cd ~/
ManufacturingSensorsHardwareScaleMagic Leap

From Prototype to Mass Production: Sensor Manufacturing

The journey from a working sensor prototype to thousands of units rolling off a production line - yield, calibration, and quality at scale.

Evyatar Bluzer
3 min read

We have working sensors in the lab, and now we need to make 100,000 of them. Those turn out to be almost two different jobs.

Prototype vs Production Mindset

In prototyping you hand-tune calibration on best-in-class components, take unlimited debug time, and work with a sample size of 1. In production, calibration is automated at seconds per unit, components are cost-optimized, diagnosis has to happen in seconds, and every unit must work.

Yield: The Defining Metric

Yield = (Good units) / (Total units manufactured)

At $50 component cost and 80% yield, the effective cost is $50 / 0.8 = $62.50 per good unit, and the 20% scrap adds up to millions at scale. At 95% yield the same unit costs $50 / 0.95 = $52.63. That 15% yield improvement might be worth more than any other optimization available to us.

Cost per good unit versus yieldA line chart with yield from 70 to 100 percent on the horizontal axis and cost per good unit on the vertical axis; a curve falls from about 71 dollars to 50, with two highlighted labeled points at 80 percent and 95 percent yield and a dashed rule at the 50 dollar component cost.70%75%80%85%90%95%100%Yield (good units / units built)$50$55$60$65$70Cost per good unit, $50 of components$50 component cost80% yield: $62.50 per good unit95% yield: $52.63
Cost per good unit at a $50 component cost, computed as $50 divided by yield: $62.50 at 80% and $52.63 at 95%, the two yields worked through above.

Sources of Yield Loss

Component variation comes first: every resistor, capacitor, and optical element has tolerance. +/-1% is common for resistors, +/-5% on transmission is good for optical elements, and the distributions multiply, which is why tolerance analysis had to start with the lens design. Then assembly variation: pick-and-place placement is typically +/-50μm, die attach thermal interface quality varies, and contamination puts particles in the optical path. Process variation sits on top of all that - reflow profiles vary oven-to-oven, adhesive cure varies with batch, and temperature, humidity, and equipment drift never stop moving.

Design for Manufacturing (DFM)

A few choices we made for manufacturability. Wider tolerances where possible: if the algorithm can handle +/-10% depth error, don't specify +/-5% sensor calibration. Testability: every subsystem must be independently testable, because you can't debug what you can't probe. Redundancy: if one sensor fails calibration, can we reroute to a spare? And binary pass/fail criteria on the line, with no subjective decisions.

Calibration at Scale

Our lab calibration takes 30 minutes per unit. The production target is 30 seconds, which is the precision-versus-throughput gap I wrote about last year with a number on it. Getting there means calibrating multiple parameters in parallel, calibrating against known-good golden units rather than absolute standards, inferring some parameters statistically from others, and trimming in firmware what the hardware can't deliver.

What Happens When a Bad Unit Ships?

Field failure, customer frustration, returns, reputation damage. One field failure costs about 100x what a factory catch does, so we invest heavily in outgoing quality control: 100% functional test, sample environmental stress screening, and accelerated life testing on samples.

Learning from the Line

Production data is gold. Calibration parameters let us track component suppliers, test failures feed back into design, and customer returns expose escapes in testing. We've built dashboards tracking key metrics in real-time, so manufacturing problems become visible within hours instead of weeks.

Production ramp is Q3. We'll learn a lot more soon.

Comments