Adaptive sampling technology - the next step for factory efficiency

As the semiconductor industry moves to ever smaller geometries in its manufacturing processes, a change in the criticality of metrology data is also occurring. Through most of the history of IC manufacturing, devices could be manufactured in processes that used metrology data for monitoring and excursion control purposes. The processes ran effectively in an open-loop mode with corrections being made manually based on offline review of the collected data. In such scenarios, metrology is not a mission-critical portion of the manufacturing process.

Within the last few years, this has changed so that newer technologies can no longer be manufactured profitably with the natural openloop variation of the process. Automated closedloop control is required to tighten product distributions to meet the design constraints of advanced manufacturing. This change moves metrology into a mission-critical role. Without the metrology to feed into automated control systems, the factory cannot produce.

An effect of this move towards mission-critical metrology is the tendency to add more metrology capacity to the factories. While this makes control engineers (and metrology tool vendors) happy, it has negative impacts. The tools take up valuable factory floor space, the measurements increase cycle times of products, and the data still do not provide inherent value to the factory. These side effects do not make factory managers happy.

The key to addressing this issue is the ability to strike a balance between the measurement needs of automated control and factory monitoring systems and the needs of the factory management to make the most effective use of available factory resources. At AMD, we believe that the solution to finding this balance is in the development of adaptive sampling technologies.

When we speak of "adaptive sampling" what we are really talking about is efficient and effective use of metrology sampling capacity. In other words, the finite metrology capacity of the factory is allocated to those measurements that provide the best actionable data for the factory.

The idea of allocating metrology capacity is not a new idea. Sampling plans have been in place for years that allocate higher sampling rates to some processes rather than others. Measurements for processes with small process windows or those with high correlation to end-of-line yield may be given higher percentages than other, less critical measurements. Operations that have high wafer-to-wafer variability may measure more wafers but a smaller percentage of lots. While these strategies are effective in providing the data they are designed to collect, they have the fundamental problem of being static - and modern factories are anything but static. Truly efficient sampling strategies must be able to adapt to the changing factory dynamics.

Semiconductor manufacturing facilities are in a constant state of flux. Processing tools are being taken out of production for maintenance events and are being returned to production after their completion. New products and technologies are being introduced and old ones retired. Work in process (WIP) loadings and distributions are changing due to customer demand. Wafers take various paths through processing tools in the factory. Automated control systems are in various states of initialisation or control of production. Each of these scenarios and more has its own unique needs for metrology data and those needs change over time.

Adaptive Sampling for Tool Monitoring
An example of the need to adapt to the changing state of the factory that is present everywhere is the need to monitor the health of process tools. Every tool in every factory goes through some type of maintenance cycle - moving in and out of production availability. Likewise, every factory develops a methodology for monitoring that tool (usually through a combination of qualification events and product metrology). A typical toolbased sampling plan would be to request that a fixed percentage of lots that are processed through the tool be measured on a defect inspection tool.

Experience tells us, though, that the most likely times for a tool to fail for defects are either immediately after an intrusive maintenance event or immediately preceding a scheduled maintenance event. If this is the case, why should sampling be held constant throughout the maintenance cycle? The critical times where metrology provides the most actionable data is either before or after a maintenance event. At all other times, the likelihood of failure is comparatively small. By using an adaptive sampling system, sampling rates can be adapted to match the amount of information and/or risk associated with either measuring or not measuring a lot from the tool. At the beginning of a maintenance cycle, product sampling rates can be set high to ensure that the maintenance event was done correctly and that the product is on target. During the mid portion of the cycle, sampling rates can be held low. Finally, towards the historical end of the maintenance cycle, sampling rates could again be increased.

Predictive maintenance through adaptive sampling and FDC
Another opportunity for adaptive sampling related to tool monitoring is in coupling sampling with fault detection systems to perform not just preventive, but predictive maintenance on tools. Currently, most maintenance events are scheduled in a conservative fashion and tools taken down for scheduled maintenance are still able to produce quality product. This is done because the ability to accurately predict when the tool will no longer be able to produce quality product is hampered by a lack of data.

Adaptive sampling of metrology by itself is not capable of providing the information needed for this. Once a lot has been sampled, measured and seen to be bad the tool should have already been removed from production. Similarly the implementation of multivariate fault detection on a tool may not by itself be sufficient to provide a good prediction of when maintenance will be required. The sensors available on the tool may not be able to accurately reflect everything that happens on a wafer.

Combining the capabilities of adaptive sampling and multivariate fault detection does show promise in the ability to predict when a tools performance has degraded to the point where it is negatively impacting product. This, in turn, allows the factory to move from a conservative process of scheduled preventative maintenance to the more efficient process of predictive maintenance.

Product-based adaptive sampling
Perhaps equally common to the need to monitor tool status is the need to respond to the introduction of new technologies or products into the factory. Only a few years ago, it was possible for factories to run with perhaps one or two concurrent technologies and only a few generalised products on each technology. Today, IC manufacturers in all sectors must provide a wide array of products to respond to customer requirements - power options, cache sizes, single vs. multi-core processors, etc.

In addition to the need for quick introduction of new technologies and products, customers expect fast execution to their demands for existing products. As an example, in the second quarter of 2004, AMD planned to produce a certain number of units of a particular product. By mid-quarter, customer demand had shown that there was a stronger market demand. Through flexible manufacturing and scheduling, AMD was able to increase the availability of the high-demand product by 30% by the end of the quarter. The need for this type of production flexibility is becoming the rule, not the exception.

Responding to customer demands for new technologies or product distributions creates another need for adaptive sampling methodologies. As new products are introduced, sampling rates must be comparatively high (up to 100% of lots in some cases) to ensure that enough data is collected to accurately target and manufacture the product when it is run in volume. However, maintaining this high level of sampling when the product is in full production is often infeasible due to limited metrology capacity. Likewise, it is also often unnecessary as the volume of material being processed can provide an adequate stream of data for monitoring and control.

The solution to this problem is to adapt the metrology strategy based on the life-cycle of the products and technologies. As they are introduced, the sampling rates are held high; as the products increase in volume, the rates are decreased; and as the product volume begins to decline at the end of their life, it may be appropriate to increase the sampling rates again to ensure an adequate stream of data with the lower volume.

This type of adaptation to product life cycles has been done before in a manual manner, but was rife with inefficiencies. Frequently adaptation occurred only when a new products high sampling rate was causing capacity issues (with resulting cycle-time penalties). Additionally, sampling rates for products may need to be different for different portions of the production line. A product that is in declining volume at lot start may still be the dominant product in the back-end-of-line. An adaptive sampling system coupled to the WIP management system can take all of these factors into account and implement appropriate sampling strategies.

Adaptive sampling for control systems
A relatively new area where the need for adaptive sampling is becoming evident is to provide data to closed-loop control systems. The linkage between closed-loop control systems and metrology data is evident. In order to calculate accurate process adjustments, the control systems must maintain a good estimate of the current state of the process. That state is usually derived from metrology data. As such, sampling systems need to be able to comprehend and adapt to the needs of the control systems.

The most basic scenario of adaptation to control state is the initialisation of a new controller in the factory. When a controller is initialised, the uncertainty in its estimate of the process state is typically quite high. As such, the sampling system needs to ensure that the percentage of material processed through that controller that receives metrology data is quite high until a good state estimate is achieved. Without this ability, the controller may not be able to adequately maintain the process on target.

Another way in which sampling must adapt is in helping the control system to respond to a disturbance of some sort. If a shift is observed in the control state, temporarily increasing the sampling rate associated with that controller can dramatically reduce the impact of an exception. For example, when a shift is detected, a typical control algorithm can easily respond to the shift after having received three metrology updates. If the sampling plan is static, a larger number of lots could be processed before those three metrology updates are implemented (e.g. 10% sampling - 30 lots for three metrology updates).

With an adaptive sampling system, the response to a shift can be much faster. Immediately after the control system identifies the first point of a possible shift, the sampling system begins to identify lots than can provide more information to quantify the disturbance. Typically, the next two lots can be sampled and their information provided to the controller. In such a case, a control system coupled with an adaptive sampling system can respond to a disturbance within three lots plus the lag between the process and metrology operations.