Candidate:
AJ
Malcolm
Title:
Multilevel
Random
Telegraph
Noise
Analysis
Using
Machine
Learning
Techniques
Date:
August
13,
2020
Time:
9:30
AM
Place:
Remote
attendance
Supervisor(s):
Kim,
Na
Young
Abstract:
Random
telegraph
noise
(RTN)
is
a
noise
process
which
occurs
in
solid-state
electrical
devices
such
as
metal-oxide-semiconductor
field-effect
transistors
(MOSFETs)
and
Josephson
Junctions.
Defects
in
the
crystal
structure
of
these
devices
trap
charge
carriers,
resulting
in
modulations
of
the
devices
electrical
transport
properties
such
as
mobility
or
threshold
voltage.
This
is
observed
as
sudden
transitions
between
two
discrete
current,
voltage,
or
resistance
levels,
corresponding
to
the
occupied/unoccupied
states
of
the
trap.
The
magnitude
of
a
trap’s
effects
can
be
linked
to
its
physical
location
in
devices,
and
so
we
sought
to
apply
RTN
analysis
during
the
characterization
of
devices
such
as
MOSFETs
to
learn
more
about
its
structure
and
any
trap
dependence
on
temperature
or
bias
levels.
However,
device
measurements
demonstrated
a
large
proportion
of
RTN
signals
with
more
than
two
levels,
which
strongly
suggests
the
presence
of
multiple
charge
defects.
This
scenario
is
commonly
avoided
in
published
research
on
RTN
analysis,
with
most
literature
focusing
on
measurements
showing
the
effects
of
only
a
single
trap.
The
frequency
with
which
multiple
traps
was
observed
in
our
measurements
motivated
the
development
of
an
algorithm
to
better
characterize
multi-level
RTN,
and
to
avoid
discarding
large
swathes
of
measurement
data.
The
developed
algorithm
applies
mixture
models
formed
with
Gaussians
to
identify,
isolate,
and
analyze
RTN
signals.
This
is
accomplished
through
the
use
of
machine
learning
techniques
to
maximize
the
likelihood
that
a
constrained
combination
of
these
models
describes
the
RTN
components
at
every
step.
When
multi-level
RTN
is
present,
it
is
further
decomposed
into
its
constituent
components
which
allows
characterization
of
each
independent
defect.
The
algorithm
is
applied
to
a
set
of
cryogenic
MOSFET
measurements
taken
from
near
cut-off
and
into
saturation,
which
demonstrates
the
ability
to
characterize
trap
count,
trap
amplitude,
and
state
occupation
distribution
for
each
trap.
Although
development
of
the
algorithm
has
precluded
an
in
depth
exploration
of
the
effects
of
parameters
such
as
temperature
on
these
traps,
that
type
of
analysis
could
not
have
been
achieved
to
a
suitable
level
of
accuracy
without
it.
Thursday, August 13, 2020 9:30 am
-
9:30 am
EDT (GMT -04:00)