OPC / ILT / Machine Learning – One Page Interview Summary
1. What OPC Really Does
- OPC compensates for lithography distortions between mask and wafer
- Goal: target wafer contour after lithography
OPC = predicting wafer patterns and modifying the mask accordingly.
- Rule-based OPC: fast, empirical geometric corrections
- Model-based OPC: physics-based (Hopkins + resist model)
---
2. Rule-based vs Model-based OPC
- Rule-based: bias, serif, hammerhead (no simulation)
- Model-based: simulate → compare → iterate
Production OPC always uses a hybrid flow:
rule-based for speed, model-based for accuracy.
---
3. What ILT Is (and Why It’s Different)
- ILT formulates mask synthesis as a global optimization problem
- Not edge tweaking, but mask-level optimization
ILT asks: “What mask gives the best wafer image?”
- Highest accuracy and process window
- Extremely high computational and mask cost
---
4. Why ASML / TSMC Use Partial ILT
- Full-chip ILT → weeks of runtime
- ILT masks → curvilinear, hard to write & inspect
- Most patterns already work with model-based OPC
Partial ILT = apply ILT only to lithography hotspots.
“ILT everywhere is unnecessary. ILT where it matters is optimal.”
---
5. Why ML Is Added to OPC / ILT
- Physics-based OPC is accurate but slow
- ML provides fast pattern recognition and prediction
ML does not replace physics — it accelerates it.
- Hotspot detection
- Initial guess for ILT
- Runtime reduction (5×–20×)
---
6. ML + ILT in Production
- ML identifies difficult patterns
- ILT refines only those regions
- Physics-based verification remains mandatory
Future OPC = Physics-based models + ML acceleration.
---
7. Interview-Ready Answer (Memorize This)
“Modern OPC combines rule-based and model-based approaches.
ILT provides the best pattern fidelity but is too expensive for full-chip use,
so companies like ASML and TSMC apply partial ILT to critical hotspots.
Machine learning accelerates this flow by enabling fast hotspot detection
and near-optimal initial mask solutions, while physics-based models
remain essential for accuracy and sign-off.”
---
Final Takeaway
- Rule-based OPC → speed
- Model-based OPC → accuracy
- ILT → maximum fidelity (selective)
- ML → scalability
This is not theory — this is how production OPC is actually done.
↩ Go to OPC / RET Technical Notes Index
↩ Go to Home – Tech Notes Park