Choose Language Hide Translation Bar

Plasma Enhanced Chemical Vapor Deposition Thin Film Development using DOE - (2023-US-30MP-1516)

Plasma enhanced chemical vapor deposition (PECVD) is extensively used to deposit thin films in semiconductor manufacturing. PECVD processes are gas phase processes typically conducted at low pressures in specially designed deposition chambers. Many process parameters influence the thin film properties in ways that are difficult to characterize.

 

This presentation shows:

  • How the Generalized Regression platform was used to "rescue" a definitive screening design that had runs that couldn’t be completed.
  • How the original design was augmented.
  • How this complex design space was finally understood well enough to find a new precursor, effectively identifying a thin film that maximized the targeted film property (high mechanical strength) in the fewest number of experiments.

 

This presentation is predominantly hands-on, using JMP in real time.

 

 

Today  we 're  going  to  talk  about  how to  use  JMP  to  do  thin  film  development

using  plasma -enhanced chemical  vapor  deposition .

This  is  a  bit  different ,  I  think ,

than  some  of  the  typical  work that 's  done  in  industry ,

where  it 's  a  continuously  stirred  reactor and  you  can  always  mix  things .

Regardless  of  what  happens , you  can  always  measure  outputs .

But in  plasma-enhanced chemical  vapor  deposition ,

it 's  really  discrete  pockets  of  stability that  you  have  to  work  with .

Even  though  we  can  set  up a large  parameter  space ,

there  can  be  spots within  that  parameter  space

where  you  may  not  be  able to  strike  the  plasma

or  it  could  arc because  the  power  density  is  too  high .

Since  we  have  a  large  number of  deposition  parameters ,

we  need  to  use  a  design  of  experiments

to  effectively  explore that  parameter  space .

Even  if  we 're  able to  strike  the  plasma ,

there  are  still  issues with  thin  film  uniformity .

We 're  depositing  nanometer  films

with  nearly  perfect  uniformity across  a  12 -inch  wafer .

Once  we  get  that ,  we  still  have to  hit  the  targeted  film  properties .

We 're  going  to  talk  about   how  to  use  PECVD

to  develop  new  thin  films from  new  precursors .

The  first  thing  we 're  going  to  do is  talk  about  Precursor  1 .

From  what  I  was  able  to  read from  the  JMP  tutorials ,

the  Definitive  Screening  Design is  a  very  effective  way

to  screen  a  large  number  of  main  effects in  the  fewest  number  of  experiments .

That 's  key  to  the  work  that  we  do.

We  want  to  get  the  right  answer in  the  shortest  amount  of  time

with  a  data -driven  approach .

We  used  a  Definitive  Screening  DOE to  explore  seven  factors  in  26  runs .

What  we 'll  do  is  just  open  up that  initial  DOE .

This  is  the  setup  we  came for  the  Definitive  Screening  DOE .

Here  are  the  seven  different  factors we're  varying  for  the  deposition,

and  our  output  is  going to  be  this  parameter  Y,

and  we 're  trying  to  maximize  it .

If  we  look  at  the  range  of  parameters for  this  type  of  PECVD  processing ,

this  is  a  very  wide  range of  initial  parameters .

Again ,  we 're  trying  to  screen for  main  effects ,

and  our  outputs  are  ranging  from,  say , 9-34 ,  and  our  baseline  was  21 .

We  do  see  an  improvement  there .

One  of  the  things  that  I  always  like to  do  when  we  do  a  DOE

is  include  a  center  point  replicate   or  a  repeat  run

to  see  how  reproducible   the  instrument  is , as  well  as  to  make  sure

that  the  statistics  we  generate within  the  design  are  valid .

These  are  the  two  center -point  runs ,

and  you  can  see  we  get excellent  reproducibility .

The  other  thing  that 's  really  nice  for  us to  do  before  we  get  into  fitting  the  model

is  just  to  look  at  the  output  variables

and  try  to  identify any  trends  that  we  can  see .

Is  there  anything  we  can  identify  quickly that  we  can  attribute  the  main  factors  to ?

Here,  there 's  four  points  with  a  Y  value of  greater  than  30  within  the  DOE .

If  we  select  those  points ,

it 's  nice  to  see  if  we  can  see any  trends  associated  with  these  data .

One  of  the  fastest  ways  I  found  to  do  that is  to  quickly  do  a  multivariate  analysis ,

and  we  can  do  this  graphically .

What  we 're  going  to  do is  take  all  our  factors

and  then  our  output  variable ,

and  we 're  going  to  generate a  multivariate  analysis .

Here  in  this  graph , this  is  our  Y  value .

You  can  see  as  we  go  from  10 to  30 ,

the  four  values that  are  the  highest  are  highlighted ,

but  the  rows  are  the  various  factors .

Here  we  can  see  for  helium ,

we  have  the  highest  values at  the  high  and  low  splits .

For  precursor ,  we  have the  high  and  low  splits .

But for  temperature  and  pressure ,

we  have  the  highest  values at  the  lowest  splits .

It 's  really ,  I  think ,

a  good  indication  initially before  we 've  done  any  model  fitting ,

that  temperature  and  pressure  could  be important  variables  for  us  to  look  at .

If  we  go  back  to  the  table,

the  other  nice  thing  about  JMP , it 's  very  powerful  because...

Again before  we  do the  definitive  screening ,

we  can  use  a  predictor  screening

to  identify  what  are the  most  important  factors .

Again ,  we  use  the  standard  analysis , input  our  factors ,  our  response  is  Y ,

and  you  can  see what  the  predictor  screening  is  telling  us

is ,  yes ,  pressure and  temperature  are  very  important .

But one  thing  that  we  didn 't  catch

in  that  multivariate  analysis is  the  precursor  flow .

These   three  factors ,  pressure , precursor  flow ,  and  temperature

appear  to  be  dominant  in  giving  us the  highest  values  of  Y .

Now  I  wanted  to  fit  the  model   because  I  think

the  real  power  of  the  DOE is  not  the  runs  in  the  table ,

but  it 's  the  response  surface  model that  you  can  use

to  get  predictions  for  improvement as  well  as  directions  to  further  explore .

But when  I  went  to  analyze  it , it  wouldn 't  work .

It  turns  out  we  were  only  able to  complete  25  of  the  26  runs ,

and  I  was  not  aware that  the  Definitive  Screening  DOE ,

the  default  analysis  would  not  work if  you  did  not  complete  all  of  the  runs .

At  this  point ,  I  contacted  Jed  at  JMP to  help  me  understand

how  I  could  get  some  models   out  of  this  data

that  we  carefully  collected over  a  period  of  time .

I 'll  turn  it  over  to  Jed .

When  Bill  called ,

like  he  said ,

when  he  hit  that  script that  saved  to  the  data  table

of  Fit  Definitive  Screening , nothing  happens .

If  you  look  over  here ,

the  log  is  saying ,  there  are  runs that  are  not  fold  over  center  point  run ,

and  it 's  run  17 ,  which  is  obviously   the  run  that  was  missing

and  couldn 't  be  completed in  the  experiment .

What  Bill  wanted  was  a  way  to  still fit  that  Definitive  Screening  model .

We  came  up  with  two  different  approaches

and  had  three  models  that  came  about from  those  two  different  approaches .

The  first  one  is  related to  the  Definitive  Screening .

The  designs  of  these types  of  models,  of  these  experiments,

are  always  fold -over  pairs , where  there 's  a  pair  of  opposites .

If  we  can  find  that  fold -over  pair , or  the  twin ,  I  guess ,  of  this  row  17 ,

we  should  be  able  to  exclude  both  rows

and  then  fit the  definitive  screening  design .

We  just  needed  a  simple  way  to  do  that .

What  we  came  up  with  was  basically   to  use  a  couple  of  shortcuts .

I 'm  going  to  first  standardize the  attributes  of  these  columns

and  change  the  modeling  type  to  ordinal .

As  I  do  this ,  you 'll  notice that  my  ability  to  select  has  changed .

That  helps  when  I  look  at  a  data  filter .

Now  I  have  boxes  rather  than  histograms , so  it  just  makes  it  faster  to  select .

What  we  need  to  do  is  find the  opposite  row  of  this .

I  have  this  row  17  selected , and  you  see  that  it 's  high ,  high ,  high ,

low ,  low ,  low ,  and  then I 'm  out  of  memory  space .

I 'm  remembering  high , high ,  high ,  low ,  low ,  low .

I  need  to  find  the  opposite  of  that ,

which  is  going  to  be  low , low ,  low ,  high ,  high ,  high .

If  I  just  come  over  here  and  start  working my  way  down ,  low ,  low ,  low ,  high ,

by  the  time  I  get to  just  four  of  the  runs

so  just  more  than  half of  the  factors  selected ,

now  I 'm  down  to  just  one  matching  row .

It  just  so  happens  that  the  very  next  run was  the  fold-over  pair  in  this  experiment .

We  can  select  both  of  those  runs , exclude  them ,

and  then  go  back  in into  the  column  properties

and  change  that  modeling  type back  to  continuous .

Now  when  we  hit that  Definitive  Screening  button ,

it  works .

We  can  run  that  model and  see  that  it 's  predicting  fairly  well .

We  can  see  the  profiler , but  we  also  were  really  aware

that  one  of  the  runs out  of  26 ,  that 's  almost  4 %.

We 're  throwing  away  roughly  4 % of  the  information  by  excluding  this .

What  we 'd  really  like  to  do is  not  throw  that  information  away ,

find  a  way  to  use  that .

We  used  the   Model Screening  platform in  JMP  Pro  to  run  a  bunch  of  models

and  then  select  the  best .

The  two  that  came  out  the  best were  a  Neural  and  Stepwise  model ,

and  I  can  walk  through  those really  quickly .

The  Neural  model  was  with  our  response and  our  factors .

Since  this  was  a  DOE , we 're  going  to  do  the  minimum  holdback ,

and  I 'm  just  going  to  choose a  random  seed  so  this  is  repeatable .

The   Model Screening  platform generally  suggests  about  20  boosts .

If  I  hit  Go  here , I  get  a  pretty  good   Rsquare  across  this.

Maybe  I  might  try  to  tune  this  model by  adding  some  more  parameters ,

but  when  I  do ,  I  can  see  that  R square is  not  really  changing .

I  don 't  think  I  want to  add  more  parameters

and  risk  overfitting  the  model .

That  was  one  extra  way  to  do  it .

Then  the  second  model that  showed  up  easily  for  us

using  the   Model Screening  platform was  to  do  Stepwise .

The  way  we  did  that  was we  put  our  output  here

and  then  use  the  shortcut to  do  a  response  surface .

That  includes  all  main  factors ,

all  squared  terms , and  all  two -way  interactions .

Then  if  we  change  this  to  Stepwise  here , we  can  hit  the  Run  button  and  Go .

Now  JMP  is  going  to  enter and  exit  everything

until  it  finds the  model  that  fits  the  best .

We  can  go  ahead  and  run  that .

Now  we  have  three  models that  we  want  to  compare .

What  I 'm  going  to  do  is  I 'm  going to  take  this  first  model  and  save .

I 'm  going  to  publish that  prediction  formula

to  the  Formula  Depot .

I 'm  going  to  give  it  a  really  quick  name , and  we  will  call  this  DSD .

Whatever ,  I  can 't  type .

We 'll  call  it  DSD ,  and  then  close  it .

We 'll  do  this  with  the  Neural  as  well .

We  will  publish  that  prediction  formula ,

give  it  a  name ,

and  do  the  same  with  this  final  model

where  we  will  publish that  prediction  formula .

This  last  one  was , we  called  it  Stepwise .

Now  I  have  these  three  models and  I  can  compare  them .

We  can  run  the  Model  Comparison  platform for  all  three  of  them

from  within  the  Formula  Depot.

We  can  get  a  rank of  the   Rsquares  of  those  models .

We  can  look  at  the  actual   versus  predicted ,

and  we  can  see  that  they 're all  predicting  about  the  same .

We  can  also  look  at  the  predicted  by  row ,

and  we  can  see  this  one  point from  the  Definitive  Screening  Design

is  the  one  that  was  left  out  when  we  fit the  original  Definitive  Screening  Design .

It  seems  to  be  important , and  probably  most  importantly ,

we  can  look  at  a  profiler for  all  of  these  against  each  other .

If  I  turn  off  the  desirability ,  we  can  see how  these  models  compare  to  each  other .

For  example ,  we  can  look at  that  Definitive  Screening  Design

and  see  that  it 's  showing  some  curvature where  the  other  two  models  are  not .

Maybe  we  can  look  and  see  over  here

that  the  curvature  is  different for  each  of  those .

Then  the  question  becomes , which  model  is  best  and  how  do  I  know ?

Then  that  brings  us  back  to  Bill .

Thanks ,  Jed .

Let  me  share  my  screen .

Can  you  see  my  screen ?

Yes .

I 'm  going  to  just ...

Jed  saved  all  that  to  the  Formula  Depot, so I 'm  just  going  to  execute  the  script

that  will  take  us  to  the  Formula  Depot that  he 's  already  saved .

Again ,  we 'll  then  just go  right  to  the  profiler .

We 're  going  to  fit  all  three  of  these .

Then  we  have  the  profiler .

This  is  what  the  real  power ,  I  think , of  the  DOE  is,

because  the   Prediction Profiler ,  we  can optimize  and  maximize  the  desirability .

The  response  surface  models  will  tell  us what  combination  of  factors  we  need

to  get  the  highest  elastic   or  highest  value  of  the  output  parameter .

What  was  really  eye -opening  for  me is  if  you  look  at  the  values  that  we  get

when  we  do  this  optimization , two  of  the  predictions ,

the  Neural  net and  the  Definitive  Screening  DOE,

are  giving  us  values  of  the  Y  parameter

that  are  greater  than  anything we  saw  in  our  initial  data  table .

We  had  a  maximum  value  of  34 .

Let 's  see.

I 'm  sorry , I  just  got  to  get  that  screen  back .

Typically ,  it 's  very  unusual for  me  to  see  this  with  the  DOE .

Typically ,  the  model ,  if  you  maximize  it ,

is  generally  close  to  what you  see  as  in  the  table .

But in  this  case ,  it  looks  like we  really  have  some  low -hanging  fruit .

We  needed  to  test this  combination  of  parameters

and  really  see  if  that  prediction was  valid  or  not .

If  we  go  back  to  our  JMP  journal...

I  just  want  to  show  you  what  happened .

We  took ,  I  think  this  is  the  prediction from  one  of  the  Neural  network  fits .

Again ,  the  highest  value in  the  Definitive  Screening  DOE  was  34 .

The  model  prediction  was  42 , but  when  we  actually  ran  it ,

we  saw  some  artifacts  in  the  film that  were  not  acceptable .

The  plasma  itself  was  stable .

There  was  no  way  to  see  this until  the  wafer  came  out  of  the  reactor .

But you  can  see there 's  a   bullseye  pattern ,

which  is  due  to  film  non -uniformity .

In  this  case it 's  very  thick in  the  middle  and  thin  at  the  edge ,

which  gives  us  this  bullseye  pattern .

Then  if  you  look  carefully ,

you  can  see  all  these  small  dots over  the  wafer ,

which  are  actually  the  holes under  the  shower head .

The  shower head  has  thousands of  small  holes  where  the  gas  comes  out .

In  this  case ,  we  have  a  shower head  pattern  and  a  bullseye .

I  think  the  model  is  telling  us what  direction  to  go .

But again ,  plasmas  are  challenging  to  use .

Even  though  the  model  was  telling  us

this  film  should  have the  highest  value  of  Y ,

the  film  itself  was  unacceptable .

Then  we  have  to  rely  on  our  process and  theoretical  knowledge  of  the  process .

We  know  that  argon has  a  lower  ionization energy,

and  if  we  substitute  argon for  helium  and  the  plasma ,

we  can  get  a  higher  plasma  density ,

which  may  help  us overcome  these  challenges .

What  we  did  is  switch  to  argon ,   and  you  can  see ,

although  the  film  is  not  perfect , it 's  much  more  uniform

and  certainly  good  enough  for  us to  get  the  physical  properties

of  the  film  that  we  can  use .

In  this  case ,  we  were  able to  hit   a  Y  value  of  46 ,

which  again , is  much  greater  than  34 .

We 're  certainly  trending in  the  right  direction .

What  we  really  wanted  to  do  is ,

are  there  any  opportunities  for  us to  further  improve  the  film ?

Again ,  that 's  where the   Prediction Profiler

or  the  response  surface  models are  very  powerful .

If  we  just  look  at  the  trends that  we  see  here...

I 'm  just  going  to  blow  these  up

so  we  can  see  them  a  little  bit  better for  each  of  these  cases .

The  data  is  really  telling  us,

in  certain  cases ,  there 's  things that  we  really  want  to  investigate .

Lower  temperatures

look  like  it 's  definitely favoring  the  highest  value  of  Y ,

pressure  appears  to  be  a  key  parameter ,

and  low -frequency  power for  this  initial  DOE

looks  like  in  two  cases ,

you  want  to  go to  higher  low -frequency  power .

The  Stepwise  is  giving  us  the  opposite .

But you  can  see that  this  is  really   a  blueprint

for  us  to  do  an  additional  design to  see  how  far  we  can  push  it .

Can  we  go  to  lower  flow  rates, can  we  go  to  lower  temperatures,

can  we  go  to  lower  pressures and  further  improve  the  film  properties ?

It 's  really  sequential  learning ,

and  that  for  me , is  the  real  power  of  the  DOE .

We  don 't  really  have  time to  go  through  all  of  that ,

but  what  I  did  is  put  together a  new  JMP  table  with  the  results

from  our  sequential  learning for  this  set  of  experiments .

Here  is  the  same  data  we  saw in  the  Definitive  Screening  DOE .

Here  are  our  Y  values , and  we 're  ranging  from  9 to  34 .

The  different  colors are  the  different  DOE .

Here 's  the  next  DOE  that  we  did .

What  you  can  see  is  based on   the  trends  in  the  Prediction  Profiler

and  the  response  surface  models ,

we  fixed  the  low -frequency  power at  the  highest  setting  we  could .

It  turns  out  physical  limitations for  this  plasma  chemistry  prevented  us

from  adding  any  more than  20 low -frequency  power .

We  also  fixed  the  temperature .

We  can 't  operate for  this  chemistry  below  200 .

We  know  that  lower  temperature gave  us  the  highest  value ,

so  we  fixed  these  two

and  then  did  a  five -factor  DOE focusing  on  lower  precursor  flows ,

various  spacings ,  the  higher  powers , and  certainly  the  lower  pressures ,

which  was  indicated  as  one of  the  most  important  parameters .

If  you  look  at  the  Y  values  here ,

you  can  see  we 're  definitely trending  in  the  right  direction .

Now  we 're  going  from  mid  20s  up  to  56 .

We 're  certainly  above the  46  we  saw  there .

Then  we  did  the  same  learning .

Again ,  the   Prediction Profiler  indicated what  parameters  we  should  explore .

We  did  another  DOE .

In  this  case , we  fixed  different  parameters ,

but  you  can  see that   the  trend  is  the  same .

Now  we 're  hitting  up  to  66 in  terms  of  our  Y  value ,

and  we  did  one  final  experiment ,

and  in  this  case  you  can  see basically  the  sum  of  all  the  knowledge

that  we  gained .

It  turned  out  when  we  switched  to  argon , you  could  add  more  low -frequency  power .

You  could  go  from  zero  to  40 %.

In  our  final  analysis , this  DOE  showed  that  unlike  the  first  DOE ,

after  we  finetuned  everything and  switched  gasses ,

the  low -frequency  power had no  statistical  impact  on  the  Y  value .

We  set  that  to  zero .

We  found  out  the  lowest  spacing was  the  most  important .

Our  sweet  spot  for  pressure  was  2 .3  torr ,

and  we  did  want  to  operate at  the  lowest  temperature .

Really  we  had  a  three -factor  DOE   between  total  power ,  precursor  flow ,

and  argon  dilution to  really  dial  in  the  films ,

and  we  could  hit a  maximum  value  of  84 .

I  summarized  all  that  in  a  box  plot  here , which  I  think  really  shows

the  ability  of  the  DOE and  sequential  learning ,

where  we  started  out  with  a  seven -factor Definitive  Screening  DOE  with  26  runs

and  ended  up  with  a  three -factor I -optimal  design  with  16  runs ,

but  you  can  see our  continued  improvement .

This  was  our  reference  target , so  we  still  have  more  work  to  do .

But this  is  a  very  powerful  way  for  us to  screen  seven  factors

with  three -level  designs in  a  very  short  period  of  time .

I  do  think  it 's  worthwhile just  to  point  out

how  efficient  these  new  modern  DOE s  are .

If  we  looked  at  what  we 're  really  doing ,

we  have  seven  factors that  we  started  with .

All  of  these  are  three -level  designs .

For  a  three -level ,  seven -factor  design , that  would  be  over  2 ,100  runs .

We  could  run  90 out  of  the  experimental  designs

and  achieve  this  increase  in  the  Y  value .

I  think  these  modern  designs ,

the  optimal  designs  combined with  the  Definitive  Screening  DOE s

are  a very  powerful  tool  for  us to  get  the  most  value

with  the  fewest  number  of  experiments .

The  final  thing  I  want  to  touch  on

is  when  we  switched to  a  different  precursor .

This  is  really   a  different  challenge  we  faced .

The  goal  here  was  to  evaluate   different  precursors

to  compare  how  it  stacked  up against  the  initial  baseline  film .

What  we  tried  to  do  is  use all  of  our  learning  from  those  four  DOEs

and  become  even  more  efficient.

Instead  of  90  runs, can  we  do  this  in  52  runs ?

With  Jed 's  help ,  we  put  together an  eight -factor  A -optimal  design ,

but  what  we  found  is  that the  chemistry  was  shockingly  different .

All  of  the  parameter  space  that we  could  operate  easily  with  Precursor  1

was  not  the  case  here .

In  fact ,  I  put  together  a  slide to  show  you

how  bad  some  of  these  films  could  look , so  we  could  get  perfect  films .

You 'd  be  hard -pressed to  tell  there 's  a  film .

This  is  a  nanometer -thick  film ,

edge  to  edge  on  a  12 -inch  silicon  wafer , perfectly  uniform .

Then  we  would  have  films that  look  like  this .

Obviously ,  that 's  not a  design  that  we  wanted .

But the  challenge  that  we  faced  was , we 're  doing  an  eight -factor  DOE ,

and  we 're  trying  to  do  this quickly  and  efficiently ,

and  30 %  of  the  runs  failed .

I 'm  looking  at  a  table with  eight  different  factors .

How  do  I  pick  out  the  factors that  are  contributing  to  this ?

What  we  did  is  we  created  a  table for  our  eight  input  factors

and  then  identified  all  of  the  films

that  had  delamination ,  arcing , or  other  issues ,

and  then  created  a  metric , just  a  film  metric,  pass  or  fail .

It  turns  out   we  can  fit  this  categorical  variable

and  see  if  we  can  get  a  model that  will  help  us  understand

what  is  really  causing these  issues  through  all  these  films .

The  first  thing  we  can  do  quickly

is  again  go to  our  Model Screening platform,

Predictor  Screening ,

and  get  a  handle  on  what  factor ,  if  any , is  really  controlling  film  quality .

If  we  look  at  this , it 's  pretty  clear  and  it 's  shocking,

because  this  was  not  the  case with  Precursor  1 .

The  flow  rate  of  the  precursor was  by  far  and  away

the  most  dominant  factor   impacting  film  quality .

But what  we  needed to  run  these  experiments

is  not  just  knowing  this  factor ,

but  what  value  can  we  safely  run to  generate  quality  films.

That 's  where  we  did  the  Neural  net .

Again ,  we 'll  go  into  Predictive  Modeling, Neural  net,

we 'll  take  our  factors ,

and  we 're  going  to  fit  film  quality   as  a  categorical  variable .

I 'm  going  to  go  to  boosting  of  20 , as  Jed  mentioned .

That 's  typically  what  the   Model Screening comes  up  with ,

and  we 'll  generate  our  model ,

and  you  can  see  we  get excellent   Rsquare  values .

This  is  a  categorical  model .

I  believe  that  looking  at  the  ROC  curve

provides  insight into  how  well  the  model  fits .

If  the  curve  is  along  this  diagonal , it 's  just  basically  a  guess .

This  looks  like  a  half  square  wave .

These  are  basically  perfect  fits in  the  ROC  curve .

Then  the  question  is , how  can  we  utilize  that  data ?

The  nice  thing  about  the  Neural  net is  that  you  have  a   Categorical Profiler.

I  can  execute  the   Categorical Profiler ,

and  now  I  set  this  up  where  we know we  want  to  operate  at  lower  pressures .

We  know  from  the  previous  work we  want  to  operate  at  lower  spacings .

We  want  to  go to  the  lowest  possible  temperature.

We 'll  just  set  this  in  the  middle .

Then  basically  we  have this  profiler  that  tells  us

the  primary  factor  affecting  this is  really  the  argon  flow  rate .

If  we  can  keep  our  flow  rate below  400  sccms ,

we  can  have  a  100 %  success  rate

for  the  films that  we 're  trying  to  optimize .

With  this ,  we  set  up  a  new  DOE ,

limiting  the  total  flow  rate  to  400  sccms .

We 'll  go  back  to  our  JMP  journal .

We  were  able  to  come  up with  a  new  design

and  complete  41  of  the  42  runs , and  we 're  still  executing  that  study .

But it  just  shows  how  powerful

the  Neural  network  is for  a  categorical  variable

where  we  can  do  this  in  an  afternoon,

where  at  one  o 'clock  we  found  these  films  weren 't  working .

Three  hours  later ,  we  had  a  model that  told  us  how  to  set  up  a  new  design

and  we  were  executing  that  later  that  day .

I  think  that 's  the  material we  wanted  to  cover .

Presenter