Dear community, I have a stacked data table, and want to exclude outliers by a Label (e.g. an element for some measurements). In the robust fit outlier platform I can detect outliers and than can select or exclude corresponding rows. For my understanding, rows that are selected than or excluded, should be exactly the same ones. And that is the case for the below sample data table used in the JSL below.   // from scripting index Names Default To Here( 1 ); dt = Open( "$SAMPLE_DATA/Water Treatment.jmp" ); // similar, but stacked data and usage of by role dt_stack = dt << Stack( columns( Column Group( "Sensor Measurements" ) ), Source Label Column( "Label" ), Stacked Data Column( "Data" ), Output Table( "Water Treatment stacked.jmp" ) ); obj_select = dt_stack << Explore Outliers( Y( :Data ), By( :Label ), Robust fit outliers, select rows() ); obj_exclude = dt_stack << Explore Outliers( Y( :Data ), By( :Label ), Robust fit outliers, exclude rows() ); Show( N Items( dt_stack << get selected rows() ), N Items( dt_stack << get excluded rows() ) ); But in my case with real data, I get different rows selected and excluded, it looks to me that the platform works well in identifying the outliers, but the exclude operation fails to exclude one row (that has been selected as outlier). Any idea on this behaviour?     ... View more

What inspired this wish list request?  I've been using different ways of connect to databases, the documentation describes well the "open database" and "execute sql" way, but not the "new sql query" way. (Database Access (jmp.com)). Maybe it's hidden under "run query builder queries".   What is the improvement you would like to see?  see: Which way to go query database? - JMP User Community There is no statement about what to use in future, the different behaviour regarding "hide odbc connection string" has hurt us a lot (JMP17 behaviour on Preference "Hide ODBC Connection Strings" set: connection str... - JMP User Community). I would like to see more detailed how the one or the other method works, what are Pro's and Con's to use one or the other. There seems to be a favourite one from JMP side, one that is the future method, and others are legacy. That should be made available to others.   Why is this idea important?  Because we could streamline our scripts, use one best know method, and do not need to troubleshoot with a lot of different behaviour regarding preferences settings and script functionality. All our data comes from database, so database functions are very important by default.     ... View more

Using JMP 17 we have been a bit surprised to see what happens in detail with the above preference set, i.e. when performing a SQL Query, the connection string is removed (and not hidden like JMP16) in the result table. When starting the table script "Modify Query", the query builder opens, and we can look at the query, but pressing the "Execute Query" leads only to a generic "Error running SQL Query". I understand, that removing the connection string is a security feature to not distribute passwords, server names etc. And the good thing is, that the query definition (SQL) is still available.   This is my proposal: "Hide ODBC Connection Strings" should be renamed into "Remove ..." it should be made more transparent, that the Query will not be functional with this preference set, i.e. making visible that the SQL-Query object is not connected to any database, and throwing a more distinct error accordingly maybe there could be implemented a way to again connect that query to a database under the red triangle, i.e. "Connect to Database" to make it working again Thanks!           ... View more

Georg Raming, Senior Manager, Siltronic AG   Siltronic AG is a global technology leader in the semiconductor wafer industry. This presentation will introduce the Siltronic AG approach to preparing batch process data for modeling with JMP Pro. It will demonstrate some interactive steps to clean and rearrange the dataset before modeling using an anonymized dataset containing both historical and experimental batch data. Once the best model algorithm is found, the boosted tree model will be tuned.  The Siltronic AG team found that a technically sound model may be physically worthless, meaning it had been overfitted. Therefore, the team started with a large set of factors, gradually reducing the factor list and testing the model's behavior to find the most effective factors (step backward strategy for a boosted tree in a small JSL routine). The last step provided the best insight into which levers are the strongest to optimize the process.     Hello,  everyone.  Thanks  for  joining  in.  In  this  talk,  I  want  to  talk  about  how  we  did  prepare  our  batch  process  data  for  modeling  with  JMP  Pro  and  gaining  valuable  insights  with  a  team  approach.  My  presentation  is  detailed  in  a  PowerPoint  part.  That  is  the  first  part  and  the  details  all  here  shown  will  follow  in  JMP,  like  how  the  data  set  looks,  which  platforms  I  have  used  like  missing  data,  multi collinearity,  functional  data  explorer,  predictor  screening,  modeling  batch  data  with   Boosted Tree  and  Profiler. Summarized  data  will  be  analyzed  by   Boosted Tree  as  well,  and  then  by  a  script with   Boosted Tree  backward  selection.  At  first,  my  company,  Siltronic,  has  world- class  production  sites  all  over  the  world  like shown  here  and  about  4,000  employees.  Here  are  some  key  figures.   If  you  imagine  that  we  have  a  complex  process  flows  like  shown  here  with  silicon  mold  in  a  crucible.  Silicon  ingot  is  created  here.  That's  my  special  task.  To  make  processes  for  growing  silicon  ingots.  Then  the  ingot  is  ground  and  sliced. Edge  rounding  is  done  for  the  wafers,  laser  marking,  lapping,  cleaning,  etching,  polishing,  and  maybe  epitaxy  for  the  final  wafer  to  be  created.  Our  portfolio  is  that  we  are  selling  300  millimeter,  200  millimeter  and  smaller  diameter  wafers  for  different  applications  like  shown  here,  silicon  wafers  with  several  specifications. About  me,  my  education  is  I'm  an  electrical  engineer  and  I  did  some  Six  Sigma  education,  and  my  main  task  is  to  develop  processes  for  growing  silicon  crystals  like  shown  here,  and  I'm  as  well  responsible  for  around  500  users  at  Seltronic  JMP  users.  How  does  the  task  look  like?   What  we  see  here  is  the  final  table,  but  it  has  been  created  and  this  takes  a  lot  of  effort  as  well.  So  there  are  some  database  queries  behind  to  get  this  data  from  database. We  fetched  the  results  into  JMP  data  tables  and  enlarged  the  data  set  with  archives  from  earlier  date  and  enriched  some  information  like  details  of  experiments  and  details  on  consumables  and  wrote  some  script  for  graphs  and  evaluations.  T hen  we  have  done  the  modeling  tasks  and  of  course,  looked  for  missing  data  correlations  to  see  what  are  the  most  significant  effects  and  to  do  feature  engineering  to  see  which  features  are  important  to  generate  an  optimal  result. At  this  point,  I  will  switch  into  JMP  then.  We  can  see  here  my  journal  I'm  working  with  and  the  JMP  main  window  and  the  abstract  is  seen  here.  We  will  start  with  technical  hints.  The  use  data  set  I  show  here  is  fully  anonymized  and  standardized,  and  all  identifiers  are  generic  for  better  understanding  what  are  the  features,  what  is  the  result,  and  so  on.   The  aim  of  this  presentation  is  to  show  all  the  steps  we  needed  for  getting  an  overview,  restructuring,  and  understanding  the  data  set,  and  how  to  build  the  models  to  get  some  insights  of  the  content  of  the  data  set. I  will  show  some  results  that  we  have  discussed  in  a  team.   The  team  is  very  important  here  because  the  team  drove  a  lot  of  discussion  and  work  as  well,  how  to  analyze  and  what  features  may  be  interesting  and  what  should  not  be,  and  what  may  be  the  physics  behind.   I  will  start  with  the  data  set  here.  It's  also  a  part  of  the  contribution  in  the  community.  Here  it  is  opened  and  I  will  change  the  design  a  little  bit  to  see  how  it  looks  like.   We  have  around  80,000  rows  in  this  data  set  and  it's  a  batch  data  set,  so  we  have  a  batch  ID. T his  data  set  is  quite  challenging  because  it  has  a  mixture  of  historical  data  like  here,  POR  batches.  We  can  see  here  that  we  have…  Most  of  the  data  is  historical  data,  and  there  are  only  a  few  special  experiments  shown  here.   We  have  several  features  then,  like  one  categorical,  it's consumable.   We  have  the  batch  maturity,  it's  the  time,  also  standardized.   Then  we  have  several  features.  So  these  X  values  here,  we  have  one  result  column,  and  to  reduce  the  noise  a  little  bit,  we  have  calculated  a  new  moving  average  as  well. Let's  have  a  look  at  how  the  data  set  looks  more  in  detail.  We  can  see  here,  if  we  do  a  summary  on  the  data  like  this,  you  can  do  this  from  the  table s  menu  as  well.  Summary.  We  get  around  500  rows,  500  batches.   This  is  a  summary  by  batch,  and  we  see  that  there  is  no  variation  in  the  parameters  X 1  to  X 4,  meaning  that  they  are  constant  for  each  batch,  and  the  others  are  changing  at  different  rates.  To  have  a  look  how  the  data  looks  at  all,  we  can  see  here  the  result  parameter  like  yield,  a  long  time  for  all  the  rows  of  the  batch  data  set,  and  this  smoothing  is  done  by  JMP  Graph  Builder  platform. We  implemented  this  as  a  formula  as  it  is  available  as  a  function  in  JMP.   We  can  have  a  look  here  at  some  special  batches  as  well.  If  we  use  the  local  data  filter  and  see  here  how  the  average  works  and  what  noise  is  in  the  single  data  points,  the  blue  ones  are  the  original  data  of  yield,  and  the  orange  one  are  the  moving  average.  I  will  close  this  then,  and  next  point  may  be  to  look  at  how  much  data  is  missing.  So  we  have  this  in  JMP  as  well.  We  can  mark  all  the  columns  and  do  the  missing  data  pattern  platform  like  this. It  will  show  us  that  from  about  80,000  rows,  we  have  178  rows  with  some  missing  data  in  one  column.   This  can  also  be  shown  here  as  a  graph.  This  is  very  important,  at  least  for  the  data  creation steps,  it  was  important  to  see  where  is  some  data  missing  and  to  fix  this  missing  data  then  as  much  as  possible.  Another  step  to  look  at  the  data,  maybe  Columns  Viewer.   We  can  get  here,  I  put  all  the  columns  in,  and  here  we  can  see  again  like  we  had  before,  there  is  some  rows  missing  for  parameter  X2. We  can  see  what  the  min,  max,  mean,  standard  deviations,  and  so  are  for  all  the  parameters  we  can  see  here.   Here  we  nicely  see  that  everything  is  standardized.   For  the  yield,  it's  between  zero  and  100.  We  can  as  well  start  from  here,  distribution  platform.  So  all  the  columns  are  marked  and  we  get  by  only  one  click  for  all  the  data,  the  distributions.  We  can  see  here  what  consumables  are  used  how  often  that  we  have  most  data  from  historic  processes  and  only  some  of  some  experiments  with  special  settings. The  time,  of  course,  looks  nicely  distributed,  but  the  others  don't  look  that  nicely.  So  there  is  a  lot  of  room  between  some  settings,  and  it's  sparsely  distributed,  non- normal  distributed  for  the  most  parameters,  and  that  makes  it  even  more  challenging  to  analyze  this  data.   We  go  to  the  next  steps.  I  will  close  these  reports.   Then  we  may  look  even  more  in  detail  on  some  things  like  how  the  parameters  are  correlated.  We  can  see  this  in  the  multivariate  platform.  It  needs  some  time  to  be  calculated. You  will  find  it  here  under  the  analysis  menu,  multivariate.   It  takes  the  numeric  columns  and  generates  this  correlation  report,  and  you  will  see  that  the  parameters  like  X6  and  X5  are  highly  correlated.   This  makes  it  difficult,  like  X10  and  X9  as  well,  makes  it  difficult  to  do  feature  engineering.   What  we  want  to  know  from  the  analysis  is  which  parameter  causes  some  yield  drop.   If  two  parameters  are  correlated,  it's  not  so  easy  to  detect  or  to  find  out  which  one  is  the  responsible  one. Here  in  the  scatter  plot  matrix,  you  can  see  as  well  which  parameters  change  with  time,  like  X1,  X2,  up  to  X4  is  constant  over  time,  and  the  others  are  changing  and  how  they  are  distributed,  and  you  can  nicely  mark  some  rows  like  here.  They  are  selected  in  the  data  table  then  and  see  how  the  curves  are  for  each  parameter  over  time  or  which  parameter  over  what  parameter  combination  looks  like.  Next,  I  want  to  use  the   functional data explorer. The   functional data explorer  allows  us  to  fit  curves  for  each  batch  and  extracts  the  features  of  each  curve.   Then  we  can  have  a  look  at  which  batches  behave  similar  or  maybe  extreme  ones.  So  the  start  is  like  this.  We  can  have  a  look  at  how  I  started  the  analysis.  We  launch  analysis.   I  put  time  as  an  X  parameter,  Yield  as  the  output  parameter  Y,  and  the  ID  function  is  the  Batch  ID.   Then  we  have  here  some  informal  parts  like  Part  and  Group. This  platform  is  available  in  JMP  Pro  only,  and  when  we  start  it,  we  can  do  some  data  processing  here.  But  in  this  case,  it's  not  necessary.  We  can  have  a  look  at  each  batch,  how  it  looks.  So  there  are  a  lot  of  graphs  here  like  this.  We  can  mark  the  rows.  We  can  see here  the  marked  rows  and  the  data  table  as  well.  To  go  on  with  this  platform,  we  need  to  make  some  models  like   P-splines  for  each  batch  and  JMP  does  this  and  defines  itself  which  splines  are  used  and  how  many  supporting  functions  are  needed,  like  the  knots  shown  here. So  the  best  result  is  given  with  a  cubic  spline  with  20  knots.   You  can  see  how  each  batch  is  modeled  here  by  the  red  line  shown  here  and  how  it  looks.   We  have  here  the  shape  functions.  So  each  curve  is  added  together  by  a  combination  of  shape  functions,  and  we  get  for  each  batch  the  coefficient  for  each  shape  function.  If  we  are  looking  at  Shape  Function  1…  This  is  the  main  behavior  of  all  batches  with  a  drop  here  at  around  0.7.  We  can  see  that  here  we  have  Component  1.  This  is  a  coefficient  for  the  shape  function  one. If  we  select  these  batches,  we  will  see  that  they  have  a  pronounced  shape  like  Shape  Function  1.  We  can  see  it  here.   We  can  as  well  use  the  Profiler.  So  this  is  mostly  for  understanding  the  data  but  we  have  not  used  it  for  further  analysis  because  we  did  not  really  need  the  information  how  the  back  batch  looks  as  a  shape  for  each  curve.  We  were  more  interested  in  average  yield  of  each  batch  because  we  cannot  define  only  to  use  the  first  part  of  the  batch  and  forget  about  the  second  part.  This  would  not  work  in  our  case. As  well  to  see  again  how  this  works  together,  we  can  have  a  look  in  Graph  Builder.  The  graph  of  some  batches  we  have  seen  just  before.  Maybe  you  see  this  number  here.  We  have  seen  it  before.   Here  it  is  shown  again  together  with  the  moving  average  of  yield.  Next  step  would  be  to  start  modeling  of  the  batch  data. When  doing  modeling,  it  may  be  interesting  to  see  or  have  an  idea  which  parameters  are  most  important  for  the  variability  of  the  output.   There  we  have  the  predictor  screening  platform.  You  can  as  well  start  it  from  here.  Analysis  and  predictor  screening.  I  wrote  it  here  as  a  script  simply  to  start  it  by  pressing  a  button.   When  doing  so,  we  will  see  some  Bootstrap  Forest  analysis  going  on,  and  it  shows  us  the  importance  of  the  features  we  have  in  our  data  set.  Time  is  the  most  important,  but  this  is  useless  at  the  end  for  us  because  we  need  to  use  the  full  batch. T hen  comes  X1,  then  comes  part  X8,  X5,  and  so  on.  So  here  we  could  as  well  select  a  few  rows,  copy  them  and  put  them  into  a  model.  I  will  stop  this  here,  and  to  see  which  model  works  best,  I  used  model  screening  platform.   I  will  not  run  it  here  because  it  takes  several  minutes.  But  we  have  seen  that   Boosted Tree platform  may  perform  well.  There  is  maybe  not  so  a  big  difference  between  the  next  ones,  but  that's  the  reason  why  I  used   Boosted Tree platform. Then  we  will  run   Boosted Tree  like  this  on  the  batch  data,  and  it  works  quite  quick.  We  will  see  the  result,  and  a  nice  feature  of  the   Boosted Tree platform  as  well  is  that  you  have  column  contributions  so  that  you  can  nicely  do  some  feature  engineering.  We  can  see  here  that  we  have  71 %  R square  for  training  and  66  for  validation,  may be  okay,  and  we  have  still  all  features  in.  But  we  are  interested  mostly  in  which  features  are  reliably  the  most  important  ones. When  doing  this,  we  can  save  it  as  a  column,  save  prediction  formula  in  the  data  table.  We  see  in  the  data  table  we  have  a  formula  now  and  we  can  use  it.  We  can  maybe  have  a  look  at  how  the  model  performs  or  to  use  it  in  Graph  Builder  simply  to  see  how  the  model  data  looks  like  over  the  batch  maturity.   I  hope  the  Graph  Builder  to  show  the  graph  the  graph  soon,  and  here  it  comes.  Yes.   We  have  seen  that  this  modeling  works  quite  well,  so  we  have  a  formula  now  to  rebuild  the  data  and  we  can  maybe  work  with  it. But  especially  for  the  batch  data  modeling,  we  have  a  problem  here  that  validation  will  not  work  because  we  may  have  here  for  some  batch,  these  rows  in  training  set  and  the  rows  next  to  it  in  validation  set.  So  they  are  not  well  separated  for  the  features  that  control  the  batch  then.   Additionally,  the  model  is  not  very  stable,  so  we  will  get a  different  result.   For our  different  runs  of  the  model,  this  is  known  from  tree- based  methods  that  they  may  give  different  results  for  high  variability  data. If  we  do  something  like  running   Boosted Tree  twice,  we  get  also  different  column  contributions  and  maybe  different  order  like  we  can  see  here  for  the  part  and  X 5  are  switched  here  for  these  two  runs,  and  I  will  show  it  again  also  here.  If  we  run   Boosted Tree  twice.  Don't  know.  Yes,  here  I  should  have  the  script.  It  comes  later.  So  at  this  point,  we  have  said  that  it  may  be  better  to  model  the  summarized  data  because  we  need  to  use  the  full  batch.   Here  I  have  a  script  now  to  summarize  the  data  in  a  form  that  we  have  only  one  row  for  each  batch. There  is  a  nice  feature  statistics  column  name  format  that  we  get  the  same  columns  for  the  summarized  data  as  we  have  in  the  original  table  that  we  can  use  the  same  scripts  for  both.   Doing  so,  we  get  here  the  summary  data  table— I  can  close  the  script— with  around  500  batches.  It's  a  lot  of  easier  to  model,  and  here  I  have  summarized  the  data  for  0.6  to  0.8  time.  So  it's  where  the  yield  drop  was,  and  we  can  again  do  here  some  predictor  screening  like  this  and  see  that  I  still  have  time  in  that  data  set  to  see… It's  more  like  to  see  what  level  noise  is,  and  it's  around  these  parameters  that  are  likely  also  noise  for  the  model  then.  Then  we  can,  of  course,  do  some  model  comparison.  So  I  selected  a  few  parameter  that  we  found  to  be  responsible,  most  probably,  and  I'm  doing  two   Boosted Tree  analysis  and  then  do  some  model  comparison  for  both.   It  looks  like  this.  We  can  see  here  we  get  a  Profiler  and  compare  the  result  for  different  settings,  maybe  like  this.  Here  we  have  still  the  problem.  We  see  some  features  like  this  here  for  X10  in  this  model  and  not  in  that  model. So  it  likely  seems  to  be  noise.   At  the  beginning,  we  have  discussed  a  lot  about  these  differences.  We  have  seen  sometimes  and  sometimes  not,  and  asked  the  question,  what  is  true,  what  is  physical,  and  what  is  not.  That  brought  me  to  the  step  then  that  we  need  to  continue  with  feature  selection,  and  that's  why  we  created  this  script.  It  takes  this  summary  data  and  has  been  done  for  the  batch  data  as  well. For  each  step,  it  builds  Boost ed  Tree  model  for  the  full  parameter  set,  saves  the  model  into  the  formula  depot,  and  saves  the  model  performance  R square  and  so  into  a  data  table,  and  shows  us  the  column  contribution.   Here  we  can  see  something  that  we  have  seen  often  that  with  the  higher  number  model,  it  is  the  model  with  less  parameters  like  we  can  see  here  with  the  column  contributions,  we  have  the  best  result.   It  looks  different  for  each  run,  but  the  tendency  we  see  in  most  times.  So  here  we  can  see  for  more  or  less  sure  that  part  and  X1  and  X5  are  the  most  important  parameters. This  one  may  be  there  sometime  and  maybe  not,  so  we  will  focus  on  these  three  parameters.   As  well,  we  can  have  a  look  in  the  Formula  Depot.  There  we  can  start  model  comparison.  We  maybe  can  compare  the  first  model.  So  we  do  it  like  this,  model  comparison.  This  is  our  data  table.  Take  the  first  number.   The  numbers  here  are  shifted  by  one,  and  maybe  the  5th  should  be  that  one,  and  the  last  one.  This  will  not  work.  I  think  it's  number  three,  and  the  last  one. The  ones  with  the  highest  validation  score  and  maybe  compare  them  here.  We  see  the  model  comparison  dialog.   We  see  that  the  last  model  is  between  the  best  models  we  could  fit  at  all,  and  can  see  here  the  Profiler,  for  example,  and  as  well,  may  use  extrapolation  control.  We  have  seen  that  we  have  sparse  data,  so  not  behind  every  point.  There  is  some  data.  Let's  look  where  it  is.  Here  it  is.  Extrapolation  control  warning  on.  So  it  shows  us  when  there  is  no  data  between  the  points. Here  we  can  maybe  compare  the  models  and  we  see  here  that  there  is  no  variability  on  the  X  factors  that  haven't  been  used  here.  To  sum  up,  let's  close  some  tables  first  and  some  dialogs.  To  sum  up,  we  have  prepared  a  workflow  for  modeling  this  data  and  have  done  several  steps  and  additional  script  to  enhance  understanding  and  to  drive  the  discussion  about  what's  important  and  what's  not. I  have  a  proposal  for  a  model  and  some  tasks  we  can  focus  on  to  improve  the  year  yield  of  our  process,  and  you  will  find  the  data  and  the  presentation  in  the  user  community.   If  you  have  other  ideas  how  to  explore  this  data  set  and  how  to  find  the  final  best  model,  you  can  contact  me  or  post  something  on  my  contribution  in  the  in  the  community  for  this  Discovery  Summit.   Thanks  for  your  attention  and  bye.  That's  it,  Martin. ... View more