Latest / Hardware Startups with Fexingo: Devices, Robotics, and Manufacturing Tech Companies / How Hardware Startups Use Design of Experiments to Optimize Tolerances
Transcript
- Lucas: You know that moment when your first prototype works perfectly, and you breathe a sigh of relief — and then the next ten units all fail in slightly different ways? Luna: Oh, I can feel that. It's like the prototype was a fluke, and suddenly you're staring at a pile of boards that don't fit in the enclosure. Lucas: Exactly. And a lot of founders' first instinct is to just tighten every tolerance they can, make everything more precise. But that's expensive — tighter tolerances mean more costly machining, more rejects, longer lead times. Luna: And sometimes it doesn't even fix the problem, right? Because you're just guessing at which dimension matters. Lucas: Right. That's where a method called Design of Experiments — DOE for short — comes in. It's a structured way to test multiple variables at once so you can figure out which tolerances actually drive the failure, and by how much. Luna: I've heard of it in chemical engineering and pharma, but less in hardware startups. Is it practical for a small team? Lucas: Yeah, and that's exactly what we're going to get into today. We'll walk through a real scenario: say you're building a temperature sensor that goes inside a plastic enclosure. You're getting warpage — the lid doesn't seat right, seals fail. You've got five or six different dimensions that could be the culprit. Luna: So normally you'd test one at a time, change the wall thickness, run a batch, see if it helps, then change the rib height, another batch. That's slow. Lucas: And statistically inefficient. With a full factorial DOE — testing every combination of every variable — you'd need, say, two levels per variable. With six variables, that's two to the sixth power: 64 different combinations. That's a lot of prototype runs. Luna: But there's a smarter way, right? I remember something about fractional factorials. Lucas: Exactly. If you assume that higher-order interactions — three-way, four-way interactions between variables — are unlikely to matter, you can run a fraction of those combinations and still estimate the main effects and two-way interactions. A one-quarter fraction of that 64-run experiment is just 16 runs. Luna: So instead of 64 prototype batches, you do 16. That's a huge time and cost savings. Lucas: And in many cases, you can even go to an 8-run experiment if you're willing to alias some effects. The key is to design the experiment so that the effects you care about — the main effects — are not confounded with each other. Luna: So let's talk about the actual numbers. You mentioned a scenario with a temperature sensor enclosure. Did you have a specific example in mind? Lucas: Yeah, I was thinking of a startup I worked with a couple of years ago — they make industrial temperature sensors. Their enclosure was injection-molded ABS. They had a 15 percent reject rate at final assembly because the lid wouldn't snap shut evenly. They were about to redesign the whole mold, which would have cost fifty thousand dollars and delayed launch by three months. Luna: Ouch. So what did DOE show them? Lucas: They ran a 16-run fractional factorial experiment. They varied five factors: wall thickness, rib height, gate location, cooling time, and filler percentage. After measuring warp on each part, they found that only two factors mattered: filler percentage and cooling time. And there was an interaction — if you increased filler but kept cooling time short, warp actually got worse. Luna: So they could loosen the tolerances on wall thickness and rib height, which were expensive to control, and focus on those two. Lucas: Exactly. They tightened the spec on filler percentage and extended cooling time by 8 seconds per cycle. That increased cycle time by about 12 percent, but the reject rate dropped from 15 percent to under 2 percent. They saved the fifty-thousand-dollar mold rework, and the overall cost per unit went down because they weren't throwing away so many parts. Luna: That's a great example. And it's counterintuitive — sometimes loosening a tolerance can actually improve yield because you're not fighting an interaction you didn't understand. Lucas: Yeah, that's a really important point. Over-engineering tolerances is a trap. You think tighter is always better, but if you don't know which tolerances interact, you might be making things worse. DOE lets you find the 'sweet spot' — the cheapest combination that still works. Luna: It reminds me of something we talked about a few episodes back — design for assembly. The idea that you should design for the process, not for theoretical perfection. Lucas: Exactly. And DOE is a tool that makes that practical. You don't need to be a statistician. There are free online calculators, and most engineering software like Minitab or JMP has DOE modules. Even Excel add-ins can handle simple fractional factorials. Luna: But what about startups that are moving really fast — they're iterating on prototypes every week. Does DOE slow them down? Lucas: It can feel that way at first, because you have to plan before you build. But in the long run, it's faster. One DOE run can replace weeks of trial and error. The key is to integrate it into your prototyping cycle. When you're ordering a batch of prototype parts anyway, have the supplier vary the dimensions according to your DOE matrix. It costs almost nothing extra because the tooling is already set up. Luna: So you're basically getting the data for free. Lucas: Right. The incremental cost of varying a dimension in a CNC program or a 3D print file is zero. The only cost is measuring the parts afterward. And if you're already doing quality checks, that data is almost free too. Luna: That's a really practical angle. And it aligns with the lean startup idea of building to learn. You're not just building to get a product — you're building to understand your process. Lucas: Exactly. And that's the kind of thinking that separates hardware startups that scale from ones that get stuck in prototype hell. If these conversations are useful for what you're building or running, listener support is what keeps the show ad-free and focused on practical advice like this. If you find value in episodes like this one, you can support the show at buy me a coffee dot com slash fexingo. It's a simple way to say this matters, and it helps us keep bringing you real-case breakdowns without any sponsors. Luna: Yeah, and it genuinely makes a difference. We hear from listeners who say they've saved thousands by applying what they learned here. That's what keeps us going. Lucas: Alright, back to the topic. So once you've run your DOE and identified the critical factors, the next step is to set your tolerance specs based on the response surface. You can actually model the relationship and find the optimal operating window. Luna: That's essentially a regression model, right? You fit a polynomial to the data. Lucas: Yes. And with that model, you can predict yield for any combination of tolerances within the range you tested. So you can ask: what's the widest tolerance I can allow on wall thickness while still keeping yield above 98 percent? That's a powerful question for procurement. Luna: So you can negotiate with suppliers from a position of knowledge. You're not just asking for the tightest tolerance they can hold — you're asking for the tolerance that gives you the best cost-performance trade-off. Lucas: Exactly. And suppliers appreciate that because it means fewer rejected parts on their end too. It's a win-win. Luna: One thing I've wondered: does DOE work for electronics assembly? Like solder paste volume or reflow oven temperatures? Lucas: Absolutely. In fact, that's one of the classic applications. You can vary solder paste volume, reflow peak temperature, soak time, and cooling rate in a DOE to optimize solder joint quality. Many PCB assembly houses use DOE internally to set up their processes. But if you're a startup doing your own assembly, you can run a simple DOE to dial in your reflow profile. Luna: That's great. So DOE is really a versatile tool that applies across mechanical, thermal, and electrical domains. Lucas: Yes. And the common thread is that it forces you to think in terms of variables, not just parts. That mental shift alone is valuable for hardware founders. Luna: So if someone wants to start using DOE tomorrow, what's the first step? Lucas: Identify one recurring defect — a dimension that's out of spec, a fit issue, a performance variation. List the variables you suspect matter. Keep it to five or six. Then look up a fractional factorial design table — there are plenty online — and build the samples accordingly. Measure the response, analyze with a simple ANOVA or even just plot the averages. You'll be surprised how much you learn. Luna: I like that it's a systematic way to stop guessing. And for a startup, time is the scarcest resource. DOE saves time. Lucas: Exactly. It's one of those tools that seems academic until you actually use it once. After that, you wonder how you ever did without it.