Electronics Guide

Why One Factor at a Time Fails

The instinctive way to investigate a process is to hold everything constant except one setting, vary it, record the result, and move on. This procedure, called one factor at a time or OFAT, is inferior to a factorial design on every criterion that matters. The decisive objection is that it cannot detect an interaction, and interactions are the normal condition in manufacturing rather than the exception. An interaction exists when the effect of one factor depends on the level of another. In reflow, raising peak temperature improves wetting and reduces voiding when the time above liquidus is short, but does nothing useful and begins to damage the laminate when that time is long. An OFAT experiment run at the current soak measures that effect at one soak setting and reports it as though it were general.

The second objection is efficiency. In a two-level factorial every run contributes to every effect estimate, since a main effect is the average of all runs at the factor's high level minus the average of all runs at its low level; in sixteen runs that comparison averages eight observations against eight. Statisticians call this hidden replication, and it costs nothing. OFAT has none, because each comparison uses two runs.

The third objection follows. OFAT optimizes along the coordinate axes and halts when no single change improves the response. On a surface with a ridge running diagonally across the factor space, which is exactly what an interaction produces, that procedure stops on the side of the ridge rather than at its peak, while a factorial followed by a path of steepest ascent climbs it directly.

Full Factorial Designs

A factor is an input the experimenter sets, a level is a value it takes, and the response is the measured output. The main effect of a two-level factor is the average response at its high level minus the average at its low level, and a two-factor interaction is half the difference between the effect of the first factor at the second factor's high level and the same effect at its low level. Factors are written in coded units, low as minus one and high as plus one, which makes the design matrix orthogonal and each regression coefficient exactly half the corresponding effect.

A full factorial runs every combination of levels: with k factors at two levels the design has two to the power k runs, eight for three factors and sixteen for four. An eight-run design in three factors estimates three main effects, three two-factor interactions, and one three-factor interaction, which together with the mean consume all eight degrees of freedom, so nothing remains to estimate error. That is why unreplicated factorials need the graphical methods described later. Effects are computed as contrasts: form each interaction column by multiplying main-effect columns element by element, take the dot product of each column with the response vector, and divide by half the number of runs.

Interpretation follows a firm rule: a main effect involved in a large interaction must never be quoted alone. Suppose a wave solder experiment gives preheat a main effect of plus twelve percentage points of barrel fill and a preheat-by-flux-volume interaction of plus eight. Raising preheat then helps by twenty points at high flux volume and by four at low, so reporting only the twelve misleads anyone who later runs the line at low flux volume. Present the interaction plot instead: parallel lines mean no interaction, while diverging or crossing lines show it directly, and a plant engineer reads them without statistical training.

Two levels cannot detect curvature, because two points define a line. Where curvature is expected the options are center points, which detect it without attributing it, or three levels per factor. A full three-level factorial grows faster still, twenty-seven runs for three factors and eighty-one for four, so response surface designs are almost always the better answer.

Fractional Factorial Designs and Resolution

A full factorial in seven factors takes one hundred twenty-eight runs, nearly all spent estimating high-order interactions that are, in real processes, negligible. Two empirical principles justify cutting the design down. Sparsity of effects holds that a response is usually driven by a few main effects and a few two-factor interactions; effect heredity holds that a large interaction almost always involves at least one factor whose main effect is also large. A fractional factorial runs a chosen half, quarter, or eighth of the full design, selected so it stays orthogonal and balanced.

Generators, the Defining Relation, and Aliasing

A half fraction of a three-factor design is built by writing the two-factor design in A and B and setting the third column equal to their product: C equals AB. That equation is the generator, and multiplying both sides by C, using the fact that any column times itself is a column of plus ones denoted I, gives the defining relation: I equals ABC.

The defining relation determines the entire alias structure, since multiplying it by any effect gives that effect's alias. Multiply I equals ABC by A and the result is A equals BC: the main effect of A and the BC interaction come from the same column of signs, and no analysis can separate them. What is reported as the main effect of A is in truth the sum of A and BC, and if the two have opposite signs they may cancel. Aliasing is the price of the reduction in runs; the skill lies in placing the confusion where it does least harm.

The Resolution Numbering

Resolution summarizes how bad the aliasing is. A design of resolution R confounds no effect involving p factors with any effect involving fewer than R minus p factors, and the numeral equals the length of the shortest word in the defining relation.

Resolution III. Main effects are clear of one another but aliased with two-factor interactions. These designs are extremely economical, packing seven factors into eight runs or fifteen into sixteen, and they are the workhorses of screening. Their accepted use is to identify which factors deserve further study, on the understanding that a large apparent effect may be a main effect, an interaction, or a mixture. Treating a resolution III result as a final model is a common and serious error.

Resolution IV. Main effects are clear of two-factor interactions, but two-factor interactions are aliased with one another in pairs. A sixteen-run design in eight factors is resolution IV and is often the best compromise available. If AB and CD share a column showing a large effect, subject-matter knowledge or a follow-up experiment must decide which pair is responsible.

Resolution V. Main effects and two-factor interactions are all clear of one another, with two-factor interactions aliased only against three-factor interactions assumed negligible. The sixteen-run half fraction of a five-factor design, generated by setting E equal to ABCD, is the classic example, and it is the natural choice once screening has narrowed the field to four to six factors.

Augmenting a Fraction

Fractional designs are meant to run sequentially. The most useful augmentation is the fold-over: a second block in which every sign is reversed, converting a resolution III design into a resolution IV design at the cost of doubling the run count. A partial fold-over, reversing the signs of a single column, de-aliases one factor and its interactions in half as many additional runs, and is the right response when a screen produces one ambiguous but important result.

Screening Designs

Screening is the first stage of an experimental program, undertaken when a process has more candidate factors than anyone can afford to study properly. Its purpose is to reduce twenty suspects to the three or four that drive the response, so its output is a shortlist rather than a recommendation.

R. L. Plackett and J. P. Burman published The Design of Optimum Multifactorial Experiments in Biometrika in 1946, describing two-level designs whose run counts are multiples of four. The twelve-, twenty-, twenty-four-, and twenty-eight-run members fill the gaps between the powers of two available from regular fractions, and a design in N runs accommodates up to N minus one factors. Where N is a power of two the design is a regular fractional factorial; where it is not, it is non-regular. In the twelve-run design, each main effect is partially aliased with every two-factor interaction not involving it, at a correlation of one third in magnitude, rather than completely confounded with a few of them. The bias from any single interaction is diluted, but the contamination spreads across many terms rather than concentrating in identifiable ones.

Bradley Jones and Christopher Nachtsheim introduced definitive screening designs in the Journal of Quality Technology in 2011, in A Class of Three-Level Designs for Definitive Screening in the Presence of Second-Order Effects. They use three levels per quantitative factor and require as few as twice the number of factors plus one run. Main effects are orthogonal to two-factor interactions and unbiased by any second-order effect, no two-factor interaction is completely confounded with another, and quadratic effects are estimable. One design can therefore sometimes serve as both screen and optimization.

The most common cause of a failed screen is not the design but the factor ranges. Levels set too close together bury the effects in noise; levels set too far apart produce runs that fail outright, leaving holes in the design matrix.

Randomization, Blocking, and Run Order

Randomization means deciding run order by a random mechanism rather than by convenience, and its purpose is to protect the conclusions from unknown, time-varying disturbances. Manufacturing equipment drifts: a solder pot loses tin to dross across a shift, a plasma chamber's wall condition evolves from run to run, a stencil accumulates paste residue. Suppose a reflow experiment is executed in the textbook order, all low-temperature runs before all high-temperature runs, and the oven's belt bearings warm through the session so the actual dwell shortens. The temperature effect and the drift are then perfectly confounded, and no analysis can disentangle them. Randomizing converts that systematic bias into random error, which widens the confidence intervals rather than moving the estimates, and it validates the significance test itself, whose probability statements rest on independent errors.

Some nuisance variables are known in advance and cannot be held constant: an experiment may need two lots of prepreg, two operators, or three days. Blocking groups the runs so each block is internally homogeneous, balanced across blocks so the block effect does not contaminate the factor effects. An eight-run design in three factors splits into two blocks of four by assigning runs according to the sign of the three-factor interaction column, which confounds the block effect with that interaction. That is an acceptable trade, because a three-factor interaction is unlikely to be real and a lot difference is very likely to be real.

Complete randomization is sometimes physically impossible, since a reflow oven takes ten to twenty minutes to stabilize after a setpoint change and a lamination press cycle is measured in hours. When randomization is restricted, that restriction is a feature of the design and must appear in the analysis, which is what the split-plot structures below provide.

Center Points and Curvature

A two-level design fits a model of linear terms and cross-products. It cannot represent a peak or a valley and cannot warn that one exists. Runs at the center of the design space, where every quantitative factor sits at the midpoint of its range, remedy both problems cheaply.

The curvature test is a comparison of averages. If the response surface is planar, the average of the factorial corner runs and the average of the center runs should differ only by chance; if it is curved, the center sits systematically above or below the plane fitted to the corners, and the difference between the two averages estimates the sum of the pure quadratic terms. The test is incomplete, since all the pure quadratic terms are confounded in a single degree of freedom, so it cannot say which factor causes the curvature. Separating them requires the axial or edge runs of a response surface design, which is why a significant curvature test is the standard trigger for augmentation.

Center points serve two further purposes. Replicated center runs give a model-free estimate of pure error at a fraction of the cost of replicating the whole design, supplying the denominator for significance tests in an otherwise saturated experiment. They also provide a stability check: if center runs executed at intervals drift systematically, the process was not stable. Three to five, distributed through the run order, is the usual allocation.

Response Surface Methods

Screening and factorial designs answer which factors matter; response surface methodology answers where to set them. Introduced by George Box and K. B. Wilson in 1951 in the Journal of the Royal Statistical Society, it treats optimization as a sequential search: use a first-order design to find a direction of improvement, follow it until it stops paying, then fit a second-order model near the optimum.

The Path of Steepest Ascent

When a factorial fits a plane with no significant curvature, the process is on a slope rather than at a peak, and the fastest route to improvement is the gradient. In coded units the direction of steepest ascent is proportional to the fitted first-order coefficients, so a coefficient twice as large as another moves that factor twice as far per step. The experimenter runs single trials along the path; the response improves and then stops or reverses, and that turning point becomes the center of a new factorial design.

Central Composite and Box-Behnken Designs

A central composite design augments a two-level factorial with axial points and center points. The factorial portion, a fraction of at least resolution V, supplies the linear and interaction terms; two axial runs per factor, at a distance denoted alpha along each coordinate axis, supply the pure quadratic terms.

Making alpha the fourth root of the number of factorial runs makes the design rotatable, so the variance of a prediction depends only on its distance from the center and not on its direction. For three factors with an eight-run factorial core that value is approximately 1.682 in coded units, which places the axial runs outside the cube. That is sometimes impossible, because a plasma will not strike or a pot will not reach temperature. Setting alpha to one gives a face-centered design, with axial runs on the faces of the cube and only three levels per factor; it loses rotatability but never requests a setting outside the declared feasible range, which makes it common in manufacturing.

Box and Behnken published their alternative in Technometrics in 1960 as Some New Three Level Designs for the Study of Quantitative Variables. It places runs at the midpoints of the edges of the experimental cube together with center points, giving fifteen runs for three factors, and its distinguishing property is that no run sets all factors to an extreme at once. In a central composite design the corner runs place every factor at its limit simultaneously, and in some processes that is the combination that fails. Box-Behnken designs require at least three factors and do not build up sequentially from an earlier factorial.

Reading and Using the Surface

A fitted second-order model is interrogated through contour plots and canonical analysis. Contour plots, drawn two factors at a time, are the most useful form for a process engineer because they show the operating window rather than a point, and a wide, flat region of acceptable response is worth more than a narrow optimum. Canonical analysis classifies the stationary point as a maximum, a minimum, or a saddle; a saddle, which is common, means the surface rises along a ridge and the useful output is the ridge direction.

Most real problems have several responses to satisfy at once: a reflow profile must produce adequate wetting, acceptable voiding, and a peak temperature below the package rating. The standard technique is desirability, in which each response is mapped to a scale from zero for unacceptable to one for ideal and the individual desirabilities are combined into a geometric mean that the optimizer maximizes. A geometric mean is used because a zero on any single response drives the whole to zero, correctly refusing any solution that violates a hard requirement.

Taguchi Methods and Robust Design

Genichi Taguchi's parameter design methods entered Western manufacturing in the early 1980s. The reception among statisticians was sharply divided: the engineering ideas were widely praised, the statistical machinery widely rejected, and both halves of that judgment have held up.

The Quality Loss Function and the Orthogonal Arrays

Taguchi's most durable contribution is conceptual. Conventional thinking treats quality as conformance: a part inside the limits is good, a part outside is bad. Taguchi argued instead that loss increases continuously as a characteristic departs from target, and modeled it as proportional to the square of the deviation. Reducing variation around target therefore has value even when every part already passes inspection, which reframes quality improvement as variance reduction rather than defect elimination. That idea underlies the capability indices used throughout Quality Control and Inspection.

Taguchi packaged designs as a catalog of orthogonal arrays named L4, L8, L9, L12, L16, L18, and L27, where the number is the run count. Several are familiar designs renamed: L8 is the eight-run resolution III fraction in seven two-level factors, L9 a three-level design in four factors, L12 the twelve-run Plackett-Burman design, and L18 a mixed-level array estimating main effects only. The catalog made designed experiments accessible to engineers who would never derive a defining relation, but it also encouraged the use of highly saturated resolution III arrays as final experiments rather than as screens.

Inner and Outer Arrays

The robust design idea is genuinely valuable and has been absorbed most completely. Control factors are those the manufacturer can set and hold: oven setpoints, stencil aperture dimensions, bond force. Noise factors vary in use or production and cannot economically be controlled: ambient humidity, incoming lot variation, board warpage. The product-array structure places control factors in an inner array and noise factors in an outer array, running the complete outer array at every combination of the inner, so each inner-array row yields a distribution of responses across deliberately varied noise conditions.

The mechanism that makes robustness possible is the interaction between control and noise factors. A control factor with no noise interaction shifts the mean but cannot change sensitivity to noise; one that does interact has a level at which the noise response flattens, and that level is the robust setting. This insight is the durable technical core of the method, developed further in Robust Design Methods. The cost of the crossed array is multiplicative: an L18 inner array crossed with an L4 outer array requires seventy-two runs.

Signal-to-Noise Ratios and the Statistical Criticism

Taguchi condensed each inner-array row into a signal-to-noise ratio, defined differently for three cases. For smaller-the-better characteristics such as void percentage, it is minus ten times the base-ten logarithm of the mean squared response. For larger-the-better characteristics such as pull strength, it is minus ten times the logarithm of the mean of the reciprocal squared responses. For nominal-the-best characteristics, it is ten times the logarithm of the squared mean divided by the variance. The analysis treats that ratio as the response and picks, for each factor, the level with the highest value.

Here the objections concentrate, and they are specific rather than rhetorical. George Box addressed the nominal-the-best measure in Signal-to-Noise Ratios, Performance Criteria, and Transformations, published in Technometrics in 1988, and showed that the criterion minimizes mean squared error about a target only where the standard deviation is proportional to the mean, and that in that case the same conclusion follows more simply from analyzing the logarithm of the response. Where the relationship between variance and mean differs, the ratio is the wrong summary. Box set out the broader critique the same year in An Explanation and Critique of Taguchi's Contributions to Quality Engineering in Quality and Reliability Engineering International.

Four further objections are standard. Combining location and dispersion into one number discards information, since the factors that shift the mean and those that shift the spread are frequently different. The heavy use of resolution III arrays leaves main effects confounded with two-factor interactions, while picking the best level for each factor independently assumes an additivity the design cannot verify. The analysis as originally taught relies on picking winners rather than on significance tests or residual analysis. And it is presented as one decisive experiment rather than a sequence.

The consensus resolution keeps the engineering and discards the arithmetic. A combined array places control and noise factors in one design rather than two crossed ones, fitting a single model containing control main effects, noise main effects, and the control-by-noise interactions. Robust settings are read from that model, and the mean is adjusted separately using factors that affect location without affecting dispersion. The combined array needs far fewer runs than the crossed array and supports significance testing and diagnostic checking.

Mixture Designs

An ordinary factorial assumes factors vary independently. In a formulation problem they cannot, because the components are proportions of a whole and must sum to one; raising the flux content of a solder paste necessarily lowers the fraction of something else. Henry Scheffé addressed this class of problem in Experiments with Mixtures, published in the Journal of the Royal Statistical Society in 1958.

The constraint has geometric consequences. With q components the feasible region is not a cube but a simplex of dimension q minus one: a triangle for three components, a tetrahedron for four. Because the proportions are linearly dependent, the usual polynomial is over-parameterized, so Scheffé's canonical polynomials drop the intercept and the pure quadratic terms, leaving linear blending terms and cross-products. A positive cross-product coefficient indicates synergistic blending; a negative one indicates antagonism. The standard designs are the simplex-lattice, a regular grid across the simplex, and the simplex-centroid; where the formulation is constrained further, extreme vertices designs place runs at the vertices and edge midpoints of the resulting polytope.

Electronics applications concentrate in materials development rather than assembly. Solder paste formulation balances alloy powder loading, flux vehicle, activator, rheology modifier, and solvent against print transfer efficiency, slump, tack life, and voiding. Conductive adhesive development trades silver filler loading against resin content, and encapsulant work balances filler loading against viscosity, coefficient of thermal expansion, and glass transition temperature. Where both formulation and process settings are under study, a mixture-process design crosses the two.

Split-Plot Designs for Hard-to-Change Factors

Complete randomization presumes every factor can be reset between runs at negligible cost, and in electronics manufacturing that presumption fails constantly. A split-plot design divides the factors in two. Whole-plot factors are hard to change: oven zone temperatures, a press thermal program, a chamber gas chemistry, each requiring a long stabilization or a cleaning cycle after every change. Sub-plot factors are easy to change: which stencil printed a board, which pad geometry a coupon carries, which position in the press book a panel occupies. The whole-plot factors are set, and several sub-plot conditions run within that setting before they change again.

The essential statistical fact is that such an experiment has two error terms. Whole-plot factors are effectively replicated only as often as the whole plots themselves, so they are tested against the variation between whole plots, which is large and carries few degrees of freedom. Sub-plot factors and their interactions are tested against within-whole-plot variation, which is smaller and carries many more. Sub-plot effects are therefore estimated with better precision than in a completely randomized design of the same size, an asymmetry worth exploiting by placing the factors you most need to resolve in the sub-plot.

The corresponding error is among the most common in industrial experimentation. An engineer runs a reflow experiment the only feasible way, changing oven setpoints a handful of times and processing many boards at each, then analyzes the data as though every run were independently randomized. The analysis pools both error terms, applies the small sub-plot error to the oven factors, and reports oven setpoints as overwhelmingly significant when the evidence supports no such claim. Correct analysis needs a mixed model with random whole-plot effects, and correct design needs the hard-to-change factors declared before generation.

Sample Size, Power, and Replication

For a two-level factorial with N runs in total, each main effect is the difference between two averages of N divided by two observations each, so the standard error of an effect is twice the process standard deviation divided by the square root of N. Detecting an effect at a five percent significance level with eighty percent power requires it to be roughly 2.8 standard errors, which rearranges to a run count of about thirty-one times the square of the ratio of process standard deviation to the effect sought. Detecting an effect equal to one process standard deviation therefore takes on the order of thirty-two runs. Halving the detectable effect quadruples the run count.

Three consequences follow. Reducing measurement variation is often cheaper than adding runs, because measurement error adds directly to the standard deviation in the numerator of that ratio. An experiment that cannot detect an effect worth acting on should not be run at all, since it will return a non-significant result that will be misread as evidence of no effect. And the power calculation needs a prior estimate of process standard deviation, for which existing control charts are the natural source.

Replication and repeated measurement must be distinguished rigorously. A genuine replicate requires the process to be reset: the oven returned to setpoint, the paste re-printed, the chamber re-seasoned. Measuring one panel at five locations gives five repeated measurements, which characterize within-panel uniformity but say nothing about run-to-run variation; treating them as replicates manufactures significance from nothing. Where both sources matter, and in wafer processing they almost always do, the correct treatment is a nested analysis estimating within-run and between-run components separately.

Reading the Results

Analysis of Variance and Lack of Fit

Analysis of variance partitions the total variation into components attributable to each effect and a residual. In a two-level design each effect carries one degree of freedom, and each mean square is divided by the error mean square to form an F statistic. Where replicated points exist, the error mean square should come from pure error rather than from pooling small effects into the residual, and the lack-of-fit test then compares the residual variation not explained by pure error against pure error itself. A significant result means the model is missing something real, usually curvature or an omitted interaction.

Effect Plots and Normal Probability Plots

A main-effect plot shows the mean response at each level of a factor, and a Pareto chart ranks effect magnitudes; neither substitutes for a significance test. Interaction plots are indispensable wherever an interaction is significant: in a wave solder experiment where conveyor speed interacts with preheat, the plot shows directly that speed matters greatly at low preheat and hardly at all at high preheat.

Unreplicated factorials have no error term, and the classical solution is graphical. Cuthbert Daniel introduced the half-normal plot for this purpose in Use of Half-Normal Plots in Interpreting Factorial Two-Level Experiments, published in Technometrics in 1959. Under the hypothesis that no factor does anything, the estimated effects are independent draws from a normal distribution centered on zero, so plotting them against normal scores produces a straight line through the origin, and real effects fall off that line. The slope through the inert effects also estimates the experimental standard deviation without any replicate runs.

Russell Lenth supplied an objective counterpart in Quick and Easy Analysis of Unreplicated Factorials, published in Technometrics in 1989. Lenth's method computes a pseudo standard error from the median of the absolute effects, trimming the largest out of the estimate so a few active factors do not inflate it, then compares each effect against a margin of error derived from it. It is the default significance test for unreplicated designs in most DOE software.

Model Adequacy and Confirmation

Every conclusion rests on assumptions about the residuals, and four plots test them. A normal probability plot of residuals checks normality. Residuals against fitted values check constant variance, where a funnel shape signals that the response needs a logarithmic or square-root scale. Residuals against run order check independence and reveal drift. Residuals against each factor check for unmodeled curvature.

Transformations matter in electronics data. Counts of defects per board are usually better analyzed after a square-root transformation, proportions such as barrel fill or yield after a logistic transformation, and strictly positive measurements whose spread grows with their mean, including etch rates and pull forces, after a logarithmic transformation. Analyzing a strongly skewed response on its raw scale inflates the influence of the largest observations and can invert the ranking of factors. The coefficient of determination also deserves skepticism, since it always rises when a term is added; the predicted version, computed by leaving each run out in turn, is the honest one, and a large gap between the two signals overfitting.

The final step is to run the process at the recommended settings and compare the result against the model's prediction interval. Confirmation is not a formality; it catches extrapolation beyond the experimental region, model misspecification, and process changes that occurred in between. A confirmed optimum is then handed to production, where the relevant characteristics go onto control charts and the experiment's estimate of process standard deviation becomes the starting point for the control limits.

Applications in Electronics Manufacturing

Five processes account for a large share of the designed experiments run in electronics plants, and each illustrates a different structural feature of experimental design.

Reflow Profile Optimization

The candidate factors are zone setpoints, conveyor speed, and atmosphere, which translate into the profile characteristics that govern the metallurgy: ramp rate, soak time and temperature, peak temperature, time above liquidus, and cooling rate. For the tin-silver-copper alloys that dominate lead-free assembly, liquidus sits near 217 degrees Celsius. The responses are void area fraction measured by X-ray, joint strength by shear or pull, wetting and fillet formation against the IPC acceptance criteria, the temperature difference across a populated board, and defect counts for tombstoning and head-in-pillow. The constraint bounding the design space at the top is the package moisture-sensitivity classification reflow profile, which sets the maximum peak body temperature a device is qualified to survive. IPC-7530A, Guidelines for Temperature Profiling for Mass Soldering Processes (Reflow and Wave), released in March 2017 as a revision of the 2001 original, supplies the profiling methodology.

The awkward part is randomization. An oven takes ten to twenty minutes to stabilize after a setpoint change, so a twenty-run design with full randomization spends hours waiting. The honest design is a split-plot: oven setpoints and belt speed as whole-plot factors changed a modest number of times in random order, with board-level factors such as paste type, stencil thickness, and pad design as sub-plot factors varied within each setting. Curvature is expected in peak temperature, so center points belong in the first design and a face-centered augmentation is the usual second stage.

Wave Solder Parameter Tuning

Wave soldering has more factors than reflow and noisier responses, which makes it a natural screening problem. The candidate list runs to a dozen: flux type, specific gravity, and deposition volume; preheat topside temperature; conveyor speed and angle; solder pot temperature; wave height and contact length; and nitrogen inerting. The responses are barrel fill against the vertical fill criteria of IPC-A-610, together with bridging counts, solder balling, icicling, and dewetting. A twelve-run Plackett-Burman or a sixteen-run resolution IV fraction is the standard first experiment, and in most plants it leaves preheat, conveyor speed, and flux volume.

The important structural feature is that the responses conflict directly: the settings that maximize barrel fill, namely high preheat and slow conveyor, also increase bridging and thermal stress. Wave soldering is therefore a textbook case for multi-response desirability optimization, where the deliverable is a window in which all responses are acceptable rather than a point at which one is optimal. Barrel fill and defect counts are proportions and counts respectively, so both should be transformed before analysis, and the machine drifts within a shift as dross accumulates. The underlying alloy and flux behavior is covered in Soldering Technologies and Materials.

Plasma Etch Recipe Development

Etch development has the highest cost per run of the five and the strongest case for statistical design. The factors are radio-frequency source power, bias power, chamber pressure, gas flows and their ratios, electrode gap, and substrate temperature. The responses are etch rate, across-wafer uniformity, selectivity to the mask and to the underlying film, sidewall profile angle, critical dimension bias, and aspect-ratio-dependent etch lag.

Three features shape the design. The interactions are strong and physically fundamental, since pressure and power together determine ion energy and ion density, so any design incapable of estimating two-factor interactions is unsuitable. Parts of the design space are physically inaccessible, because a plasma will not strike at the combination of lowest power, lowest pressure, and lowest flow; Box-Behnken designs are attractive here precisely because they never request an all-extreme combination. And chamber condition drifts with accumulated process time and resets after a wet clean, so blocking by chamber and by time since clean is standard. Such experiments sit inside the integration work described in Semiconductor Wafer Fabrication.

Wire Bond Strength

Wire bonding is a small-parameter, high-replication problem, the mirror image of etch development. The controllable factors are ultrasonic power, bond force, bond time, stage temperature, tool geometry, and, for ball bonding, the free air ball size. The primary response is bond strength, measured by destructive pull testing according to the bond strength method of MIL-STD-883 or by ball shear testing.

The instructive feature is that the numerical strength is not the whole response. Every pull test also produces a failure mode: lift at the ball bond, lift at the wedge, break at the neck, break at mid-span, or cratering of the pad dielectric beneath the bond. A high pull force accompanied by ball lifts is a worse outcome than a lower force accompanied by mid-span breaks, because a mid-span break means the bonds are stronger than the wire and the measured force is a property of the wire. Cratering is worse still, indicating that ultrasonic energy is damaging the structure underneath. A designed experiment therefore analyzes the force and the failure mode jointly; optimizing on force alone pushes ultrasonic power upward until pad damage surfaces in reliability testing months later.

The design structure is favorable in one respect and treacherous in another: hundreds of bonds can be made per parameter setting, so the response is precisely estimated, but those bonds are repeated measurements within one machine setup rather than replicates of it, and the setup-to-setup variation from tool wear and lead-frame lot changes is what production will experience. A correct design changes the setup more than once per condition. Bond development connects to the processes covered in IC Packaging and Assembly.

PCB Lamination

Lamination has the largest and most expensive experimental unit of the five and is the clearest case for split-plot design. The press factors are the temperature ramp, the pressure profile and the timing of its application, vacuum level, dwell time at temperature, and cooling rate. The material and construction factors are prepreg resin content and flow, glass style, copper weight and distribution, book construction, and panel position within the press opening. The responses span electrical, dimensional, and reliability domains: dielectric thickness and its uniformity, which governs controlled impedance; resin flow-out and starvation at high-copper areas; degree of cure by differential scanning calorimetry; layer-to-layer registration; warp and twist; and delamination resistance under thermal stress.

A press cycle takes hours and consumes a whole book of panels, so the number of whole plots is severely limited, often to six or eight for an entire program, and the press parameters are unavoidably whole-plot factors. Panel position within the book is a sub-plot factor, and also a blocking variable in its own right, because thermal gradients across a press opening are real and systematic. Treating each panel as an independent run is the standard error here, producing confident conclusions about press settings from what is in truth a handful of independent observations. The wider context appears in PCB Manufacturing Processes.

Summary

Design of experiments is the off-line counterpart to statistical process control: it perturbs a process deliberately to find and set the operating point, where SPC monitors the process in place to keep it there. Its central claim is that structured experimentation extracts more reliable information from fewer runs than the one-factor-at-a-time habit it replaces, resting on the ability to detect interactions and on the hidden replication that lets every run contribute to every effect estimate.

The design families form a progression rather than a menu. Screening designs, including resolution III fractions, Plackett-Burman arrays, and definitive screening designs, reduce a long factor list to a short one at the cost of heavy aliasing. Fractions of resolution IV and V characterize the survivors, with the resolution numbering stating which confusions the design accepts. Center points detect curvature. Response surface designs, principally central composite and Box-Behnken, fit the quadratic models needed to define an operating window. Mixture designs handle formulations whose components sum to a constant, and split-plot designs handle the case in which some factors cannot be reset between runs.

Taguchi's contribution divides cleanly. The quality loss function and the control-versus-noise robust design framework are sound and now mainstream, the key insight being that robustness arises from control-by-noise interactions. The signal-to-noise ratio analysis has not survived scrutiny, since Box showed in 1988 that the nominal-the-best criterion is justified only when the standard deviation is proportional to the mean. Current practice keeps the engineering idea and implements it through a combined array.

Analysis is a sequence rather than a single test: analysis of variance against a proper error term, graphical assessment through interaction and normal probability plots, an objective criterion such as Lenth's method for unreplicated designs, residual diagnostics, and confirmation runs at the recommended settings. Around the mathematics sit organizational habits that matter as much: state the objective as a question with a measurable answer, build the factor list with the operators who run the process daily, confirm the measurement system before committing to a run count, and document the run sheet as executed rather than as intended.

Four structural themes recur in electronics. Ovens, presses, and process chambers have long settling times, which makes split-plot design the norm. Responses conflict, which makes the search for an operating window more useful than a single optimum. Repeated measurements within a setup masquerade as replicates and must be separated from genuine run-to-run variation. And parts of the design space are physically inaccessible, which is why designs that avoid extreme corners earn their place.

Related Topics

The following articles cover the monitoring, improvement, and analysis practices that surround designed experimentation: