Effective Implementation of APC

Advanced process control (APC) is a technology that has been widely accepted within the semiconductor industry as a powerful technology tool. This widespread acceptance stems from the numerous benefits achieved through the implementation of an effective APC program, which includes sensor integration, run-to-run process control, and fault detection and classification (FDC).^1-5 The most concrete example of this level of acceptance is the inclusion of a requirement for APC in the International Technology Roadmap for Semiconductors (ITRS).⁶

To meet the roadmap's capability requirements, process corrections eventually will need to be made on a lot-to-lot, wafer-to-wafer, field-to-field and site-to-site basis. These levels of granularity will require APC-enabled process tools, as well as improved access to data coming off the tool and to the recipe inputs required for process adjustments.

A single approach will not be effective for all fabs. Each fab will have unique processes, different levels of existing automation infrastructure, and different requirements for implementation of APC. Additionally, each supplier has various levels of capability to provide. Sometimes, even within the same supplier, different products will have different levels of APC functionality. Compounding these variations is the fact that requirements will change over time.

1. The four levels of APC architecture encompass control in situ (L0), at the tool-level (L1), supervisory level (L2) and fab level (L3).

As such, the most effective solution is an open, flexible system to support APC applications. Multiple levels of implementation are likely to occur. Some capabilities will be implemented at the factory level, others at the module level and at the tool level, and some capabilities will be required at all levels. It is unlikely that APC will be implemented identically across all process technologies. All the systems must be capable of communicating with one another in an open, interactive environment. Seamless integration is the ultimate goal.

Closer cooperation between IC manufacturers and their suppliers also is needed for effective development and deployment of APC technology. Because there is no single right solution, manufacturers and suppliers need to collaborate on how they will implement APC to meet the unique requirements for that fab and that supplier team. This paper discusses two collaborations between TEL and two of its customers, AMD Fab 25 and TI KFAB.

The global goal for APC at AMD is higher product value and lower manufacturing costs. We achieve this through the following: reduction in process variability, rework and scrap, increasing levels of automation (leading to increased tool utilization and higher throughput), and control methodologies that are flexible and comprehensive.

AMD's APC architecture contains four tiers or levels (Fig. 1 ). These levels are interactive and open. In general, they can be defined as the following:

• L0—In situ APC controllers providing process-level stability, repeatability and manufacturability.
• L1—Tool-level APC controllers with an open, interactive interface for wafer-level fault detection/classification and process control functionality. Adjustments to the tool potentially are made on a wafer-to-wafer basis, and include downloaded process targets and control limits, or alternatively specific recipe values.
• L2—Supervisory-level APC controllers that are cascaded and cooperate with one another, and tool-level controllers to provide process stability at the device feature level, such as isolation, transistor and interconnect.
• L3—Fabwide APC control applications that facilitate synchronization of larger segments of the production flow, up to the factory level, such as optimization of device processing based upon electrical parametrics and agent-enhanced scheduling for product mix balancing.

2. New tool-to-factory communications are required to facilitate the level of interaction and cascading control needed for successful implementation.

For successful deployment of APC applications, a minimum level of infrastructure and capability must be present in both the manufacturing tools and factory systems. At the highest levels of the factory, two distinct tiers of APC will be required: one tier that optimizes device processing based upon key electrical parametrics (L3), and another that ties together key processes into a device module and controls them in a coordinated fashion (L2). At the top level are the APC applications responsible for optimization of overall yield, sort and device characteristics. This is achieved through adjusting the targets of the individual process modules.

The level that supervises the creation of a device module (a complete critical layer) typically is referred to as supervisory control. Supervisory controllers are responsible for controlling key device parametrics associated with specific process modules, such as the gate module and shallow trench isolation module. Between these two levels, all information from tools, metrology, electrical test, scheduling and forecasts can be combined to deliver the correct mix of products, at the needed performance specification and on the optimal schedule. The final result is the on-time delivery of the correct die out for each product the fab produces. Around each tool is the need for run-to-run and FDC controllers that comprehend the needs of the supervisory controller, the wafer state of incoming material, the current process state of the tool module(s), and the tool/module's equipment state (L1 and L0).

Current tools collect data and configure sensors through a static method, generally using the same sample rates, sensor settings for all wafers and, often, all modules. With wafer-to-wafer control, the settings for sensors need to be configurable at the wafer level. Sensors need to be programmed through data collection plans that are dynamically selected by using the wafer context. In addition to dynamically setting up sampling plans, data analysis strategies need to be configured by wafer context to perform data filtering, trimming, summarization and analysis. The results of the data collection plan or output from tool-level analysis need to be available to factory-level APC, or to tool-level applications that contain the limits and rules defined by the equipment and process engineers. If wafer results are out of control, the tool module needs to pause within a wafer, after a wafer or after a lot, depending on the severity and confidence of the fault detected.

3. Process data typically are analyzed by multivariate statistical methods, neural networks or graphical visualization techniques.

With integrated metrology, the specification of wafer sampling (example: slots 1, 5, 15) will become part of the recipe from the factory, and must be combined with the sampling needs of the integrated tool (wafer-to wafer) controller. When the controller is running in a stable region and all the material is consistent, the number of wafers needing sampling will be dynamically defined at run-time to a minimum level. The number and location of measurements on the wafer can be adjusted to maximize the sampling of data needed to represent the wafer state. For example, the factory control system will know that, for a given lot, wafers 3, 4 and 5 were processed on a previous tool under less than optimal conditions, and that these wafers will need 100% metrology at the appropriate downstream processing step. The factory, supervisory and tool-level systems must coordinate this activity.

With tool-level APC available in some equipment, factory-level APC collaboration (roles and responsibilities) begins to evolve. By employing standard communication technology such as XML (extensible markup language), the tool can communicate directly with the factory system APC to coordinate changes to recipes at a lot or wafer level. Supporting tool-level APC requires agreements between the factory and tool regarding what will change, limits and ranges of allowable change, and notification of changes. Without collaboration with the tool-level APC monitoring for faults, the factory could change recipe set points lot-to-lot, and the tool APC would detect these as faults. To control wafer-to-wafer variability, the process state needs to be modeled, and adjustments need to be made based on the process stabilization. With wafer-state models evolving, allowing the prediction of physical attributes (such as depth and critical dimension), physical target information is supplied to the controller. Integrated metrology can aid the controller by reducing time between processing and measurements.

New tool-to-factory communications are required to facilitate the level of interaction and cascading control needed for successful implementation (Fig. 2 ). The tool communicates directly with the factory system APC to coordinate changes to recipes, data collection plans or control models at a lot or wafer level. For example, the process controller and the fault detection applications must be coordinated to prevent one from making a false-positive decision based on a true outcome of the other. All levels of APC must have the capability to embed and cascade the business rules logic necessary to react to out-of-control points. These reactions will be different for each fab, and will depend on how they wish to respond to deviations.

TEL started with the goal of building a foundation for tool-level APC. The objective was to enable wafer-to-wafer applications such as process control, fault detection and fault classification. This required an understanding of the data required, data sources, data formats, data rates, and analysis and storage requirements.

The foundation starts with all tool control data. This information is available from internal control loops (L0). Basic tool data is not enough to understand the complete process. Additional sensors are added to record information about the environment or the process. The software includes the capability to synchronize tool events, tool data and external sensor data. The merged raw data then is trimmed and filtered. This allows unique data collection and analysis plans to run for specific products or processes. This step is critical for robust modeling.

The data is summarized or analyzed by various techniques and applications. These typically are multivariate statistical methods, neural networks or graphical visualization techniques. The information from the analysis application is used in various applications, such as statistical process control (SPC), tool management models, model-based process control applications, fault detection models, or APC models using metrology data (Fig. 3 ).

A key part of APC is the ability to provide two-way communications with the factory. Ideally, the tool-level APC system will reduce the data to a level suitable for factory-level applications. Information and models from the factory level also will be plugged in at the tool level (open architecture). To implement this, AMD and TEL are working together to create an XML interface, in addition to the existing HSMS (high-speed SECS message service control) interface (Fig. 4 ).

Etch process equipment provides a unique challenge due to a lack of first-principle process models. This requires extensive measurements (data collection on the order of megabytes per minute, per chamber) of equipment and plasma state, followed by the development of empirical models to correlate equipment performance related to the process state. The etch process requires chambers to be cleaned and consumable parts replaced. This adds a large variable that must be included to make the models robust in high-volume manufacturing. Optimal performance requires the support of multiplexing three or four chambers running at the same time. Chambers must produce predicted process results, regardless of their point in the maintenance cycle. Additional sensors are required to understand and characterize the plasma state in an etch reactor. The sensors TEL is implementing are optical emission spectroscopy (OES) for full spectrum analysis and the VI probe to analyze rf harmonics. These sensors can be used independently or in conjunction with one another for control or process characterization. TEL's APC program incorporates these sensors to create additional process knobs for the etch reactor.

4. A key part of APC is the ability to provide two-way communications with the factory.

TI and TEL collaborated to reduce the number of in-line process monitors used for checking health of the etch rate. TI's current practice for qualifying the etch rate for an oxide etcher is to run one monitor wafer every seven days, and one monitor wafer after a wet clean is performed. If the first monitor wafer fails, the standard practice is to run another one.

The potential new practice is to minimize the qualification procedure by 50%. Successful implementation of this methodology directly translates into benefits including: 1) increased throughput (reduced set-up time, reduced number of quals and qual time, decreased post-maintenance recovery time, and reduced cycle time); 2) cost reductions (conserved monitor wafers, and improved productivity of operators and engineers); and 3) risk management (chamber-to-chamber or tool-to-tool correlation, fab-to-fab standardization, knowledge of chamber condition at real time, process and product understanding, and increased confidence in alarms).

Employing the new methodology, the TEL tool-level APC product captures the real-time wafer data and performs multivariate analysis (MVA) to predict etch rates after the wafer is processed.

Figure 5 shows graphical representations of the observed vs. the predicted etch rates. Figure 6 shows the actual etch rate predictions during the production runs. The predicted values track the actual values fairly closely. Some important points to note: Points 28 and 29 had incorrect measurements and, while the actual measurement of point 30 recovered, the predicted etch rate did not show this fluctuation. It was later determined that the metrology tool gave an incorrect reading for measurement points 28 and 29.

A very small number of measurements presently is used for the training set. The next step in optimizing the model is to tighten the gap between the actual and predicted values using additional data. Other applications currently under evaluation are general process excursions, etch depth detection, and wet clean recovery. These applications are targeted toward risk management, reduced cost, and increased throughput.

There are still areas of APC that require industry cooperation. Notably, more work on standards for the data formats and data interfaces is needed. Additional discussion is needed regarding the n-tier control architecture, and for determining which applications are best run at the tool, and which at a higher level. Key factors in the discussion are data rates; time required to process the information and get back to the tool; the tool's rate of communication; where the data originate and need to go; and where it is best to store both the raw and processed data. Both the process tool and integrated metrology need to be considered.

Recipe management will become increasingly critical with tool-level process control. The recipes today are relatively simple, with the most complex having three parts (sequence recipe, module recipe, and parameter set-point overrides). Recipes with APC are more complicated. They require additional data for integrated metrology (measurement recipe and analysis recipe), FDC algorithms (models and data collection and analysis plans), and process control (algorithms, coefficients and parameter overrides).

Creation and maintenance of models (FDC or control) require information from both the factory and tool. Who creates these models and where are they created/stored? For the tool supplier to provide this service, the factory must provide access to information traditionally considered private. The factory requires detailed understanding of the tool to optimize the parameter interaction for control.

While ITRS has indicated the need for APC, AMD and TI are early proponents. Collaboration is necessary; AMD and TI provide the knowledge of the product, fab interactions and factory APC, while TEL provides the tool knowledge, tool-level APC, and sensor integration. We benefit from reduced cost, increased productivity and higher-value products.