A part that fails in the field is rarely a mystery in hindsight. The drawing said one thing, the process made another, and nobody measured the interaction between them until a customer's line stopped. Design for Manufacturing (DFM), tolerance evaluation, Design of Experiments (DOE) and a clear form-fit-function view of the product are the disciplines that shorten the distance between "it failed" and "we know why, we fixed it, and we told the customer what we did."
This article walks through how these practices fit together, how to evaluate tolerances before they reach the shop floor, how to use DOE to separate real causes from coincidences, and how to report failure root causes to a customer in a form they can act on. It ends with a small Java example for tolerance stack-up and DOE run generation.
Design for Manufacturing starts before the drawing is final
DFM is the practice of checking whether a design can be made reliably, at the target cost, with the processes you actually have. It is cheapest when it happens early, because every later change multiplies through tooling, fixtures, qualification and documentation.
A useful DFM review asks specific questions:
- Can each feature be produced by the chosen process without special handling? A deep narrow pocket may need a tool the shop does not own.
- Are tolerances tighter than the process can hold? A feature toleranced at ±0.01 mm on a casting that varies by ±0.2 mm is a design that will always fail inspection.
- Can the part be fixtured and measured repeatably? Datums that cannot be reached by a probe make inspection a guessing game.
- Does the assembly sequence allow tools and operators to reach every fastener? Blind joints are a common source of incomplete assemblies.
- Are the material and surface finish compatible with the environment the product will see? Many field failures are corrosion or wear problems that the design could have avoided.
DFM reviews work best when manufacturing engineers, quality and the designer look at the same drawing together, with the process capability data open on the table.
Tolerance evaluation: worst case versus statistics
Every dimension has a tolerance, and every assembly stacks several of them. Tolerance evaluation estimates how much the final dimension can vary when the parts are combined.
There are two common approaches:
Worst case. Add the tolerances linearly, assuming every part is at its extreme in the same direction simultaneously. This guarantees the assembly will fit, but it often forces tight, expensive tolerances on every part.
Statistical (root sum square, RSS). Assume the parts vary independently around their nominal values, and combine the tolerances as the square root of the sum of their squares. This gives a realistic spread for a large production run, but it carries a risk: if the process is not centered or the parts are correlated, the actual spread can exceed the estimate.
Use worst case when the function is safety critical, when volumes are low, or when you have no process data. Use statistical analysis when you have capability data (Cp and Cpk) that show each process is stable and centered. Either way, document the assumptions, because the assumptions are what fail first.
Design of Experiments: finding the cause, not just a correlation
When a failure occurs, the natural instinct is to change the thing that looks most suspicious. That approach fixes problems by accident and misses interactions. DOE is a structured method for changing several factors at once, in planned combinations, and measuring how the response changes.
A basic workflow:
- Define the response. Choose a measurable output that represents the failure: force to failure, leak rate, bond strength, dimensional drift.
- Select the factors. List the variables that might drive the response: temperature, dwell time, fixture pressure, material lot, tool wear.
- Choose levels. Pick two or three realistic values for each factor, usually the range the process actually sees.
- Run a designed experiment. A full factorial design tests every combination. With three factors at two levels, that is eight runs. Fractional factorial designs reduce the count when there are many factors.
- Analyze the effects. Estimate the main effect of each factor and the interactions between them. Statistical significance tells you whether a difference is likely real.
- Confirm. Run the predicted optimum, and verify that the response meets the requirement across a reasonable range of conditions.
A DOE that is designed well answers questions that one-factor-at-a-time testing cannot. For example, a glue joint may fail only when cure temperature is high and surface energy is low. Neither factor alone shows the problem.
Form, fit and function as the failure lens
Form, fit and function (FFF) is a way to describe what a part must do, so that failure analysis asks the right question.
- Form is the physical shape, size and appearance the design requires.
- Fit is how the part interacts with mating parts: clearances, interferences, alignment and assembly loads.
- Function is what the part must accomplish: carry a load, seal a volume, transmit a signal, dissipate heat.
When a part fails, classify the failure against FFF. A form failure means the part is out of dimension or deformed. A fit failure means it does not assemble or it assembles with the wrong preload. A function failure means it assembles correctly but does not perform. These categories sound simple, but they separate problems that look alike. A cracked housing might be a form problem (a sharp corner that concentrates stress), a fit problem (an interference that induces stress during assembly) or a function problem (a material that does not meet the fatigue requirement). Each leads to a different corrective action.
Root cause analysis that holds up
Good root cause analysis is disciplined about evidence. Common tools include:
- Five whys, useful for simple chains but dangerous when the first answer is a guess.
- Fishbone (Ishikawa) diagrams, which organize possible causes under categories such as method, machine, material, measurement, environment and people.
- Fault tree analysis, which models how combinations of events produce the top-level failure.
- Physical analysis, including sectioning, microscopy, fractography and dimensional measurement of the failed part and of good parts for comparison.
The most common mistake is to stop at the first plausible cause. A root cause is confirmed when you can reproduce the failure by introducing that cause, and when removing it prevents recurrence. DOE is often the bridge between a plausible hypothesis and that proof.
Reporting failure root cause to the customer
Customers care about four things: what failed, why it failed, what you did about it, and how they can be sure it will not happen again. An 8D report (Eight Disciplines) is a widely used structure that answers those questions in sequence:
- Form the team and define the problem clearly.
- Contain the problem so affected product is identified and quarantined.
- Define and verify the root cause.
- Choose and verify permanent corrective actions.
- Implement the permanent corrections.
- Prevent recurrence through system changes.
- Recognize the team and close the report.
Several details make a report credible. State the failure in measurable terms, not adjectives: "bond strength below 12 N in 3 of 40 units" is more useful than "weak joints." Show the evidence: photographs, measurement data, DOE results. Separate what you know from what you suspect. Quantify containment, including units shipped, units in stock and units in transit. And give the customer a date for verification, not just a promise of improvement.
A good report also closes the loop with the design. If the root cause was a tolerance that was never evaluated, the report should say that the stack-up was added to the design review checklist. Customers trust suppliers who change the system, not only the part.
A small Java model: tolerance stack and DOE run generation
The math behind tolerance stacks and the bookkeeping behind DOE are both simple enough to automate, which makes them easy to review and test. The example below has two parts. The first calculates worst-case and RSS tolerance for a linear stack. The second generates every run of a full factorial design. It uses Java 17 and no external libraries:
import java.util.ArrayList; import java.util.LinkedHashMap; import java.util.List; import java.util.Map; public class DfmToolkit { public record Dimension(String name, double nominal, double tolerance) {} public record StackResult(double nominal, double worstCaseMin, double worstCaseMax, double rssTolerance) {} /** * Linear stack-up where every dimension contributes in the same direction. * Worst case adds tolerances; RSS combines them statistically. */ public static StackResult linearStack(List<Dimension> dimensions) { double nominal = 0; double worstCaseTolerance = 0; double sumOfSquares = 0; for (Dimension d : dimensions) { nominal += d.nominal(); worstCaseTolerance += d.tolerance(); sumOfSquares += d.tolerance() * d.tolerance(); } return new StackResult( nominal, nominal - worstCaseTolerance, nominal + worstCaseTolerance, Math.sqrt(sumOfSquares)); } /** Generates every combination of factor levels (a full factorial design). */ public static List<Map<String, String>> fullFactorial(Map<String, List<String>> factors) { List<Map<String, String>> runs = new ArrayList<>(); runs.add(new LinkedHashMap<>()); for (Map.Entry<String, List<String>> factor : factors.entrySet()) { List<Map<String, String>> expanded = new ArrayList<>(); for (Map<String, String> run : runs) { for (String level : factor.getValue()) { Map<String, String> next = new LinkedHashMap<>(run); next.put(factor.getKey(), level); expanded.add(next); } } runs = expanded; } return runs; } public static void main(String[] args) { // Three stacked lengths in one direction, in millimeters. List<Dimension> stack = List.of( new Dimension("housing_wall", 10.0, 0.05), new Dimension("spacer", 8.0, 0.03), new Dimension("cover_lip", 7.0, 0.04)); StackResult result = linearStack(stack); System.out.printf("Nominal: %.2f mm%n", result.nominal()); System.out.printf("Worst case: %.2f to %.2f mm%n", result.worstCaseMin(), result.worstCaseMax()); System.out.printf("RSS tolerance: +/-%.3f mm%n", result.rssTolerance()); // Three factors: two cure temperatures, two dwell times, three fixture pressures. Map<String, List<String>> factors = new LinkedHashMap<>(); factors.put("cure_temp_C", List.of("120", "140")); factors.put("dwell_min", List.of("30", "45")); factors.put("pressure_kPa", List.of("200", "250", "300")); List<Map<String, String>> runs = fullFactorial(factors); System.out.println("Full factorial runs: " + runs.size()); System.out.println("First run: " + runs.get(0)); } } Running this prints a nominal of 25.00 mm, a worst-case range of 24.88 to 25.12 mm, an RSS tolerance of ±0.071 mm, and 12 factorial runs, starting with 120 °C, 30 minutes and 200 kPa. The worst-case range is wider than the RSS band by design: it guarantees fit, while the RSS figure predicts typical spread if the process is stable and centered.
Two notes on the model. The stack-up assumes all dimensions contribute in the same direction; real assemblies need signed contributions, and a vector-loop method handles that. The DOE generator produces the plan, not the analysis. Its output should feed a test matrix with randomized run order, because a run order that follows the factor list lets time-dependent drift masquerade as a factor effect.
Practical recommendations
Bring tolerance analysis into the design review. Run a stack-up for every critical assembly before release, and record the method and assumptions.
Check process capability before trusting statistics. RSS estimates are only as good as the Cpk values behind them.
Use DOE before changing parts. Test the process factors first. Many field problems are fixed by a process window, not a redesign.
Classify failures against form, fit and function. The category shapes the investigation and the corrective action.
Write the customer report as a system change. Show what you changed in design, process and inspection, not only in the failed part.
Conclusion and key takeaways
Reliable products come from designs that can be made, tolerances that can be held, experiments that separate causes from correlations, and reports that customers can verify. DFM catches manufacturing risk early. Tolerance evaluation shows whether an assembly can fit under the spread of real production. DOE finds causes that one-factor-at-a-time testing misses. Form-fit-function gives failure analysis a common language. And an 8D-style report turns a field problem into evidence that the system has improved.
Key takeaways:
- DFM is cheapest before the drawing is frozen, so review it early with manufacturing and quality present.
- Worst-case stack-ups guarantee fit; RSS estimates typical spread, but only with stable, centered processes.
- DOE tests several factors at once and reveals interactions that drive real failures.
- Classifying a failure as form, fit or function points the investigation at the right evidence.
- Customer reports earn trust when they quantify containment, show evidence and change the system.
If your team handles field failures today, start with the next one: write the failure in measurable terms, classify it against form, fit and function, and check whether your tolerance stack for that assembly was ever calculated.


![Physical AI Foundations and Definitions — Ch.10: Future Directions [EN]](https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fgz7t8gzrdba0obh9ma56.png)
![Physical AI Foundations and Definitions — Ch.08: Physical AI Product Strategy [EN]](https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F4gyormoamcida6t18mdw.png)
![Physical AI Foundations and Definitions — Ch.07: Relation to AMR Manipulator Humanoid Quadruped UAV [EN]](https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fte7aqbjpqzqiay6cht0x.png)
![Physical AI Foundations and Definitions — Ch.05: Physical AI Taxonomy [EN]](https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fxeb66tbhskewzsz49i7p.png)