Domain Adaptation Based fin Height Measurement for DRAM

Fin Height is critical quality parameter of DRAM device and its measurement is destructive process. This measurement process has some disadvantages - Destroying wafer/chips, limitation on number of chips for measurement leads to less data samples, and its availability towards end of production cycle. Ability of system to predict fin height early/mid production cycle serves vital role to reduce redundancy and save losses in the production pipeline. With the advent of Spectroscopic methods like Ellipsometry (SE), Reflectometry, Interferometry, it is possible to capture spectral characteristics of devices in non-destructive framework. ML/DL modelling of spectrum data learns fin height characteristics and aids prediction. Simulation of DRAM for spectrum data enables more learning data, helps to enhance fin height prediction by not just data augmentation technique of simulated data to real data. We adopt and employ Domain Adaptation (DA) methods for regression problem where DA mostly dealt with classification problem. We demonstrate significant improvement in fin height prediction accuracy by using simulated data in addition to real data via various Domain Adaptation methods.