Issue

Fab planning: trading in your spreadsheet

10/01/1998

Fab planning: Trading in your spreadsheet

Don Baylis, Manugistics Inc., Los Altos, California

Every semiconductor production manager has the responsibility of maximizing the company`s return on assets while meeting customer expectations. The classic tool for this effort -an elaborate spreadsheet -is woefully inadequate to handle the dynamics of today`s semiconductor manufacturing. Increasingly, those responsible are turning to advanced planning and scheduling software tools. Here is a guide on how to implement these tools, what to expect, and the cautions.

IC-design engineers rely on layout and verification software. Wafer-process development engineers engage in extensive simulation before putting a wafer into a fab, but many operations groups still use spreadsheets for production planning in fabs and assembly lines. Further, scheduling shift activity is left to manual guesswork or dispatch via local rules of a manufacturing execution system (MES). At the top of the organization, enterprise planners use, if anything, expensive custom legacy applications.

Increasingly, the difficult challenges that semiconductor manufacturers face dictate a need for change. Customers demand rapid response and on-time performance; product cycles grow shorter; and fab costs make return-on-investment vitally important. It takes a new generation of planning and scheduling tools to balance these competing demands in the production environments of today and tomorrow.

Scope of planning and scheduling

A typical semiconductor enterprise uses three major levels of planning and scheduling that address different levels of detail (see figure). Supply-chain planning affects the whole enterprise, assigning demands to individual lines and projecting capacity shortfalls over future weeks. Factory planning determines work for an individual production line over a period of days to weeks, generating a starts-and-outs plan. Production-line scheduling, which may require updating several times a day due to equipment and lot dynamics, establishes day-to-day activity for the production floor.

Click here to enlarge image

The three major levels of planning and scheduling for semiconductor manufacturing provide different levels of application detail.

Semiconductor production is, in itself, a limited supply chain. Seldom does a single plant perform all the required sequences for wafer fabrication, probe, assembly, and test. More commonly, manufacturing is done on multiple continents, often by multiple companies, including wafer foundries, assembly contractors, and test houses. Additionally, a given semiconductor manufacturer may have multiple plants for each major segment of production.

For supply-chain planning, this combination of production challenges poses a series of difficult questions: Which fab should start the wafers? Which test floor has the capacity? Does an assembly contractor have the capacity, material, and competency? Does the virtual enterprise have the capacity to accept this order profitably? After satisfying bookings and safety-stock requirements, how much of the forecast can be produced? Which part of the forecast will maximize cash flow or profit?

Once enterprise planning determines the production requirements for each wafer fab or assembly line, individual lot-start schedules need to be generated. This is the realm of factory planning. Key questions include: How many lots of which product need to be started on each given day or hour to meet demand? What is the cycle time for each product or lot given the mix of products in the line? With the product mix being produced on a given line, how many hours will be used changing machine setups? Which lots should be allocated to existing orders, and when should they be started?

Scheduling requirements eventually come down to specific production lines and individual equipment, the realm of production-line scheduling. Here, the questions include: Which lot moves next? When should this machine get preventive maintenance (PM)? Which lots are batched together for this operation? What is the impact of expediting this lot? Will other lots become delinquent? Will reticles, operators, etc., be available for lots in the queue?

The role of APS

Today, all the operational details outlined above are better handled by advanced planning and scheduling (APS) software applications. APS combines linear programming, heuristics, and discrete-event simulation to generate plans and schedules.

Linear programming (LP) defines planning as a series of equations based on models of a given production process. Constraints -such as machine capacity, material availability, available inventory, and product demand -define the equations. Linear programming engines then very quickly solve these problems involving millions of variables. The objective and strength of linear programming is goal-seeking: A linear programming solution is optimized to meet goals such as minimum delinquency or maximum cash flow.

Heuristics mimic human pattern recognition and guesswork by following user-defined rules. Most heuristic programs load the entire production model into memory and can quickly determine the rough-cut feasibility of an option. This tool can test production options such as "Can I fill this order?" or "Can I meet demand if a plant shuts down for a week?"

Discrete-event simulation (DES) projects factory activities over time using defined inputs to simulate events in the factory. Each lot has a projected completion date and individual cycle time. Any delinquent lots are identified. Future bottlenecks due to equipment or operator shortages can be identified in place and time. Only DES can project temporary and short-term interaction of lots in a fab. This allows just-in-time scheduling while maximizing the throughput of both equipment and the factory.

While most commercial APS suppliers have expertise associated with one of the above technologies, APS applications commonly use them in combination. Linear programming or heuristics ("solver technologies") are often combined with the detailed grounding of DES to provide extremely powerful results.

APS solutions enable modeling of a semiconductor manufacturing production process at a level not possible with conventional methods. They allow forecasts of future behavior rather than depending on historical data. APS systems bring the benefits of increased manufacturing efficiencies and improved customer service. They include more planning factors and generate results more closely aligned with factory performance. These systems provide faster plan generation while retaining human control of the final plan. They enable the synchronization of enterprise goals, factory plans, and production-line schedules. Even operations managers at fabless semiconductor companies are finding that APS helps them meet customer requirements.

Generation of a production plan and start schedule is only the first of many steps in meeting production commitments. A plan must be "capacity feasible" for the given mix, but frequent adjustments to changing conditions of the production floor are necessary to execute a plan. Scheduling systems can download WIP location and equipment status from the MES, and generate a new short-interval schedule. Some planning and scheduling systems can even share the production model, and make the schedule adjustments needed to keep bottleneck equipment busy and get the right lots out at the right time.

Spreadsheets vs. APS

The use of spreadsheets for planning semiconductor-manufacturing lines is limited by four major concerns. First, spreadsheets use static calculations and therefore do not consider inputs that vary over time. Second, they base throughput and cycle time on historical approximations. Third, they do not consider the impact of instantaneous product mix, both historical and future. Finally, they are restricted by slow calculation speed.

Static calculations, which define input values once and use them throughout a set of calculations, lead to major inaccuracies. Consider this example. An ion implanter has processed 100,000 wafers in the previous 1000 production hours, an average throughput of 100 wafers/hr. During operation, the throughput is 150 wafers/hr, but with PM, setup time, machine failures, lack of operators, and lack of WIP, the overall throughput is 100. Static calculations can use 100, but further investigation might reveal a range of 80-122 wafers/hr during the 100-hr period. Since the planning group will be criticized if too many lots are late and the product mix is changing, caution suggests using 90 wafers/hr. However, application of this philosophy across the line can reduce planned manufacturing capacity by 10-20%.

The above example is based on historical data. Execution of such a plan may reveal the product mix creating moving bottlenecks between two workstations, making the estimate not conservative enough and leading to customer dates being missed for 10% of the wafer lots. To a spreadsheet, cycle times are inputs, not a mix-sensitive output.

The example shows the concept of effective capacity -the capacity of a factory used for planning. If it exceeds real capacity, delivery dates are not met and WIP grows. If effective capacity is less than real capacity, orders are refused needlessly. APS can narrow any differences. This author has heard users of APS planning state, "We could have taken a million dollars a month in additional orders if we had installed APS planning earlier and understood our capacity better."

Most of us have experiences with simple spreadsheets that calculate in seconds. However, a spreadsheet with enough complexity to model a wafer fab can take several hours to calculate. By comparison, APS tools used on the same computer as a spreadsheet can generate multiple what-if scenarios that can be used to determine realistic production targets. Here a convinced user stated, "Our former planning cycle was four weeks, consuming 90% of the planner`s time using static files that couldn`t comprehend the interaction of various activities and, by then, producing outdated data. Now ... one and a half hours for a plan with greater accuracy ... [that] ensur[es] agility."

Selecting a supplier

Several suppliers offer APS software solutions for semiconductor manufacturers. These range from large companies with extensive experience in other industries, to enterprise resource planning (ERP) providers, to firms that specialize in semiconductor manufacturing.

When selecting a supplier, it is important to remember that the interaction of semiconductor manufacturing`s special set of requirements (Table 1) makes it unique and demanding. Examine the record of prospective providers within the semiconductor industry. Broad-based experience in other industries may provide good integration skills, but it does not assure a provider who understands your terminology.

Click here to enlarge image

Some providers offer a monolithic application to solve supply-chain allocations and detailed production plans simultaneously for all plants involved. Other providers offer distributed solution sets using multiple technologies to best meet the differing needs of the planning and operation groups (Table 2). Semiconductor manufacturers with strong central planning staffs often find the single solution attractive. Others prefer a less-detailed demand on the central planning group, distributing it to individual plants. Selection of a supplier with an approach similar to your planning business process will ease implementation and make a happier partnership.

Click here to enlarge image

A total production model of multiple fab, assembly, and test areas is of necessity large and complex. A single model capable of both supply-chain and production-line planning will require a computer of correspondingly large capacity, and model maintenance will require the cooperation of staffs from all plant sites. On the other hand, users can maintain and operate a distributed model on a network of normal-sized computers, with local models maintained at each site. The aggregated enterprise model can then be refreshed automatically from the network of local models.

Just as a fab area uses different tools to examine wafers in a processing line, the use of multiple tools and technologies can be the most efficient approach when planning a semiconductor enterprise. It is important to match the technology of the solution to the focus and scale of the problem; a process engineer would not use a SEM to do whole wafer inspection after each diffusion step (Table 3). Therefore, a planner should not use the detailed model needed for shift scheduling to do an 18-month enterprise projection.

Click here to enlarge image

Implementing APS

Implementation of APS requires getting the upper hand on data management. Any program to model a fab, whether using LP, heuristics, simulation, or spreadsheet, includes a significant amount of data. Much of this data is already available within computer systems, computer integrated manufacturing (CIM) systems, sales-order management systems, and existing spreadsheets. An APS provider should be able to help gather and transfer this data, but the user must still collect and verify it.

APS requires data about production resources - the definition of machines, people, tooling, transport stockers, etc.; operating rules - equipment loading, lot dispatching at workstations, PM, work calendars, etc.; product demand - data extracted from order entry and demand forecast systems; product definition - product names, direct material, starting material, routings, rework and yield losses, setup and processing times, etc.; and model controls - run time span and various internal options to tool scheduling.

Also, significant amounts of data may be used to make up APS models. In the supplier selection process, visits to existing users may reveal very detailed models, leading to the frightening conclusion that it will require a lengthy implementation before recouping the investment.

However, quick APS system startups are possible. A phased implementation can provide benefits before completing a more detailed model. A model capturing the major bottlenecks and process flows can provide plans and schedules of greater accuracy than those generated by spreadsheets. For instance, if in-process inventory is lined up at wafer steppers and high-current implanters, do not start by modeling the diffusion area. A smaller implementation team can continue the process, refining and enhancing the manufacturing model for added improvements while enjoying partial benefits. This strategy will maximize the value of the implementation investment.

Expected benefits

Goals for implementing an APS system installation are:

Plan and commit to demands constrained by material and resource capacities (e.g., neither exceeding a stepper`s capacity nor failing to improve throughput if setup requirements are reduced).

Release lots intelligently into a fab area based on bottlenecks and minimizing queue lengths (e.g., process lots through photolithography so they go to an empty diffusion area, not an overloaded implanter).

Predict bottlenecks based on current WIP, product mix, and equipment status (e.g., know beforehand that new products will reduce throughput of implant by 20%).

Schedule WIP movements based on WIP location, customer due dates, and production availability (e.g., move lots that will satisfy customer orders, not production numbers).

Schedule PM to minimize effects on production schedules (e.g., perform PM early rather than just as 10 lots arrive on schedule).

Balance the conflicting demands of overall equipment effectiveness (OEE), quick cycle times, and customer satisfaction, while raising the standards for each (e.g., low WIP levels are key to quick cycle times, but an enemy of OEE).

Provide an integrated solution for WIP tracking and material purchasing systems.

Caveats

Successful application of APS requires management buy-in to support possible corporate-culture changes. Since it uses a wider range of inputs compared to the simpler system, APS may produce unexpected output, which users must be prepared to accept.

The need for human input is not eliminated. APS tools free human planners from routine adjustments to concentrate on extraordinary cases. While the tools identify these cases, only a person can balance the inputs to judge the most complex decisions.

A paradox of scheduling a complex environment is that local optima do not lead to a global optimum. For instance, if an ion implanter limits production, the temptation is to maintain a large WIP in front of the machine and minimize setup times by processing all lots for one setup before switching. This policy will alternately flood and starve every system later in the flow, increase factory WIP and cycle time, and ultimately lead to late deliveries. Proper optimization of factory performance involves balancing WIP, minimizing delinquencies, and minimizing cycle time to avoid quality accidents and processing obsolete material. This may reduce OEE at the implanter, but improve customer satisfaction and factory return on assets.

Compared to spreadsheets, APS tools and their integration are expensive. However, if fab output is increased a single percent the first year and another percent the second year, the investment has a multimillion dollar return; for a $2 billion fab, a single percentage point equals $20 million. Consider that APS systems have been used to create successful plans with 10-20% higher "outs" than legacy systems.

Conclusion

The growing complexity of semiconductor manufacturing, increasing pace of change, and intensifying competitive pressures require a major improvement in production planning and scheduling. APS tools are now available to generate capacity-constrained plans and schedules, integrated with MES. Early adopters of these technologies have demonstrated increased production, improved customer satisfaction, and higher manufacturing efficiencies.n