As ICs grow in size, finding a failure's cause has become more and more difficult. Data Mining and Diagnosing IC Fails by Leendert M. Huisman describes how statistical methods can help uncover the source of a problem using data provided by IC test.

This book discusses failures during IC manufacture, rather than at the board or assembly level or in the field. Diagnosis here means diagnosis in the broad sense: Although the book does discuss diagnosis with respect to pinpointing failure locations within a die, most of the book concerns the detection of patterns and indicators of process problems that cause many die on many wafers to fail. So, even though the title does not indicate it, much of this book is devoted to yield enhancement, which I think makes it far more interesting than if it were completely devoted to traditional diagnosis.

Huisman begins with a brief review of statistical distributions and likelihood. He then moves to yield statistics, with emphasis on the variance of wafer yield and the use of differences in the probabilities of first failure (that is, the first test step that fails) to indicate that wafers or lots have different process histories. With that done, it's possible to consult the manufacturing histories of the lots, attempting to determine what caused the variation.

Chapter 4 discusses the area dependence of yield, models of this area dependence, and the impact of nonuniform defect distribution on yield. This chapter, as do many of the chapters, contains examples with real data. Its emphasis is on identifying a potential problem to enable further investigation, rather than on directly determining the problem's root cause. The mapping of process parameters to defect types is not within this book's scope, which keeps it focused on diagnosis.

Not all defects are equally easy to detect. There are different classes of defects, and some defects exist in structures that are easier to test than others. In particular, it's possible to thoroughly test objects such as memories and scan flip-flops, and discovering defects in them can provide insight into the causes of failures in general. Chapter 5 describes how to estimate failure probabilities of the objects and how to find correlations between them.

Defects on dies can be random or systematic. Systematic defects have become the major cause of failures, and finding them is crucial because they are correctable by process changes. The next two chapters describe how to distinguish random and systematic defects. Chapter 6 shows how failure signatures can help find commonalities between failures, and thus find failure clusters. Chapter 7 shows how to find patterns in the data and explains how these patterns can help identify systematic problems.

With Chapter 8, the book's focus moves to the die—the chapter relates test fallout to test coverage. Chapter 9 gives a brief summary of logic diagnosis, and Chapter 10 describes single location at a time (SLAT), a fault-model-independent diagnosis strategy developed by the author and his colleagues. This method finds patterns for which a defect affects only a single location, and uses these patterns to determine a set of possible locations for the defect. This has the advantage of working for bridges and unmodeled defects as well as stuck-at faults. Chapter 11 concludes with a discussion of data collection requirements.

The book's treatment of diagnosis is quite mathematical, though the author helpfully places many of the details in a set of appendices, which improves the flow of the chapters. I found the notation difficult. Although there is a list of symbols and abbreviations in back of the book, the symbols do not easily map to the concepts, and some are nonstandard. For instance, the book uses c to represent test coverage, instead of the more common f for fault coverage or d for defect coverage.

Though the title of the book mentions data mining, the book contains relatively little discussion of the topic, most of which appears in Chapters 6 and 7. Madge and Daasch et al. have stressed the notion that test is the source of data from which you can make decisions about the quality of ICs. I was expecting more discussion of this topic, given the book's title, but their work receives no mention at all and does not appear in the references. In fact, data mining does not appear in the index, and I can find only one mention of it in the book.

Huisman references other work in the area but does not review those other methods in the text. A casual reader might think there is only one way to do diagnosis. For instance, although the well-known works on escape rates by Williams and Brown and by Seth and Agrawal appear in the references, I find no mention of them in the chapter on yield and coverage.

Finally, although there are examples of data and graphs, I would have liked to see some example applications of the formulas described in the book. Such examples would help someone new to this area apply the techniques in this book. And although many of the techniques are too complicated for text examples, a Web site or CD with sample (fictional) data and worked-through examples would be useful for the learner.

All in all, this book is a valuable contribution to the essential art of finding clues about what goes wrong during IC manufacturing.