Application of Machine Learning and Neural Networks for Generation of Pre-CMP Profiles of Advanced Deposition Processes for CMP Modeling

Abstract:
Due to the shrinking of both transistor size and the space between them, isolation of transistors from one another with high-quality dielectrics leads to the invention of new advanced deposition technologies such as flowable CVD (FCVD) and enhanced highaspect- ratio processes (eHARP) that allow void-free deposition in features with an aspect ratio of 12:1 or greater. The post-deposition profile on the patterned wafer is usually non-uniform. It may contain large variations that can affect on-surface planarity after chemical-mechanical polishing (CMP). Complicated pre-CMP profile height dependence on the underlying pattern is observed for spin-on dielectric (SOD), highdensity plasma CVD (HDP-CVD), FCVD, eHARP, and other processes. While it is possible to construct compact models for SOD and HDP-CVD processes, building compact models for FCVD and eHARP is challenging, due to complex nature of these processes. In this paper, we present the application of machine learning and neural networks to pre-CMP surface profile generation for SOD, HDP-CVD, FCVD, and eHARP processes. Neural networks are trained on measured line scan data and synthetic data generated by models using the Calibre CMP Model Builder tool. We show that two hidden layers are enough to capture the complexity of deposited surface profiles. The simulated surface profiles show good agreement with measurements. Keywords: chemical-mechanical polishing, CMP, CMP modeling, deposition process modeling, spin-on dielectric, HDP-CVD, flowable CVD, eHARP, machine learning, neutral networks