Deep reinforcement learning assisted reticle floorplanning with rectilinear polygon modules for multiple-project wafer

Abstract

In multiple-project wafer (MPW), the cost of mask tooling can be significantly reduced by optimizing the floorplan of the reticle. However, the generation of an ideal reticle floorplan can be very challenging due to the unusual constraints and high complexity in the placement of the dies. In this paper, we propose a low-complexity method to extract the placement of dies in a reticle. The proposed algorithms are capable of placing die modules in rectilinear polygon shapes and achieving a high reticle area usage. The aspect ratio of the reticle is also optimized for fabrication consistency. We first design a heuristic method to decide the placement of all modules with given width and height of the reticle. Then based on this method, we propose a searching framework to minimize the area of the reticle while keeping its aspect ratio close to 1. Finally a deep reinforcement learning based recursive neural network (RNN) model is developed to optimize the sequence of modules so that the area usage of the reticle can be further improved. The experimental results demonstrate an up to 88.61% area usage obtained solely from our heuristic method, and an additional 9.15% usage is improved after implementing the RNN model.