Sub-100μs Laser Control for LPBF Defect Prevention
OverviewAnalysisSolutions
Complete
·Feb 2, 2026
The Core Insight

You don't need to classify defects to correct them

  • The industry frames this as 'detect defect type → select appropriate correction.' But keyhole and LOF are opposite boundaries of the SAME energy density window.
  • If brightness is above target → reduce power (regardless of whether it's 'keyhole' or not).
  • If brightness is below target → increase power.
  • This transforms a classification problem (50ms ML inference) into a tracking problem (analog PID in <100μs).
Viability
Solvable with Effort
  • The physics supports <10ms closed-loop control; no fundamental barriers exist.
  • The gap is architectural—digital processing can't meet timing requirements, but analog alternatives exist.
Key Decision

If you prioritize immediate results with zero hardware risk, start with scan strategies (concept 6). If you're building toward production-grade sensing-based control, validate the stability tracking paradigm first (concept 5) via back-reflection characterization (concept 2).

Solution Paths
01NEEDS VALIDATION

Self-Correcting Scan Strategies with Built-In Redundancy

Pure software change using overlapping remelting passes; blocked by unknown healing rate for deep keyhole porosity; trades 30-50% build time for potentially 10x defect reduction

02NEEDS VALIDATION

Stability Tracking via Analog Feedback

Paradigm shift: track stability instead of classifying defects; blocked by need to validate brightness-to-process-state mapping across materials; enables 1000x faster response than ML

Recommendation
  1. If this were my project, I'd start building test coupons next week with self-correcting scan strategies—50% overlap on bulk regions, standard scan on thin walls.
  2. Zero hardware investment, proven EBM physics, and I'd have CT results in 6 weeks showing whether I can cut defect rates in half through pure software.
  3. That buys time while the sensing work proceeds.
  4. In parallel, I'd get a back-reflection photodiode and beam splitter on a research machine within 30 days.
  5. The $20K hardware investment answers the most important question: does brightness deviation direction correlate with defect type? If yes, the entire analog control path opens up.
  6. If no, I know early and can pivot to ratio pyrometry or multi-modal fusion.
  7. I'd resist the temptation to jump straight to the ASML-style two-loop architecture.
  8. It's the right production solution, but it's 12-18 months and $200-500K.
  9. The simpler approaches—scan strategies + back-reflection threshold—could achieve 50-80% defect reduction in 6 months for under $100K.
  10. Prove value fast, then scale sophistication.
  11. The one thing I'd do differently than most AM teams: I'd hire an analog control engineer, not another ML researcher.
  12. The industry has been throwing GPUs at a problem that's fundamentally about control latency.
  13. A good analog designer with optical disc or arc welding background would see solutions that the software-centric AM community has missed for 15 years.

This site uses cookies to improve your experience.