Optimization of Multi-Fins FinFET Implemented on SOI Wafer Based on SiGe and Gaussian Process Regression

Despite advancements in mitigating the short channel effect using high-k materials, multi-gate structures, and silicon-germanium (SiGe) alloys in three-dimensional FinFETs, performance trade-offs remain. This study introduces a novel machine learning framework utilizing a Gaussian process regression model (GPRM) and surrogate optimization (SO) to optimize design parameters of n-type and p-type SiGe FinFETs. With this approach targeting switching ratio (SR), the optimal mole fractions of the n-type FinFET are Si0.7Ge0.3 for source extension (
$S_{ext}$
), Si0.2Ge0.8 for channel (
$L_{g}$
), and Si for drain extension (
$D_{ext}$
), achieving an SR of
$6.9\times 10 ^{9}$
. For p-type FinFET, the optimal configuration is Si0.9Ge0.1 for
$S_{ext}$
, and Si for
$L_{g}$
and
$D_{ext}$
with an SR of
$5.81\times 10 ^{7}$
. The optimization of n-type and p-type multi-fin FinFETs (NmFinFET and PmFinFET) was also investigated, considering varied input parameters such as
$L_{g}$
, fin height (
$F_{h}$
), fin width (
$F_{w}$
),
$S_{ext}$
,
$D_{ext}$
, and the number of fins (numfin). The optimized devices for NmFinFET and PmFinFET, prioritizing speed, have the same dimensions:
$L_{g} = 10$
nm,
${F} _{h}$
=42 nm,
$F_{w} = 10$
nm,
$S_{ext} = 3$
nm,
$D_{ext} = 4$
nm, and numfin =5. An inverter constructed using these optimized parameters showed a simulated propagation delay of 2 ps. This machine learning-driven approach demonstrates remarkable effectiveness in optimizing FinFET designs. The framework’s ability to simultaneously optimize multiple objectives showcases its potential for advancing semiconductor device engineering.