Technical Viewpoint

Trace data: yesterday's news or key to vastly superior equipment performance?

Jon Goldman

Recently, I visited a facility at which my company, Jon Goldman Associates, was installing demonstration statistical process control software for a cluster tool. We had previously installed data collection software and a trace-data viewer for the same equipment. When I asked the customers how they liked the software's trace-data capability, they replied that it is of little benefit because it is too difficult and time-consuming to pore through reams of trace data looking for potential problems that may not even exist. I can readily sympathize with anyone who feels overwhelmed by the unremitting flood of new information lobbying for a slice of one's attention. But it was a little unnerving to be considered part of the problem—a purveyor of junk mail in the guise of useful process information!

We live in an age in which information is more readily available than at any other time in human history. I recently read a little book by David Shenk called Data Smog: Surviving the Information Glut (New York: HarperEdge, 1997). Shenk writes that we are experiencing an information glut that contributes to all manner of social ills, such as a disinterested electorate, deteriorating standards of public education, and even increasing numbers of people with attention deficit disorder. The book is a good read, even though its thesis is not an entirely new idea. He suggests steps that each of us can take to address the problem of information glut, such as turning off our cell phones in movie theaters and restaurants. But he has nothing to say about the semiconductor industry and whether trace-data generation constitutes our own local form of data smog. I don't think it does, but the effect of living in an information-rich environment is that we have become unwilling to do anything that is laborious or time-consuming, like digging through trace data looking for subtle variations that adversely affect semiconductor processes. If gaining knowledge requires expending time, we often choose to remain ignorant. We want easy explanations—a microwave oven approach—when what we really need is to become gourmet cooks.

Telling a Heck of a Story with Trace Data

Most thin-film engineers can bore you to tears with accounts of process subtleties that they believe have major effects on deposition processes, such as little pressure waves when gases are turned on or wafer-to-wafer and chamber-to-chamber temperature variations. Some of that knowledge is folklore, some of it science. But whether it is folklore or science, that knowledge was gained the hard way: through laborious manual data captures or tearing apart machines and finding nothing. Today, we have better tools than ever for detailed diagnosis of equipment performance, but the tools do not operate by themselves. Invariably, they require the engagement of that greatest tool of all—the human brain—to make them useful. As process equipment grows more complex, understanding how to make it perform at higher levels of productivity and control requires detailed study and analysis by human beings. Sorry, there's no short cut.

Consider one of the few industries with a higher level of work in progress than the semiconductor industry: air transportation. Every time there's an accident, the National Transportation Safety Board goes through monumental gyrations to retrieve the voice and flight data recorders (FDRs). But what do those instruments provide? Trace data. Airplanes and thin-film process equipment are similar in some ways. They're both highly complex systems, computer controlled, involve high-temperature chemical reactions, utilize precision-machined components, and occasionally fail. While people's lives are (usually) not at stake when thin-film processing equipment fails, the principle is the same: trace data tell a heck of a story if you take the trouble to read them. (Incidentally, I've learned that FDR data are often used for routine aircraft maintenance and troubleshooting.)

Since I acquired most of my hands-on experience lying under diffusion furnaces, I'd like to share a couple of "toaster" anecdotes to illustrate my point. Figure 1 tells an interesting story. It shows a trace taken from a diffusion furnace during a power outage caused by a tornado. It reveals the exact instant when the main power to the furnace dropped out and the temperature began falling. The trace goes on to show that power was later restored momentarily, only to fail a second time. (Fortunately, the process control and data-capture computers had battery backup, allowing the sorry episode to be documented in its entirety.)

 

Figure 1: Diffusion temperature trace showing effect of power outage during a tornado, followed by temporary power restoration.

I think that one of the unwelcome side effects of Moore's law is that we have become obsessed with "new." Yes, the industry does move toward ever-smaller design rules, larger wafers, and new processes, but I don't believe we have wrung anywhere near the last ounce of productivity-enhancing information out of our trace data. In many instances, we never even try to collect them, let alone look at them!

Digging for Process Variations

Before we rush headlong into implementing techniques such as neural networks, multivariate analysis, and "advanced process control" in our fabs, it would be wise to once again review the pioneering work of Walter Shewhart and W. Edwards Deming. Their techniques are well understood and have been proven effective time and again across a wide variety of industries. For example, Figure 2 shows the average temperature recorded across many runs during a 1270°C soak for five different furnaces. These averages were derived by manipulating data extracted from the traces. There is nothing esoteric about the techniques used to obtain this information. All it required was a little bit of busywork, yet it reveals a significant disparity in operating temperatures between furnaces. This is profound information; these types of differences, I suspect, are quite pervasive among many types of process equipment, but we have to look for them in order to find them. Such disparities should make most product engineers quake in their boots.

Figure 2: Mean spike TC temperatures for five different diffusion furnaces running the same process. Note the ~4°C furnace-to-furnace spread in each zone. Each bar represents an average of at least 30 runs.

Mother Nature loves normal distributions. Ask every male in your office his height and plot the data on a histogram. Unless you work at a basketball summer camp, the data should be reasonably represented by a normal distribution. And Mother Nature did not ignore wafer fabs. Figure 3 shows average flow data from a single mass-flow controller (MFC) plotted as a histogram. The averages for individual runs were again extracted from trace data. The data appear to reflect a normal distribution, meaning that in this instance run-to-run differences are caused by random events. Doesn't this type of information make you want to run out and buy a $10,000 data-capture system to attach to your $4 million cluster tool? You might learn that the silane MFCs in different chambers exhibit quite different wafer-to-wafer variations.

Figure 3: Distribution of average flow rates for 180 runs. Both the experimental data and the normal distribution are shown.

Process equipment is composed of different components, many of which provide measurable output signals that, when recorded, contain useful information. In many cases, the performance of those components is not expected to change over time. For example, Figure 4 shows gas-flow traces exhibiting large overshoots followed by rather lengthy settling times before the flow stabilizes. This is clearly aberrant behavior, and it doesn't require sophisticated analytical techniques to detect or rectify it—just a willingness to dig through data traces.

Figure 4: Abnormal flow data traces. Note serious overshoots and slow flow-settling times.

Figure 5 shows traces from a particle monitor and a gas valve. It seems clear that the valve opening, which connects the process tube to a vacuum source, triggered the particle event—a simple explanation that jumps out of a casual perusal of the trace data. Therefore, before proceeding to fancy data analysis techniques, try gleaning the story within the trace-data record. Doing so takes time, and the effort will sorely test your computer literacy skills, but you will find it interesting and illuminating.

Figure 5: Integrated furnace trace data and particle data. Opening of half-gate valve has caused a particle event.

Many processes routinely change over time. Pads on chemical-mechanical polishing tools wear out, vacuum systems deteriorate as they accumulate reaction products, and the thermal properties of furnace elements change as they age. Accounting for that kind of variability can be tricky, but the key is to understand that even variation can be consistent. Even subtle run-to-run systematic variation can be revealed to those willing to spend a portion of their over-committed time to studying the record of the traces. The watchword is "change." As long as things happen as they have happened in the past, it is safe to assume that all is right with the world. Your challenge is to sort out the truly anomalous wafer-killing behavior from the gray haze of normal process variation.

Trace data analysis is a little like archaeology. Just as the record of an entire civilization may be painstakingly revealed a layer at a time, the record of a process can be revealed one trace at a time. (The semiconductor industry could use a few of its own Heinrich Schleimanns.) But don't run out and buy your green eyeshades just yet. It is possible to automate the process of trace-data analysis (as I have done with some of the data presented above), although it does require that you are familiar enough with your processes and equipment to program an automated system that will find what you're looking for. And how do you develop that familiarity? By spending many hours forming a close working relationship with your trace-data viewer.

Jon Goldman is the founder and president of Jon Goldman Associates (Orange, CA), which has provided data-mining software solutions to the semiconductor industry since 1987. The company's data analysis tools are used in the industry by end-users and equipment manufacturers to reduce downtime and wafer scrap by implementing and improving statistical process control. Goldman joined Motorola in 1973, working on low-pressure chemical vapor deposition (LPCVD) process development, then in its infancy. Subsequently, he was technical director at Thermco Systems. He holds several patents in the fields of silicon nitride LPCVD, low-temperature oxidation, and borophosphosilicate processes. He received a PhD in materials science from MIT. (Goldman can be reached at 714/283-5889 orjon@jga-inc.com.)

MicroHome | Search | Current Issue | MicroArchives
Buyers Guide | Media Kit

Questions/comments about MICRO Magazine? E-mail us at cheynman@gmail.com.

© 2007 Tom Cheyney
All rights reserved.