Machine Learning-Assisted Statistical Variation Analysis of Ferroelectric Transistor: From Experimental Metrology to Adaptive Modeling

Abstract

A novel machine learning (ML)-assisted approach is proposed for investigating the variability of ferroelectric field-effect transistor (FeFET) to shorten the loop of technology pathfinding. To quantify the ferroelectric (FE) domain variation, the atomic intragranular misorientation of Si-doped HfO <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$_{\text{2}}$</tex-math> </inline-formula> thin film is measured by transmission Kikuchi diffraction (TKD) and is transformed into a polarization map. With the metrology data, polarization variation (PV) of FE domains on the gate-stack is modeled in technology computer-aided design (TCAD) to assess the impact of PV on the FeFET performance and to obtain datasets for ML-assisted analysis. A neural network model is trained using the datasets (input: polarization maps; output: high/low threshold voltage, ON-state current, and subthreshold slope) for the 28-nm bulk FeFET analysis. Our trained network, if used for inference to obtain three-sigma statistics, shows <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$&gt;$</tex-math> </inline-formula> 98% of accuracy of the device features and significantly faster simulation time than TCAD. In addition, we used the transfer learning technique to reduce the number of training datasets by 83% for the fully depleted silicon-on-insulator (FDSOI) FeFET by applying the pretrained model from the bulk FeFET.