Optimization of dummy pattern for mask data size reduction

Abstract

Dummy ("waffling") features on masks improve pattern density on wafers, helping chemical-mechanical polishing (CMP) achieve uniform thickness of residual oxide. However, getting the required pattern density may substantially complicate mask creation algorithms, e.g., for metal lines at loose pitch not matching the waffling grid. Without proper optimization, multiple waffling passes with decreasing waffle sizes increase design database and mask writing time, not always achieving the planarization goal. In this work, we investigated the effect of waffle type, size, spacing, and number of waffling runs (passes) on die database with variable line density and pitch. We correlated the extracted pattern density with the simulated residual oxide thickness and proposed to use partial waffles created from the large size, original on-grid waffles, by subtracting the sized geometries of drawn lines running over them. We found that, surprisingly, multiple waffling passes do not significantly impact neither pattern density, nor the database file size; they are also not efficient in improving die planarity. In contrast, routines using partial waffles would produce very uniform, high density pattern, resulting in superb wafer planarity confirmed by the CMP simulation. However, this was accomplished at the expense of a threefold growth of the database size, as compared to the standard waffling routines, or about 30 time increase compared to the original drawn design. We then proposed an alternative approach using a "waffle add" CAD layer to achieve a significant reduction in waffle count and parasitic capacitance.