One practical ILT tuning that consistently improved yield for me without blowing up mask write time was aggressively constraining curvilinear complexity early, rather than trying to smooth it later. The single most impactful setting in my flow was a tight shot count cap enforced during ILT optimization, paired with controlled curvilinear smoothing. Early on, I let the solver chase maximum image fidelity, which produced beautiful contours but an explosion in shot count and fractured edges that were painful for mask write and inspection. Yield did not improve enough to justify the cost. What worked better was setting a realistic shot count ceiling up front and forcing the solver to optimize within that budget. That constraint pushed the ILT engine to make smarter tradeoffs instead of brute force corrections. Curvilinear smoothing mattered, but only when applied with intent. I found that moderate smoothing with minimum radius constraints preserved process window while dramatically reducing small edge perturbations that do not print reliably at high NA anyway. Over smoothing hurt CD control, but light, geometry aware smoothing removed features that only added mask complexity, not silicon value. Another key adjustment was weighting hotspots by process sensitivity rather than pure EPE. By prioritizing regions with poor NILS or tighter focus exposure margins, I could relax ILT effort in less critical areas. That reduced overall pattern complexity while improving yield where it actually mattered. The biggest lesson for me was that ILT is as much about constraint design as it is about optimization power. When the solver understands your manufacturing limits from the start, you get patterns that print better, write faster, and inspect cleaner. That balance made the difference between an impressive layout and a manufacturable one.
The tuning that helped most was putting a hard "manufacturability leash" on ILT output. We tightened curvilinear smoothing and enforced a shot-count cap early, then only let ILT get aggressive in known hotspot windows instead of the full chip. That preserved the yield benefit where it mattered, without exploding mask write time. The single setting that mattered most in our flow was the shot count cap paired with curvature regularization. Once those were locked, everything else became incremental instead of chaotic.