One of the fundamental challenges with using COTS parts is that the manufacturer can change their process, packaging, or foundry without telling you and that change can dramatically alter the radiation response. You tested lot A and it survived 50 krad. Your flight build uses lot B from two years later, and it fails at 15 krad. This is not a theoretical concern; it has happened repeatedly.
Documented examples of significant lot variability:
- Linear voltage regulators are notorious for lot-to-lot TID variation. Multiple studies have shown 3-5× differences in failure dose between lots of the same part number from the same manufacturer.
- Some COTS ADCs have shown lot-dependent SEL thresholds, with some lots being SEL-immune and others having LET thresholds in the 20-30 MeV-cm²/mg range.
- Optocoupler CTR degradation under displacement damage can vary by factors of 2-4 between production lots due to LED efficiency differences.
Root causes:
- Foundry migration (same part number, different fab)
- Process node shrinks or optimizations
- Changes in epitaxial layer thickness
- Different passivation materials or oxide growth conditions
- Wafer-to-wafer variation within a single lot
- Die revisions that don’t change the part number
What can you do about it?
- Lot Acceptance Testing (LAT): Test samples from each flight lot. This is the gold standard but adds cost and schedule. How many samples per lot is enough? TM1019 suggests 10 for TID; is that sufficient for statistical confidence?
- Date code tracking: Maintain records of which date codes you’ve characterized. At minimum, you know when your data might be stale.
- Manufacturer engagement: Some manufacturers will disclose process changes under NDA. Build those relationships.
- Design margin: If you can’t guarantee lot consistency, your design margins need to account for the variability. This argues for larger RDM on COTS parts.
How does your program handle lot variability? Do you do LAT on every flight lot? What’s your threshold for deciding a part needs lot-level screening vs. accepting characterization data as sufficient?