Investigation on MBOPC convergence improvement with location-dependent correction factors aided by machine learning

Model-based optical proximity correction (MPOPC) has been well adopted in subwavelength lithography for integrated-circuit manufacturing. Typical MBOPC algorithms involve with iteratively moving the layout polygon edges to reduce the edge placement errors (EPEs) predicted by the lithography model. At each iteration, the amounts of movement are mainly determined by the values of the EPEs and the correction factors (CFs). Since full-chip lithography simulation is very computation intensive, it is highly desirable to minimize the number of iterations for acceptable run times, by selecting suitable CFs. In practical applications, the CFs are usually heuristically determined and applied globally throughout the correction regions. This approach efficiently reduces the EPEs at most of the target points but the entire convergence can be hampered at a relatively small number of hot-spot locations. This work investigates the effectiveness of improving the overall convergence by introducing both global and local CFs, and approaches to utilize machine-learning techniques to estimate the hot-spot locations and associated local CF values.