GEORGE W HORN, Middlesex industries Inc ([email protected])
During the last 40 years, front end fabrication fabs have evolved parallel with wafer form-factor developments. Yet basic logistics principles in use have not changed. We move wafer lots through the process one by one, via discrete operators or automated vehicles. This ancient logistics needs to change in light of modern tools available to optimize manufacturing. This review looks at some of the problems.
For clarity we adopt definitions. For DISPATCH: as managing WIP to comply with processing resources. Its objective is optimum throughput/utilization. And for AMHS: as managing WIP complying with transport and storage resources. Its objective is optimum cycle time. Then, in the following we analyze AMHS, as the executor of fab logistics. We also define MANUFACTURING CYCLE TIME to consist of 1) tool process time, and 2) wafer inter process transport time. For each lot. (Where transport time includes and may be segmented and interrupted by storage times.) The domain of this inter process activity is then modeled as inter process state space. This state space model is the subject of this review. Considering this model as a dynamic system, submitted to modern analysis, and questioned if it may be chaotic or may be improved via machine learning.
AMHS In the inter process state space
AMHS, the Automated Material Handling System, provides the infrastructure for product flow between process tools and storage tools. It enables execution, and planning, of process sequences. Historically, the primary methodology for this inter process sequencing was discrete substrate moves via discrete vehicles or people. A fundamentally limited method, due to its inability to move substrate lots on demand. Yet adopted by industry consortia for today’s 300 mm wafer format factories. Few exceptions to this formula exist, in the form of conveyor networks, which liberate wafer flow to be at will, independent of vehicles or people transports.
Recursion of substrate flow, combined with the forever increasing number of mask layers, burdens the AMHS with increased complexity and delivery volume. This results in poor flow factors, meaning that more time is spent by each product being moved then being actually processed. Hence comes the urgency for updating AMHS. Which became a major contributor to factory cycle time and throughput. To keep things in control, forecasting wafer lot moves becomes a necessity, and by modeling it we aim to predict arrival times and enhance utilization of the processing assets.
How do AMHS execute product flow? 1) Todays conveyor network type AMHS (with direct equipment Interface) is a simple case. It requires no scheduling of the AMHS system. It provides a route from all tools to all other tools in the form of a network of paths, similar to a NY city street map. Direct carrier placement onto the network with a destination will move the wafer lots autonomously to destination. And at street crossings, the lots are moved FIFO. This simplicity, while still allowing in-flight reassigns of destinations, is achieved via the embedded controls throughout the conveyor tracks. 2) The discrete vehicle type AMHS demands more complexity in logistics. Individual vehicle assignments are made, or planned, as substrates appear at the output of process tools. This demands forecasting in order to maximize vehicle service efficiency and tool utilization. Yet, there is no perfect mathematical solution to this moving vehicle type of distribution system. Thus, various heuristic solutions are in use.
Chaos
We like to model our systems with mathematical formulas. We do this primarily to predict future states of those systems. It is in this forecasting context that the mechanism of Chaos becomes pertinent to dynamic systems. Chaos allows the rapid growth of uncertainty in those mathematical models and thus limits our ability to predict. One condition for chaotic development in dynamic system models is small variation in initial conditions. Where chaos provides us with a description of their rapid growth. Another fundamental condition for the system model is its nonlinearity. So outcomes are not proportional to changes in initial conditions. And lastly, Chaotic system models are deterministic, in the sense that their current state completely determines their future state, i.e. computer models. Do these conditions apply to the formulation of our state space: the inter process wafers lot space, where AMHS models operate?
Click here to read the full article in Semiconductor Digest magazine.
