experimental design
Experimental design is the disciplined science of planning experiments so that the data they produce answer the engineer's question with maximum efficiency and minimum bias, and it is the single most important methodology in semiconductor process development because it determines whether the measurements made in a fab can be trusted to reveal the true effects of the process variables. When an engineer changes one process parameter while holding all others fixed, the naive one-factor-at-a-time approach, the data can easily be confounded by unintended drift, by the variability between wafers, and by the interactions that occur when two parameters do not act independently, and the experiment can consume enormous resources while still failing to identify the true drivers of performance. Experimental design replaces this ad hoc approach with a principled plan that randomizes the assignment of treatments to experimental units, replicates the treatments to quantify the noise, and blocks the experimental units into homogeneous groups to remove known sources of variation. The reward is that the resulting data support valid estimates of main effects and interaction effects, valid tests of significance, and predictive models, all from a carefully sized number of runs. This document develops the principles, the standard designs, and the analysis of experiments as they are practiced in a semiconductor fab, where the goal is to maximize the information gained from every expensive wafer and every scarce lot.
**The three principles that underpin all of experimental design are randomization, replication, and blocking, and together they guarantee that an experiment measures what it claims to measure.** Randomization assigns the treatments to the experimental units in a random order, so that the effects of the treatments are not systematically mixed with the effects of uncontrolled variables that vary with time or position, and it is the basis of the validity of the statistical analysis because it makes the observations exchangeable under the null hypothesis. Replication applies each treatment to more than one experimental unit, so that the experiment can estimate the pure error against which the treatment differences are tested, and without replication there is no way to know whether a difference is real or simply noise. Blocking groups the experimental units into homogeneous blocks, such as wafers from the same lot or positions on the same wafer, so that the known variation between blocks can be separated from the variation due to the treatments, which makes the estimate of the treatment effects more precise. The three principles work together: randomization protects against unknown biases, replication provides the error estimate, and blocking removes the effect of known nuisance variables. An experiment that follows these three principles yields data whose analysis is straightforward and whose conclusions are credible.
**The vocabulary of experimental design is precise, and it is worth fixing the terms before discussing the designs themselves, because the words carry the structure of the whole subject.** A factor is a process variable that the engineer deliberately changes in the experiment, such as temperature, pressure, or gas flow, and a level is one of the specific values at which a factor is set. A treatment is a particular combination of factor levels applied to an experimental unit, and an experimental unit is the smallest division of the material that can receive a treatment independently, such as an individual wafer or a single die site. The response is the measured outcome, such as film thickness, etch rate, defect count, or yield, and the experiment seeks to explain the variation in the response as a function of the factors and their interactions. A main effect is the change in the average response that is attributable to a change in a single factor, while an interaction is the change in the response that occurs when the effect of one factor depends on the level of another factor. The design matrix is the tabular specification of which treatment each run receives, and the entire enterprise of experimental design is the choice of a design matrix that estimates the main effects and interactions as precisely and as cheaply as possible.
**The history of experimental design is the story of how agriculture and industry learned to ask questions that their data could actually answer, and it reaches back to the statistical revolution of the early twentieth century.** Ronald Fisher, working at the Rothamsted agricultural station, introduced the principles of randomization, replication, and blocking in his book on the design of experiments in 1935, and he developed the analysis of variance and the factorial design that remain the core of the subject. Frank Yates, Fisher's collaborator at Rothamsted, developed the Yates algorithm for computing factorial effects efficiently and contributed the fractional factorial and confounding concepts, and George Box, working in the chemical industry, developed the response surface methodology and the class of designs that bear his name. William Cochran and Gertrude Cox wrote the classic textbooks that codified the standard designs, and Douglas Montgomery's text became the standard industrial reference that brought designed experiments into process engineering. Genichi Taguchi carried the ideas into quality engineering with his orthogonal arrays and robust parameter design, while Robin Plackett and John Burman provided the economical screening designs. This lineage shows that experimental design is a mature, battle-tested methodology, and the semiconductor fab is one of the modern settings in which it pays the greatest dividends, because the cost of an experimental run is so high that the design must be chosen with care.
**The simplest valid design is the completely randomized design, in which each experimental unit is assigned at random to one of the treatments, and it is the appropriate design when the experimental units are homogeneous, as when a batch of wafers is expected to be uniform.** In the completely randomized design the treatments are replicated, so that each treatment appears a fixed number of times, and the entire set of experimental units is randomized, so that any difference between the treatment groups can be attributed to the treatments and the random error alone. The analysis of a completely randomized design is a one-way analysis of variance, which partitions the total variation in the response into the variation between treatment groups and the variation within treatment groups, and it tests whether the group means differ by more than the within-group noise. The completely randomized design is simple and provides the maximum flexibility for unequal replication, but it is only efficient when the experimental units are genuinely homogeneous, and it wastes the opportunity to remove known sources of variation. In a fab the completely randomized design is used when the engineer can randomize the assignment of wafers to chambers or recipes without regard to lot structure. When the experimental units fall into natural groups that differ from one another, the completely randomized design is not the best choice.
**The analysis of a completely randomized design illustrates the general logic that runs through every designed experiment, and it is worth following the reasoning once in detail because the same structure reappears everywhere.** The total variation in the response, measured by the total sum of squares, is split into the between-treatment sum of squares, which measures how much the treatment averages differ from the overall average, and the within-treatment sum of squares, which measures the variation of the individual observations about their own treatment averages. Each sum of squares is divided by its degrees of freedom to give a mean square, and the F statistic for the treatment effect is the ratio of the treatment mean square to the error mean square. If the treatments have no real effect, then the two mean squares both estimate the same underlying error variance and their ratio is near one, whereas if the treatments differ, the treatment mean square is inflated and the F statistic is large. The p value associated with the F statistic is the probability of seeing a ratio at least that large when the treatments have no effect, and a small p value is the evidence that the treatments matter. This analysis of variance is the statistical heart of every designed experiment, and the engineer who understands it can read the output of any statistical package.
**The power of a factorial experiment can be concentrated on the effects that matter by the deliberate pooling of negligible terms, and this practice, known as pooling or collapsing, is a standard part of the analysis of the larger designs.** When the analysis shows that a high-order interaction is negligible, its sum of squares and degrees of freedom are added to the error, which increases the error degrees of freedom and makes the F tests for the remaining effects more powerful. The guiding assumption is that high-order interactions, such as a three-factor interaction, are usually small, an assumption that is the basis of the sparsity of effects principle that underlies the fractional factorial designs. The sparsity of effects principle states that in most systems only a few of the many possible effects are large, so that the vast majority of main effects and interactions can be treated as negligible, and this principle is what makes the screening of many factors practical. The practical consequence is that the engineer should not clutter the model with every possible term, but should retain only the effects that are clearly significant and pool the rest into the error. In a fab this disciplined reduction yields a parsimonious model that is easier to interpret and more powerful than a model cluttered with noise. The sparsity of effects principle is one of the most powerful ideas in experimental design.
**The design of a block experiment depends on the balance between the number of treatments and the size of the natural blocks, and when the blocks are too small to hold a full replicate of every treatment, the engineer must use an incomplete block design.** In a balanced incomplete block design, not every treatment appears in every block, but the design is arranged so that every pair of treatments appears together in the same number of blocks, which keeps the estimates of the treatment differences balanced and precise. The analysis of a balanced incomplete block design adjusts the treatment means for the block effects, and it recovers information from the blocks to estimate the treatment differences efficiently. The incomplete block design is the tool for the common situation in which a lot holds only a few wafers but the experiment must compare many recipes, so that each lot can carry only a subset of the treatments. The balanced incomplete block design generalizes the randomized complete block design and is the natural next step when the block size is limited. In a fab the incomplete block design compares many recipes across lots that are too small to hold them all.
**The analysis of a designed experiment is only as good as the model that fits the data, and the model for a designed experiment is fitted by the method of least squares, which is the same regression machinery that underlies the analysis of variance.** In the regression representation of a designed experiment, the response is written as a linear function of the factor settings and their products, and the least squares estimates of the coefficients are the effects that the analysis reports. The analysis of variance is in fact a special case of linear regression, and the sums of squares, the F tests, and the p values all follow from the least squares fit of the model to the data. This unity means that the engineer who understands regression understands the analysis of designed experiments, and it means that the powerful tools of regression, such as residual analysis and prediction intervals, apply directly to the designed-experiment data. The residuals from the fitted model are examined for patterns that reveal a missing curvature, a nonconstant variance, or an outlier, and the model is refined until the residuals look like pure noise. In a fab the least squares fit of the designed-experiment model produces the predictive equation that the process engineer uses to set and to control the process.
**The randomized complete block design improves on the completely randomized design by grouping the experimental units into homogeneous blocks, each of which contains one replicate of every treatment, and it is the design of choice when the experimental units fall into natural groups such as lots or spatial positions.**** In the randomized complete block design each block is a set of experimental units that are expected to be as alike as possible, the treatments are assigned at random within each block, and each treatment appears exactly once per block, so that the differences between blocks can be estimated and removed from the comparison of the treatments. The analysis is a two-way analysis of variance with block and treatment as the two factors, and the blocking effect soaks up the between-block variability, leaving a smaller error with which to test the treatment differences. The randomized complete block design is more efficient than the completely randomized design whenever the blocks differ, because it removes the block variation from the error, and it is the standard design for comparing several recipes across several lots or across the positions of a wafer. The Latin square design extends this idea to two blocking variables at once, arranging the treatments in a square so that each treatment appears once in each row and once in each column. In a fab the randomized complete block design compares process recipes across multiple lots, and the Latin square design handles two nuisance dimensions such as position on a wafer and position in a load.
**The factorial design is the heart of modern experimental design, because it allows the engineer to study the effect of several factors simultaneously, including the important phenomenon of interaction, and it is far more efficient than changing one factor at a time.** In a full factorial design with two factors, one at $a$ levels and one at $b$ levels, the design runs every one of the $a \times b$ combinations of factor levels, so that every main effect and every interaction is estimated from the data. A 2^k factorial design, in which each of $k$ factors is set at two levels, commonly called low and high, runs $2^k$ experiments and estimates $k$ main effects and all of their interactions, and it is the workhorse of screening experiments because it detects the important factors with a modest number of runs. The factorial design reveals interactions that the one-factor-at-a-time approach cannot detect, and it estimates the effect of each factor with the full precision of every run, rather than wasting runs on the uninformative repeated centers of a one-factor approach. When a factor does not matter or when its effect is negligible, the engineer can pool its contribution into the error to gain power for the remaining factors. In a fab the 2^k factorial design is the standard way to identify which process parameters, such as temperature, pressure, and flow, actually drive a response.
**The analysis of a factorial design is built on the sums of squares and the F test, and it identifies which effects are real against the estimate of the experimental error.** The total variation in the response is partitioned into a component for each main effect, a component for each interaction, and a component for pure error, and each effect is tested by comparing its mean square with the error mean square. The F statistic for an effect is the ratio of the effect mean square to the error mean square, and under the null hypothesis that the effect is zero it follows the F distribution, so that a sufficiently large F statistic is evidence that the effect is real. The degrees of freedom for each source of variation count the number of independent pieces of information that it uses, and they must sum to the total degrees of freedom, which is one less than the number of runs. The analysis produces a table of sources, sums of squares, degrees of freedom, mean squares, F statistics, and p values that is the standard summary of a designed experiment, and it tells the engineer which factors and interactions matter. The mean square of an effect also provides an estimate of the magnitude of that effect, so that the analysis identifies both which effects are significant and how large they are. In a fab the analysis of variance of a designed experiment reveals which process parameters matter, how strongly they act, and whether they interact.
**The interaction between two factors is the phenomenon that makes factorial design indispensable, because it is the effect that a one-factor-at-a-time experiment can never see.** Two factors interact when the effect of one factor depends on the level of the other, so that changing factor A changes the response differently at the low and high levels of factor B. An interaction is detected in the analysis of variance as a term with its own sum of squares and F test, and it is interpreted graphically as a crossing or fanning pattern in the plot of the response against one factor at the two levels of the other. When a strong interaction is present, the main effects alone are not a reliable guide, because the best level of one factor depends on the level of the other, and the engineer must choose the combination of levels that maximizes or minimizes the response. The two-way interaction is the most common and the most important, but higher-order interactions among three or more factors also occur, and they are typically smaller and harder to interpret. In a fab interactions explain why a recipe that works at one temperature fails at another, and why the optimal gas flow depends on the pressure. The detection of interactions is the principal reason that the factorial design replaced the one-factor-at-a-time approach, and it is the reason that modern process development is built on designed experiments.
**The full 2^k factorial design runs every combination of the two levels of each of $k$ factors, and it provides the best possible estimates of all effects, but its cost grows exponentially, so that a design with seven factors requires $2^7 = 128$ runs.** When the number of factors is large, the full factorial is often too expensive, and the fractional factorial design comes to the rescue by running only a fraction of the combinations while still estimating the important effects. A one-half fraction of a 2^k design runs $2^{k-1}$ combinations, a one-quarter fraction runs $2^{k-2}$, and so on, and each fraction sacrifices the ability to estimate some higher-order interactions, which are usually assumed negligible, in exchange for a great reduction in cost. The price of a fractional design is that some effects are confounded, meaning that their estimates are algebraically entangled and cannot be separated, and the confounding pattern is summarized by the resolution of the design. A design of resolution III estimates all main effects but confounds them with two-factor interactions, a design of resolution IV estimates main effects free of two-factor interactions but confounds two-factor interactions with one another, and a design of resolution V estimates main effects and two-factor interactions free of one another. In a fab the fractional factorial is the standard screening tool when many process parameters are candidates and the engineer must identify the few that matter with a small number of runs.
**The resolution of a fractional factorial design is a single number that encodes how the effects are confounded, and it is the key to choosing a fractional design that can answer the engineer's question.** A design of resolution III is generated by defining the alias structure so that main effects are confounded with two-factor interactions, which is acceptable when the engineer only needs to screen many factors to find the few that matter, on the assumption that interactions are small. A design of resolution IV confounds two-factor interactions with one another but keeps the main effects clear of them, which is better for separating real effects from interactions. A design of resolution V keeps the main effects and the two-factor interactions all clear of one another, at the cost of more runs, and it is the smallest fractional design that fully estimates the first-order model of main effects and pairwise interactions. The general principle is that the higher the resolution, the less confounding and the more runs, and the engineer chooses the resolution that matches the goal of the experiment and the budget for runs. In a fab a resolution III screening design with many factors identifies the active parameters in a handful of runs, and a follow-up resolution V design then quantifies the interactions among the active parameters. The systematic use of fractional factorials lets an engineer screen broad process spaces cheaply and then focus expensive runs where they matter.
**The orthogonal array is the combinatorial structure that underlies both the fractional factorial and many classical designs, and it is the organizing principle that the engineer can exploit to design experiments that are balanced and efficient.** An orthogonal array is an array of factor settings in which, for any pair of columns, each combination of the two factors' levels occurs an equal number of times, and this balance is what makes the columns independent and the effects estimable. The Taguchi method, developed by Genichi Taguchi, popularized the use of orthogonal arrays for robust parameter design, in which the engineer chooses factor settings that make the product insensitive to noise variation, and it introduced the concept of a signal-to-noise ratio as the response. The orthogonal arrays are denoted by notation such as L9 and L16, which describe the number of runs and the number of factors they can accommodate, and they provide ready-made balanced designs for a wide range of factor counts and level counts. The use of orthogonal arrays makes it possible to assemble a balanced experiment without deriving the design from scratch, and it is especially useful when factors have different numbers of levels. In a fab the orthogonal array designs are used both to screen process factors and to tune a process so that the product quality is robust to the inevitable variation in the operating conditions.
**The experiment must be sized before it is run, and the power analysis determines how many runs are needed to detect an effect of a given size with a given probability, and this sizing is as important as the choice of the design itself.** The power of a test is the probability that it will detect an effect of a specified magnitude when that effect is real, and it depends on the effect size, the variance of the response, the sample size, and the significance level. To compute the required number of replicates, the engineer specifies the smallest effect that must be detected, the estimate of the noise standard deviation, the significance level, and the desired power, and the sample size formula or a power curve then gives the number of runs per treatment. Increasing replication increases power, but the gain is subject to diminishing returns, because the standard error of an effect shrinks only as the square root of the number of replicates. In a fab, where each run costs an entire wafer or lot, the power analysis is essential to avoid the double waste of an underpowered experiment that cannot detect the real effect and an overpowered experiment that wastes scarce resources. The power analysis is the quantitative bridge between the engineering goal and the design size. A properly sized experiment detects the effects that matter with a pre-specified confidence and no more runs than necessary.
**The confirmation run is the final check that closes the loop of a designed experiment, and it is the run that establishes whether the fitted model really describes the process rather than the accidents of a particular dataset.** After the analysis identifies the significant factors and the model predicts the response at a chosen set of settings, the engineer runs one or more new wafers at those settings and compares the observed response with the prediction interval of the model. If the confirmation response falls inside the prediction interval, the model is confirmed and can be trusted to set the process; if it falls outside, the model is suspect and the experiment must be examined for a missing factor, a curvature, or an error in the recorded settings. The confirmation run is the safeguard against the danger of overfitting, in which a model describes the noise of one experiment and fails on the next, and it is the reason that a designed experiment concludes with a decision rather than a mere table of statistics. In a fab, where the cost of a wrong recipe is measured in lost lots, the confirmation run is the difference between a confident process change and a risky gamble. The disciplined experiment therefore always ends with the empirical test of its own prediction.
The process of carrying out a designed experiment in a fab follows a disciplined workflow that moves from a clear question to an actionable model, and it is worth stating as a sequence of steps that the engineer repeats for every experiment.** The workflow begins by defining the objective and choosing the response, then listing the candidate factors, their ranges, and their levels, and selecting the design, the resolution, and the number of replicates based on the power analysis. The next steps are to randomize the run order, block the runs when necessary, and actually execute the runs while carefully recording the actual factor settings, because the settings that were planned are not always the settings that were achieved. After the data are collected, the engineer plots and examines them, runs the analysis of variance, identifies the significant effects, and fits a model that relates the response to the active factors, then validates the model with confirmation runs at the predicted optimum. The final step is to document the experiment, the model, and the decision, so that the knowledge is captured for the process and for future experiments. This workflow keeps the experiment honest at every step and turns a set of wafers into a quantitative model of the process. An engineer who follows this workflow extracts the maximum information from every run.
**The response surface methodology extends the factorial design beyond the identification of significant factors to the optimization of a continuous response, and it is the toolkit for finding the factor settings that maximize or minimize a response such as yield or minimize a response such as defect density.** A response surface design is a two-level factorial augmented with center points and axial points, so that the engineer can fit a quadratic model that describes the curvature of the response, and the central composite design is the most widely used of these designs. The central composite design consists of a factorial portion, a set of center points that estimate the pure error and the curvature, and a set of axial points that make the quadratic model estimable, and it requires only a modest number of runs beyond the factorial. The fitted quadratic response surface is visualized as a contour plot or a three-dimensional surface, and the optimum, whether a maximum, a minimum, or a saddle, is located by examining the fitted surface and using the method of steepest descent or ascent. The response surface approach is iterative: the engineer first screens with a factorial, then fits a first-order model and moves toward the region of the optimum, and finally fits a quadratic model in the optimal region. In a fab the response surface methodology tunes a process to its optimum, such as the temperature, pressure, and flow that maximize the uniformity of a film or minimize the etch bias. The result of a response surface study is a predictive equation that the engineer uses both to set the process and to understand its sensitivity.
**The response surface is quantified by the fitted quadratic model, and the interpretation of that model is the practical payoff of the methodology, because it gives the engineer a predictive equation for the response.** A quadratic response surface in two factors $x_1$ and $x_2$ is a model of the form $y = \beta_0 + \beta_1 x_1 + \beta_2 x_2 + \beta_{11} x_1^2 + \beta_{22} x_2^2 + \beta_{12} x_1 x_2 + \epsilon$, in which the linear terms describe the slope, the squared terms describe the curvature, and the cross term describes the interaction. The fitted surface is a paraboloid, and its stationary point, where the gradient is zero, is found by setting the partial derivatives to zero and solving the resulting linear equations, and this stationary point is a maximum, a minimum, or a saddle point depending on the coefficients. The engineer examines whether the stationary point lies inside the experimental region, whether it is a genuine optimum rather than a saddle, and how flat or sharp the surface is near the optimum, because a flat surface means that the response is insensitive to the factor settings and robust. The contour plot of the fitted surface makes all of this visible, and it is the standard way to communicate the results of a response surface study. In a fab the fitted response surface gives a quantitative recipe for the optimum and a map of how the response changes as the factors move away from the optimum. This predictive equation is the reason that response surface studies are worth the additional runs.
**The screening design is the deliberate use of a small experiment to identify the few factors that matter from a large set of candidates, and it is the natural first step in a sequential strategy for process development.** When there are many candidate factors and no prior knowledge of which ones matter, a resolution III fractional factorial or a Plackett-Burman design can estimate the main effects of many factors in a small number of runs, on the assumption that most factors are inert and interactions are small. The Plackett-Burman design, developed by Robin Plackett and John Burman, is a class of two-level screening designs that accommodate factors in runs that are multiples of four, and it provides a very economical way to screen a large number of factors. The output of a screening experiment is a ranking of the factors by the magnitude of their main effects, usually displayed in a Pareto chart, which identifies the few factors that dominate the response. The active factors found by screening are then carried forward into a follow-up experiment, often a full or fractional factorial with higher resolution, that estimates the interactions and optimizes the response. In a fab screening is used when a new process has dozens of tunable parameters and the engineer must find the handful that control the yield. The screening design is the efficient front end of a two-stage strategy that spends few runs on discovery and then concentrates runs on the important factors.
**The choice among the many possible designs is governed by the goal of the experiment, the number of factors, the budget, and the anticipated presence of interactions, and the following table organizes the standard designs by their purpose and their typical run counts.** The table makes it easy to select a starting design for a given engineering question, and it shows how the designs progress from the simplest comparisons to the most sophisticated optimization. The engineer reads the table by matching the experimental goal to a design, and then sizes it with the power analysis.
| Design | Purpose | Factors | Typical runs | Notes |
|---|---|---|---|---|
| Completely randomized | compare treatments, homogeneous units | any | k × r | simplest, flexible |
| Randomized complete block | remove one nuisance dimension | any | k × b | one blocking factor |
| Latin square | remove two nuisance dimensions | 3+ | k × k | rows and columns block |
| Full 2^k factorial | estimate all effects and interactions | 2–6 | 2^k | no confounding |
| Fractional 2^(k−p) | screen many factors cheaply | 5+ | 2^(k−p) | some effects confounded |
| Central composite | fit quadratic, find optimum | 2–6 | 2^k + 2k + center | response surface |
| Plackett-Burman | very cheap screening | 7–30+ | multiples of 4 | main effects only |
| Box-Behnken | quadratic fit, fewer runs than CCD | 3–7 | economical | response surface |
| Taguchi / orthogonal array | robust parameter design | mixed | L9, L16… | signal-to-noise response |
**The selection of a design is guided by a decision tree based on the goal, the number of factors, the need for blocking, and the budget, and the following flowchart routes an experiment to the appropriate design.** The first question is whether the engineer is comparing a small set of treatments, screening many factors, or optimizing a response; the second is whether the experimental units fall into blocks; and the third is whether interactions are of interest. Working through these questions selects a design that is both efficient and capable of answering the question.
```flowchart
A([Design a new experiment]) --> B{Primary goal?}
B -- compare few treatments --> C{Units blockable?}
C -- yes --> D[Randomized complete block]
C -- no --> E[Completely randomized]
B -- screen many factors --> F{How many runs allowed?}
F -- very few --> G[Plackett-Burman]
F -- moderate --> H[Fractional 2^(k-p), resolution III/IV]
B -- estimate effects & interactions --> I[Full 2^k or resolution V]
B -- optimize a response --> J[Central composite / Box-Behnken]
B -- robust to noise --> K[Taguchi / orthogonal array]
B -- two nuisance dimensions --> L[Latin square]
```
**The connection between experimental design and the other keywords in the series is strong, and it completes the applied statistics toolkit that the series has been assembling.** The statistics basics keyword supplies the descriptive measures and the concept of variation that the designed experiment quantifies, while the inference statistics keyword supplies the analysis of variance, the F test, and the concept of significance that the experiment uses to decide which effects are real. The probability stats keyword supplies the distributions, including the F distribution, that the tests rely on, and the nonparametric statistics keyword provides the rank-based alternatives that are used when the designed-experiment data are not normal. The stochastic processes keyword supplies the time-ordered structure that underlies the randomization of run order and the drift that blocking controls. Experimental design, in turn, is the engine that produces the clean, informative data on which all of the statistical inference rests, so that it is the indispensable partner of the entire statistical toolkit. The engineer who combines a well-designed experiment with the appropriate statistical analysis extracts the maximum information from the expensive wafers that a fab produces.
**A concrete example ties the tools together and shows how a designed experiment is actually run, and the example of optimizing the uniformity of a deposited film illustrates the complete workflow.** The engineer selects the response, the non-uniformity of film thickness, and identifies temperature, pressure, and gas flow as the candidate factors, each with two levels, and because the experimental units come from different lots, the engineer chooses to block by lot. The engineer runs a full $2^3$ factorial with two replicates, randomized within each lot block, and the analysis of variance shows that temperature and the interaction between pressure and flow are significant while the other effects are not. The engineer then augments the design with center points and axial points in the region of the interesting settings, fits a quadratic response surface, and locates the temperature, pressure, and flow that minimize the non-uniformity. The engineer validates the predicted optimum with confirmation wafers, which confirm that the non-uniformity at the predicted settings matches the model, and then writes the recipe into the process. This single example shows that experimental design is not a statistical ritual but the disciplined method by which a fab turns a handful of wafers into a verified quantitative recipe.
**The closing lens for experimental design is that it is the methodology that determines the information content of the data before a single wafer is run, and the value of the subject is not the machinery of the analysis of variance or the F test in isolation but the recognition that the design, chosen before the data exist, governs what the data can and cannot reveal.** With this lens the engineer sees randomization, replication, and blocking not as administrative niceties but as the guarantees that the measured effects are real and the error is honestly estimated, sees the factorial design as the way to detect the interactions that a one-factor-at-a-time approach can never see, and sees the fractional factorial and the response surface as the tools that spend the scarce runs where they matter most. The mastery of experimental design is the mastery of extracting the maximum information from every expensive wafer and every limited lot, which is precisely the challenge that governs semiconductor process development every day. Read experimental design through a data-economy lens rather than a formula-catalog lens.