Optimization of Power VDMOSFET's Process Parameters by Neural Networks

This paper presents the most important aspects of a process/device optimization technique based on neural network models. A reverse model neural network, trained to learn the inverse dynamic of semiconductor manufacturing processes effectively determines the optimum process conditions for realizing the desired device performances. Moreover, this optimization technique enables the optimization of a large number of process parameters, that is the major drawback of classical RSM optimization procedure. Furthermore, forward model neural networks could be introduced in a computer-integrated manufacturing system in a parallel with process and device simulators to save computational time. This is especially significant for the fast estimation of relationships between output characteristics (process and device design performance) and input process parameters. In order to demonstrate the efficiency and practicality of an TCAD system with integrated forward and reverse model neural networks a complex manufacturing process flow of low-voltage power VDMOSFET is optimized. Six input parameters (epi-layer concentration N
_epi
, body implantation doses Q
_p+
and Q
_p-
, body diffusion time t
_p
, source implantation dose Q
_N+
and diffusion time t
_N
) are optimized in order to obtain desired doping profile properties (channel length l
_ch
, boron concentration at the source end of the channel N
_ch
and depth of source junction x
_js
) and electrical device characteristics (threshold voltage V
_th
and breakdown voltage V
_b
). Results obtained by the forward and reverse model neural networks in predicting of the design performance is compared with the simulation results obtained by the process simulator MUSIC 2 and modified version of the device simulator MINIMOS 6.