. The second method uses coupled, injected signals to perform an electrical test. This method requires that conductors be present, so it cannot be used as an after-development test in a lithography cell. Both methods allow simple repair of the lines because there are no layers on top of the lines.

Array Functional Test. The array test is a 100% electrical test that is used to find actual bad lines and pixels. Typically, orders-of-magnitude-more defects are identified by AOIthan are found by array functional test. Review can be reduced to the actual electrical defects and any optical-defect types of interest. Repair can similarly be limited to actual electrical defects.

Array functional testing is performed in three main technologies. Voltage imaging, which uses a small liquid-crystal-enabled electro-optic sensor to map the voltages on the array, is the industry standard. It has the advantage of simultaneously mapping the entire array surface under the sensor, and thus addresses any subpixel structures. Voltage imaging is performed at atmospheric pressure and easily integrates into in-line-automation structures. Shorting bars are used to drive the panel, lowering the contact-structure costs and providing higher-reliability contacts.

Another technique is full-contact probing. It has the advantage of good signal resolution, but the reliability and cost of the full-probe frames have limited its usage in later-generation fabs.

The final technique is Ebeam imaging. It has good spatial resolution, but it is a spot measurement so test time per pixel increases as multi-domain dual-drive pixels are incorporated in displays. Signal resolution is limited by the charging noise of the surroundings. In addition, the measurement is performed in vacuum, so it is harder and more expensive to integrate it into in-line fab flows.

Repair. Repair is the most important step in yield management. Very low ramp yields can be brought up to near-production levels by repair. This adds revenue early in the fab's life during the high-ASP period.

The highest-performing fabs always over-buy repair in their first phase to ensure that there is adequate capability during the ramp. As the fab adds capacity, the initial extra repair capacity is consumed by normal production usage.

The savings to the fab are extremely large. In a 6-month ramp, with raw (unrepaired) yield starting low and improving to a mature production level over time, the recovered revenue is roughly 50% more than it is without repair. This can add $35,000,000 in revenue for a typical TV-fab startup.

In production, repair allows the baseline yield to be kept at a higher level. It can also mitigate excursions if the quantity of defects per panel is reasonable, and it can be used to classify defects for yield learning. Again, the savings are much larger than the cost of repair. Over a 60-month depreciation period, the savings at a single fab can easily reach hundreds of millions of dollars of revenue.

Chemical Vapor Deposition (open repair). Open circuits are usually repaired by performing localized chemical vapor deposition (CVD). The technique is slow compared to cut/weld repair, but panels typically have far fewer opens than shorts. Scheduling and fab flow are difficult to determine because even a few percent of the defects being open will ensure that at least one glass per cassette requires open repair.



Fig. 3: In this general Poisson model, LCD fab yield is presented as a function of typical fab defect densities and sensitive areas. As a pixel becomes more complex, its sensitive area increases. Unfortunately, as panel sizes increase and fabs become larger, particulate levels also increase. This interaction has a profound effect on yield.


Zap (cut/weld repair). Cut/weld repair is accomplished by cutting open the defect usinglasers of various wavelengths. It can also weld redundant traces that effect repairs. A review of the defects is incorporated into these machines so that binning verification can be performed. Data from both the array functional test and AOI are fed into systems that permit review before the repair is made. Multi-wavelength laser arrays are used because the system must be able to selectively remove particular layers.

Color Filter

Automated Optical Inspection. Pattern inspection is conducted with machines capable of working with both transmitted and reflectedlight, and separate RGB channels are used. The resolution is slightly relaxed compared to that of array inspection because the structures being inspected are larger.

Repair. Color-filter repair involves ablative processes [zap (cult/weld) and/or grinding] and deposition of color materials. As in the array, the ablative repair is more prevalent than the deposition. Some tools involve combining techniques that offer simpler fab scheduling and transport. As with arrays, the repair is of high value because it permits the recovery of lost filters. It is also valuable in avoiding a fatal defect, protrusions on the common plane that could short the array.

The Cell. The cell is defined as the integration of the array, liquid crystal, and color filter, along with the separation of the panels from the original motherglass.

Automated Visual Inspection

Automated visual inspection (AVI) is being applied to the cell as TV fabs become more prevalent because larger panels are humanly difficult to inspect. The machine capture rate is higher and much more uniform than the human capture rate. Uniformity is important because it is impossible to implement yield learning if the variability of the measurement is equal to the variability of the defects. In addition, the automated systems can analyze and report their data much more thoroughly to allow repeatable, accurate binning and root-cause analyses.

Human Visual Inspection. Despite the advantages of AVI, humans are still used to perform some cell inspection. The number of inspectors is being reduced, however, because they are used mainly to perform quick inspections in order to make close binning decisions. A typical line is likely to have its own auto-mated inspector, capable of inspecting multiple display sizes, supported by a manual backup station. The total number of AVI and HVI lines is dictated by the production rate, but full production fabs can have more than a dozen.

Repair. Repair at the cell level is purely zap. It is typically done through the back of the panel to avoid damage to the color filter and to permit the use of the broadest possible spectrum of lasers for repair. Because the laser light must pass through a complex backplane structure, multi-wavelength capability is critical. The efficacy of this repair is much lower than in process or in the array. Nonetheless, because the cell has almost one-half the materials cost of the panel incorporated, attempting repairs at this level is a critical component of yield management.


Table 1: Causes of systematic yield losses
Spatial Edge defects, radial uniformity issues, mask defect
Causal Under etch, over expose, film contamination
Temporal Shift (AM/PM) related, Process tool maintenance timing related



Fig. 4: In new fabs that are producing new-generation products, the average selling price (ASP) will start high and degrade over time (left). For a fab that is being built to address capacity in a mature market, the ASP is much more constant and only driven by general market conditions (right). In either case, yield-management techniques applied during a fab's startup and production phases can expect improved yields, as shown by the improved yield traces.


The Module

Driver-Install Test/Repair. When the drivers and the backlight unit (BLU) are added to the completed cell, the resulting assembly is called the display module. A separate inspection is made after the driver circuitry is installed. Because any defects from bad connections will appear as line defects, automated inspection is very effective. The speed, reporting, and accuracy of AVI will lead to its rapid replacement of HVI for this inspection.

Final Test. At this stage, the BLU is added to the panel, inspected, burned-in, and given a final inspection. As in cell inspection, a mixture of AVI and HVI is used. At present, the mix leans toward HVI because large, low-contrast mura (stain) defects, bezel defects,etc., are reviewed and final binning is verified. As in cell inspection, there can be over a dozen AVI plus HVI lines in production with multiple HVI stations associated with each AVI station.

Predicting the Future

A firm lesson from the semiconductor industry's implementation of yield management is that the data collection should be automated as much as possible.5 This leads to close integration of the inspection, test, and repair data. Photon Dynamics is already involved with customers to link this data and analyze the results.

The benefits are many. By seeing what defects are escaping from array to cell, array inspection, test, and repair recipes can be optimized. Similarly, color-filter repairs can be checked to ensure that their repair recipes are optimized. Module data can be fed back to optimize cell-inspection recipes, especially for the lower-contrast mura defects.

It is clear that cell and module AVI will increase because it enables more-effective yield management by providing complete, accurate data from the back end, which can then be fed back to the front-end processes. All major manufacturers are investigating or incorporating this approach in their most advanced fabs. As the industry matures, cell and module AVI will also be back-fitted into older lines to optimize the fabs and obtain higher operating margins. The functionality of the inspection may be increased, as individual panel characterization such as speed, contrast, and color are incorporated into the highest-performance panels.

Another integration trend is the increase in functionality of the repair systems. A review function, for instance, has recently been incorporated. Other functions are being evaluated, including combining open and short repair and adding functional tests to the repair systems to perform immediate verification of the repair.

A longer-term trend is the optimization of factory automation. A move will be made toward flow or in-line organization of the fab. As mentioned previously, in-line manufacturing benefits from atmospheric processes and measurements. This issue will become even more pressing when roll-to-roll manufacturing begins.


1Display Search Fab Database, July 2005.

2J. Hawthorne, "Size Matters but Yield Kills," SID 2005 Business Conferenece.

3A. Ferris-Prabhu, Introduction to Semiconductor Device Yield Modeling (Artech House, Inc., 1992), pp. 28–30.


5ibid. •



Fig. 5: When applied to the array, color filter, cell, and module processes, yield-management techniques can significantly improve panel yields and plant profitability.


Tom Pye is Director of Marketing at Photon Dynamics, 5970 Optical Court, San Jose, CA 95138; telephone 408/226-9900, e-mail: tom.pye@photondynamics.com.