Automated Construction of Semi-Physical CMP Models via Embedded Neural Networks

Abstract

The planarization of chip surfaces after chemical mechanical planarization (CMP) is becoming increasingly crucial as it can lead to problems such as depth of focus (DOF), voltage drop (IR drop), timing closure and electromigration (EM) problems. To enhance production yield, the industry requires an accurate CMP model to detect, localize, and control topography nonuniformity caused by layout dependent effects (LDE) prior to fabrication. However, existing semi-physical models heavily rely on manually specified empirical relationships during the calibration process, limiting their ability to meet the demands of advanced process nodes in terms of automated model construction and prediction accuracy. To address this limitation, we propose to construct empirical relationships in semi-physical models using embedded neural networks. Building upon this concept, we have developed a deep-learning-assisted semi-physical CMP model that eliminates the need for manual specification of empirical relationships. Experimentation conducted on silicon data from test chips across the process nodes of 28/32/40 nm highlights the advantages of our model, including rapid training (requiring fewer than 400 epochs), automated deployment and competitive prediction accuracy compared to data-driven models (RMSE reduction for dishing (18%/79%/55%) and erosion (25%/58%/61%) over traditional semi-physical models).