Process leaps without
new hardware
The benefits that process control and integrated sensor data can bring to semiconductor device processing have been widely publicised. Difficulties in connectivity, however, have prevented wide-spread implementation of these tools. Researchers from US process control supplier MKS Instruments and Swedens data analysis specialist Umetrics describe how new software can eliminate these difficulties with a brand-blind, open architecture. The result is scrap reduction, higher yields and better process understanding. Further, the solution enables the current tool set to meet next generation process specifications without the traditional equipment development cycle and associated costs.
Historically, semiconductor device manufacturers have managed the transition to tighter process specifications by depending on process tool manufacturers to design better and faster process/hardware configurations. As device geometries shrink to the nanometre scale, however, the increasing complexity of manufacturing processes has changed the landscape that must be negotiated to meet and maintain process specifications. Stand-alone control of process tools (based on equipment state data) will not maintain viable yields at 65nm and 45nm. Advanced device processing requires tool-level control based on the combination of advanced equipment (AEC) [1, 2] and sensor-based process control [3, 4]. Furthermore, tool-level control alone cannot meet all of the control needs of advanced device fabrication. System-wide implementation of advanced process control (APC) that integrates AEC to e-diagnostics, fault detection and classification (FDC), and predictive mathematical models of the device manufacturing process, will be required.
Economic as well as technological drivers exist for the move to AEC/APC. The cost of purchasing (and developing) stand-alone process tools to meet advanced production specs is expected to be staggering. A cost-effective alternative to the purchase of a new generation of process equipment exists through the combined use of add-on sensors with AEC/APC in existing, legacy equipment. Sensor-based AEC/APC can drive these tools to the tighter specifications needed for nanometre-scale device fabrication. Additional cost benefits can be realised from reductions in scrap/rework (especially in 300mm wafer processing) and in lower test wafer use, since these can be reduced or even eliminated in systems using wafer- and/or process-state data for process control. Sensor-based AEC/APC can also lessen preventive maintenance downtimes and times to process qualifications, while increasing process capabilities, device performance, yields and fab throughput.
It seems surprising that AEC/APC is not already widely implemented, given such strong driving forces. However, significant obstacles to AEC/APC exist in most fab environments. SECS communications links to process tools are typically unique by OEM, slow and have narrow bandwidths unsuited to data from more sophisticated process sensors. Different OEM and add-on sensors provide data on tool and process states at very different bandwidths and at different sampling rates. The synchronisation of data capture and data transmission at the system level thus represents a significant hurdle for AEC/APC implementation. Further, even should these obstacles be overcome, meaningful application of this potential flood of data in process control is a formidable task. Simpler approaches, such as univariate statistical process control (USPC), are well established, but have limitations [5]. USPC is effective in the control of a single response parameter but advanced device fabrication requires control of multiple response variables simultaneously. Both response and independent variables typically have complex interrelationships that USPC can neither evaluate nor control.
Connectivity
AEC/APC implementation in legacy equipment requires retrofitting automated wafer and process state sensors to existing systems. A robust interface must interconnect and synchronise the data from these sensors with that from the process tool controllers and then further integrate the combined data with in situ and ex situ metrology data. The interconnection framework must interface with multiple, complex sensors from different manufacturers and synchronise the sensor with the tool recipe. It must also reduce the sensor data into a smaller array manageable by upstream applications and handle large data sets in an open architecture database accessible by other applications. The interconnection framework must be able to intercept SECS messages from the process tools, add sensor and metrology data, and make this information package available to the appropriate factory host and APC and FDC tools. Ideally, the system should employ web-based XML coding for high-speed communications with newer applications. Finally, it should be able to purge or archive old data sets.
MKS Instruments has developed a networked framework that, when combined with FDC and multivariate process control software satisfies these requirements [6]. This automated process control and e-diagnostic framework, part of MKS TOOLweb suite of products (Figure 1), collects process data from the tool interface port and then consolidates and synchronises these data with data from OEM and add-on sensors attached to the tool. The framework makes these integrated data available over a high bandwidth network and stores them in an open standard database. APC and fault detection applications, either univariate or multivariate - another part of the suite - operate on these data in real time. Additional server-based applications can also access this database, allowing an end-user to customise the system.
Key components of the TOOLweb system are an integrated server, gateway and data monitor (SenseLink) and a tool data interface and multiplexer (Blue Box).
The TOOLweb solution can access data from all types and manufacturers of sensors, instruments, and subsystems, using common communications, such as analogue, digital and DeviceNet. When necessary, the Blue Box (Figure 2) can perform intelligent data reduction on data-rich streams from sensors like residual gas (RGAs) and Fourier transform infrared (FTIR) analysers. The framework synchronises all of the incoming data with equipment-specific SECS communications from the tool or metrology equipment and stores the data in a well-maintained, secure database. The user can add or delete sensors or other data streams from the framework at any time. Sensor data can be merged into the SECS stream for upstream analysis. Data-hungry applications can access the database directly using Ethernet. An open architecture allows rapid configuration of non-standard interfaces. As a result, connectivity time is minimised and the type of tool or sensor being integrated no longer makes a distinct difference in the ability to use its data.
Users access the system to reconfigure the connections or for data analysis through a web connection. Connectivity is flexible, so several analysis tools can access data from the same process tool. Alternatively, data from a single metrology tool can be synchronised with data from several process tools as appropriate.
With the data securely synchronised and stored, fabs now have the ability to utilise many different analysis tools. Process modelling tools, run-to-run controllers, FDC tools, maintenance schedulers and SPC packages now have a single point to easily access a complete picture of the tool.
Multivariate FDC
This framework provides the hardware interconnectivity, the data synchronicity and the transmission capabilities needed to implement advanced process control in legacy and new equipment. A final component required for APC is a system capable of using this much-expanded, multivariate data stream in effective and robust process control. In this solution, Umetrics multivariate analysis (MVA) engine, called SIMCA-P+, provides the final component necessary to enable the use of legacy tools in new demanding processes.
Traditional univariate SPC, while effective for process monitoring and after-the-fact fault detection, does not fulfil this need since it does not address the inter-relationships between variables [5]. Figure 3 shows a typical situation in which SPC fails to detect a bad wafer. The univariate SPC charts show all points to be in control while the MVA clearly detects the point outside of the in-control cluster. Various approaches to multivariate analysis make use of all sensor data and their inter-relationships for building models of the process that can be used by on-line FDC engines [7]. While not a substitute for SPC in monitoring limits and alarms, MVA coupled with on-line FDC considers thousands of variables and their inter-relationships in real time. Using MVA, the system can build a process model offline using good wafer data and then export it to an FDC server. This engine generates reliable models from a few observations with many variables, many observations with many variables and anywhere in between. The on-line fault detection engine can use the model to compute Hotellings T2 and DModX statistics in real time and alarm when critical limits are reached. MVA models are relatively insensitive to missing or noisy data from individual sensors - even when an individual sensor breaks down, the APC continues. As well, MVA readily accommodates strongly correlated (collinear) variables. Typical applications for the multivariate offline analysis include fingerprinting, chamber matching and process optimisation.
Principal component analysis (PCA) [8, 9] is a technique that generates a model based on variation in the independent variables and has the potential to significantly reduce dimensionality by summarising the data. Partial least square (PLS) analysis extends PCA models to include response variables such as film properties, yields, etc. [7, 10]. PCA and PLS algorithms enable the creation of models of both the evolution of the wafer through the recipe and of the entire wafer. Metrology data and other downstream results can also be correlated with process signals. This solution creates a flexible modelling architecture and a simple interface that allows for models to be built in minutes. By using ‘local centring it creates reliable R2R (run-to-run) adaptation to setpoints. The Blue Box accesses SIMCA-P+ automatically and feeds data into the interface for model building. A trigger to create a new model might be a value of a variable (e.g. etch rate) or number of wafers from a previous model. The system builds new models silently (and automatically) in the background.
Models are automatically loaded into the TOOLweb suite Oracle database. SIMCA-QP+ handles step-by-step and wafer-by-wafer FDC by detecting process excursions using a combination of multivariate signals. Alarm and out-of-control limits rely on Hotellings T2 statistic to gauge combined trend magnitudes and the DModX statistic to analyse for correlation structure changes. The multivariate data table can be converted into simple plots that readily show out-of-control conditions. Figure 4a shows an example of a t1/t2 score plot that provides an overview of the data and sharply contrasts outliers and data clusters.
Plots of Hotellings T2 statistic (Figure 4b) show that wafers outside the limit are outliers whose measured parameters do not fit the accepted model for “good” wafers. DModX plots (Figure 4c) identify wafer data that are too far away from the model plane and exceeding a critical limit, indicating an abnormal process that has not been seen before.
Individual contribution plots reveal the root causes of the bad wafer. Contribution plots summarise the wafer evolution for each process parameter or variable identifier (VID) and possible responsible fault sources, presenting only those most likely to be causing the excursion.
