Level III

Mixing Plan for Wine Blending and Tasting of the Modalities To Validate the Recipe of a Wine (2023-EU-30MP-1313)

Margaux RENAUD, R&D engineer, Chêne & cie


A client working on blending his wine wanted to get as close as possible to a known wine with different ingredients. To help the client do this, we created a mixture design with four ingredients for 16 samples to compare with the known blend. The samples were tasted randomly, and the panel was asked to create groups of similar samples and describe the group. The result was a distance matrix developed with the help of a JMP script. This matrix was processed by a multidimensional scaling to obtain a map that was easy to describe to the panel. A K-means classification was used to find the samples close to the target. Finally, the distance between the target and the other sample was calculated and represented by a contour plot to show the best part of the mixture design. The terms used by the taster to describe the groups were processed by JMP Text Explorer and then by AFCM to show a map with samples and terms to better describe each sample's position using sensory properties.


Good  morning  or  good  afternoon  for  all  of  you.  I'm  Margaux  Renaud  and  today  we  will  talk  about  the  mixing  plan  of  wine  blending  and  the  testing  of  these  modalities  to  validate  the  receipt  of  a  wine.

First  of  all,  I  would  like  to  present  my  company.  I'm  working  for  Chêne  Company,  which  is  a  group  of  cooperage.  I t  owns  a  French  cooperage,  Taransaud, which  is  make  barrels  and  vats  from  French  oak.  An  American  cooperage,  Canton,  and  Kádár  Cooperage,  and   [inaudible 00:00:42] ,  which   make  oak  wood  sticks  and  chips,   XtraChên.

The  French  cooperage  and  the  R&D  department  are  based  in  the  Bordeaux  area  in  France.  In  our  R&D  department,  we  have   8%  with  different  background.  We  have  a  PhD  in  chemistry  and  enology,  engineer,  agronomist,  enologist,  technician. W ith  all  these  various  skills,  we  have  a  lot  of  different  trials.

From  the  forest,  for  example,  we  can  do  trials  about  DNA  of  the  oak  in  the  forests.  On  the  aging  wood,  we  thought  to  do  the  barrel  or  the  vats  in  relation  with  the  climate  change.  Two,  the  analysis  on  our  client  wine  in  our  barrel  directly  in  the  wine.

Today,  I  would  like  to  talk  with  you  about  a  mixing  plan  for  a  wine  blending.  It's  a  client  trial  and  I  will  present  you  the  problematic  of  the  client.  In  this  case,  the  client  has  different  wine  and  one  of  them,  you  want  to  keep  in  the  same  wine  style,  but  you  want  to  optimize  the  ingredients.

In  the  wine  industry  what  we  call  ingredients  it's  very  different,  very  diverse.   It  can  be  different  variety  of  wine.  In  Bordeaux  area,  we  used  to  mix  Merlot  and  Cabernet,  for  example.  It  can  be  different  quality  of  wine,  different  type  of  aging.  If  the  wine  is  aging  in  barrel  or  in  tanks  with  oak  chips  or  without.

I n  this  case,  the  client  has  four  different  ingredient  to  mix.  Our  team  follow  the  different  ingredient  during  all  the  wine  aging,  and  at  the  end,  we  have  created  a   mixing plan  and  we  taste  it.

Just  before  to  go  on  JMP  to  present  you  the  way  to  process  data,  I  just  want  to  talk  a  bit  about  the  wine  testing.  An  important  thing  in  the  wine  industry  is  all  the  wine recipe  is  decided  by  testing.  We  do  a  lot  of  analysis,  but  it's  not  the  last  point  of  a  recipe.  It's  always  the  tasting.

The  usual  way  to  taste  wine  it's  a  quantitative  tasting.  We  do  a  profile  with  grade  on  different  descriptor.  It  could  be  bitterness,  for  example,  or  the  fruity  notes  or  the  woody  notes.   All  the  taster  are  writing  the  intensity  of  the  descriptor.

Then  we  process  the data  with  an  ANOVA,  a  two- factor  ANOVA.  The  first  factor  is  the  modalities,  the  different  modalities  in  the  trial,  and  the  tester.  To  have  really  significant results  you  need  to  have  a  large  and  trained  panel  for  your  testing.

In  our  case,  when  we  do  a  trial  with  a  client  or  in  our  group,  we  have  different  different  type  of  taster.  Most  of  the  time  you  have  the  winery  team,  some  part  of  the  commercial  team,  and  some  part  of  the  R&D  team.  A ll  this  taster  doesn't  taste  the  wine  in  the  same  way.  They  don't  have  the  same  target  when  they  taste  the  wine.

The  target  for  the  client  is  not  the  same  as  the  commercial  part  and  it's  not  the  same  for  us.   Most  of  the  time  we  are  not  trained to  taste  the  wine  in  the  same  way.  When  we  analyzed  the  data,  there  is  a  really  big  effect  of  the  taster.  In  fact,  the  taster  have  not  the  same  feeling  about  the  profile  asking.

For  us,  it's  complicated  to  using  the  profile,  so  we  decide  to  use  another  type  of  testing,  the  free  sorting.  The  free  sorting  it's  a  testing  when   I  asked  my  taster  to  test  the  different  modalities  and  to  make  group  groups  inside  them.  I  put  a  little  example  on  the  PowerPoint.

In  this  case,  I  asked  the  taster  to  make  groups  if  the  wine  is  similar  and  if  there  is  difference  between  the  two  modalities,  they  put  them  in  two  different  groups.  In  this  case,  for  example,  there  is  11  samples,  and  the  taster  decides  to  make  four  groups.  A  first  one  with  four  samples,  a  second  one  with  three,  another  one  with  three  of  the  sample,  and  the  last  one  with  an  only  glass  of  wine.

I  ask  them  after  making  group  to  describe  a  bit  the  group.  In  this  case,  the  taster  decide  to  put  together  for  sample  because  they  have  some  chestnut   not  present  in  the  other  samples.  I n  this  case,  we  don't  need  to  have  a  trained  panel,  so  if  there  is  enough  big  difference  between  my  modalities,  normally  all  the  taster  will  put  together  the  wine,  the  wine  really  close  and  put  separately  the  other  wine.

This  type  of  tasting  is  really  easy  to  use  for  us  because  we  don't  need  a  trained  panel.  We  can  have  a  small  panel  too.  It  can  be  used  in  different  language.  It  doesn't  matter  if  we  have  a  French  panel  or  an  Italian  panel,  for  example.  They  just  have  to  do  groups.

The  other  thing,  thanks  to  JMP,  it's  easy  to  present  the  result  right  after  the  testing.  When  you  do  a  profile,  you  have  to  process  data  making  the  ANOVA  test,  and  send  the  result  to  the  client.  Most  of  the  time  it  takes  a  few  days  or  a  few  weeks  if  you  are  really  late.

With  the  free  sorting,  we  can ,  and  thanks  to  JMP,  present  the  result  right  after.  This  type  of  testing  will  create  a  distance  matrix  between  all  the  sample.  In  fact,  if  you  put  samples  in  the  same  group,  there  is  no  distance  between  them.  If  you  put  them  in  two  other  group,  there  is  a  distance  of  one  between  them.   At  the  end,  you  can  make  a  matrix  distance  between  all  the  samples.  It's  what  I  do  with  JMP.   I  will  show  you  just  after.

Okay,  I  will  switch  on  JMP.   To  process  this  data,  I'm  using  a  project.   I'm  using  several  data  tables  and  it's  easier  for  me  to  put  them  in  the  same  place.   Before  to  go  on  the  testing  result,  I  just  want  to  talk  a  bit  about  my  mixing  plan.

I  told  you  that my  client  has  four  ingredients.   Unfortunately,  I  didn't  make  the  mixing  table  with  JMP.  Because  when  I  began  to  work  on  the  mixing  plan,  I  was  not  really  confident  enough  with  JMP  to  do  it  on  it.  The  client  gave  us  a  lot  of  rules  in  this  mixing  plan,  a  bit  complicated.  So  we  decided  to  make  it  by  hand  and  to  treat  the  rest  of  the  result  with  JMP.

Just  to  show  you,  this  is  my  mixing  plan.  I  have  a  code  for  all  of  my  samples  and  the  ingredients  one,  two,  three,  four,  and  the  proportion  of  each  one  in  these  samples.  There  is  just  few  information.  For  my  ingredients,  there  is  a  minimum  and  maximum  proportion.

The  important  thing  is  the  ingredients  works  two  by  two.  The  ingredients  one  and  two  are  working  together.  In  fact,  the  ingredient  one  plus  the  ingredient  two  is  always  equal  to  14 %  of  the  blending.  Exactly  the  same  for  three  and  four.  The  addition  of  these  two  is  always  equal  to  86 %  of  the  sample.

That's  few  words  given  by  the  client. T hanks  to  that  we  did  a  mixing  plan  with  16  samples  and  a  target.   The  target  is  the  historical  recipe  of  the  winery,  the  typical  wine.   The  client  wants  the  other  ingredients  to  be  closer  than  the  historical  wine.

You  can  see  here  the  mixing  plan.  It's  why  I  explained  just  earlier,  they're  working  to  pay  two.   Okay,  this  is  the  mixing  plan.   We  created,  we're  blending  the  samples,  and  we  did  the  tasting  with  the  client.

This  is  my  results  data  table.  It's  in  fact,  very  easy.  I  have  a  first  column  with  my  sample  in  the  wine  testing.  Most  of  the  time  you  have  to  test  without  knowing  which  is  the  modality  in  your  glass.   To  do  that,  to  recreate  a  random  number,  sorry,  a  random  number  of  three  digits  like  that  you  can't  know  which  sample  is  it.

I  put  it  on  my  first  column  and  after  that  I  have  one  column  by  tester.  In  this  case,  I  have  five  testers.  On  each  column,  I  put  the  group  where  the  sample  has  been  put.  Just  to  show  you  with  the  distribution  we  can  see  for  the  tester  one  in  the  group  three,  for  example,  just  for  the  tester  one.  He  put  in  the  group  three  the  sample  4 74,  486  and  910.  It's  the  same  for  all  the  samples.

I'm  not  sure  I  said...  Yes,  I  told  you  that  at  the  beginning.  I  asked  to  my  tester  to  describe  the  group  with  few  words.  When  I  do  a  testing  with  my  clients,  I  don't  write  on  my  result  data  table,  group  six,  group  one.  I'm  writing  directly  the  descriptor, the term  used  by  the  tester  to  describe  the  proof.  I  will  explain  you  why  a  bit  later.

I  have  this  data  table.  To  have  it,  most  of  the  time,  I  ask  to  my  tester  to  put  the  result  on  Excel  file  on  a  tablet  like  that.  He  put  directly  all  the  results  on  the  file  and  I  just  have  to  open  it  after  with  JMP.

I  need  another  the  data  table,  which  is  called  NUMMOD.  You  can  see  that  the  first  column  is  my  random  number  and  the  second  one  is  the  modalities. Y ou  can  see  what  modality  is  behind  the  number  given.  Then  the  other  column  is  the  description  of  each  modalities.  In  this  case,  it's  the  proportion  of  each  ingredient.

I  need  these  two  data  table  and  I  need  a  script.  To  process  the  data  directly  after  the  testing,  I  have  created  a  script.  For  this  script,  I  have  to  thank  a  lot  the  JMP  communities  because  they  helped  me  a  lot  to  do  this  really  complicated  part.

In  fact,  this  script  helped  me  to  create  the  distance  matrix  just  with  the  data  result  I  show  you  earlier.  In  this  case,  this,  I  will  not  explain  all  the  line  because  it's  a  bit  complicated,  but  I  will  show  you  how  I'm  using  it.  I'm  just  checking  I  am  on  the  right  data  table  and  I'm  running  the  script.

I  can  save  the  results.  Thanks  to  the  project,  I  can  save  the  result  directly  inside  the  folder  result.  Yes,  the  folder  result.   Directly,  I  can  have  my  distance  matrix.   You  can  see  I  have  still  my  sample  number  in  the  first  column.  Then  all  the  samples  in  column  and  the  distance  with  all  the  other  samples, so   for  the   001,  it's  the  same  sample,  so  it's  0.  Then  you  have  the  distance  with  the  other  samples.

In  the  script,  I  have  also  joined  the  information  from  my  data  table  in  the  map,  so  I  can  add  the  modalities  and  the  ingredient  proportion  in  the  same  data  table.  The  best  way  to  show  the  result  is  to  create  a  map.  To  show  the  map,  I'm  using  a  multivariate  method  and  precisely  the  multi dimensional  scaling.

In  this  case,  I  will  put  in  column  my  distance  matrix.  I  didn't  show  you,  but  I  have  grouped  directly  all  my  matrix,  it's  also  in  the  script.  Like  that  I  just  have  to  select  this  group  of  columns  to  put  inside  the  process.  I  add  my  distance  matrix  on  it.  I'm  running  it  and  I  can  have  this  map.

I  can  see  all  my  sample,  the  16  plus  the  target.  I  don't  know  which  one  is  it.  What  we  can  see  is  some  samples  are  really  close.  For  example,  the  246  and  the  592  are  really  close.  They  look  really  similar  for  all  the  taster.  Not  the  same  because  they  are  not  on  the  same  point.  There's  a  little  distance  between  them,  but  really  close.  At  the  opposite,  the  246  and  the  661  are  really  far  away  from  each  other.  They  look  really  different.

At  this  point  when  I  present  the  results  to  my  panel,  I  begin  to  show  which  sample  is  it.  We  can  talk  about  if  all  the  tester  are  agree  with  the  map. I f  they  say,  okay,  I  can  find  my  group  on  this  one.  We  can  talk  about  that and  I  show  which  sample  is  it.

For  that,  I  have  just  to  label  the  modality.  I  go  back  on  my  map  and  you  can  see  there  is  the  code  of  each  sample  of  the   mixing plan  and  most  important,  the  target.   You  can  see  the  original  recipe  is  here.  W e  can  say  that  there  is  some  sample  really  close  from  this  one.

I  think  this  one  should  be  interesting  to  use  with  all  the  ingredients  to  keeping  the  same  wine  style  of  the  target.  To  be  sure  of  that,  I  will  do  clustering  to  ask  to  JMP  to  show  me  which  sample  are  really  close  from  each  other.   For  that,  I'm  doing  a  clustering  and  more  precisely  a   [inaudible 00:19:04]   cluster.

A s  I  did  for  the  multi dimensional  scaling,  I'm  using  the  distance  matrix  as  [inaudible 00:19:15],  sorry.  I'm  running  it.  Usually,  I'm  testing  three,  four,  or  five  cluster  because  I  know  in  my  testing  it's  more  or  less  the  number  of  group  usually.  In  this  one,  I  already  know  that  three  cluster  is  the  best  way.  I'm  testing  three  and  I'm  saving  the  cluster  in  the  data  table  like  that.

I  can  put  in legend  the  row  state  of  the  cluster  and  the  map  will  be  colored  with  the  different  cluster.   You  can  see  we  have  three  cluster  really  well  separate.   One  looks  very  interesting,  the  green  one.  You  have  the  target  and  four  sample  really  close  of  the  target.

I  can  start  the  process  now.  I  can  say  to  the  client,  okay,  you  can  use  one  of  these  four  samples  from  the  mixing  plan  to  keep  the  same  quality  or  the  same  type  of  wine.  They  are  really  close.  Maybe  you  can  choose  this  one,  it's  the  closer  one.

But  if  I  want  to  give  more  information  to  the  client  about  where  it  can  play  inside  the   mixing plan,  I  did  another  treatment.  I  would  like  to  know  the  distance  between  each  sample  from  the  target.   For  that,  I  saved  the  coordinates  of  each  sample.  You  can  see  they  are  right  here,  the  dimension  one  and  the  dimension  two.  I  have  just  calculated  the  distance  between  the  target  and  all  the  others  in  sample.

To  go  a  bit  faster,  I  have  already  created  a  script  with  just  adding  a  new  column  and  a  formula  to  calculate  the  distance  between  the  target  and  the  sample.  I  will  just  running  it.  You  can  see  here  the  new  column  with  the  distance.

To  represent  the  best  part  of  the   mixing plan,  I  will  do  a  graph  builder.  A s  I  said,  the  sample  is  working  two  by  two.  I  can  represent  it  in  two  dimensions.  For  that,  I  will  put  the  ingredients  three  here  and  the  ingredient  one  here.  As  they're  working  two  by  two,  we  know  that  the  complement  of  the  ingredient  one  is  the  ingredient  two,  and  the  complement  of  the  ingredient  three  is  the  four.

We  don't  need  to  show  the  target,  so  I  will  hide  and  exclude  it.  I  have  my  16  sample  right  here.  I  will  put  the  distance  in  color  and  I  will  represent  it  with  the  contour  and  the  points.

To  be  easier,  I'm  just  changing  the  color.  I  will  take  this  one,  the  green,  yellow,  red.  Like  that  the  sample  is  close  from  the  target  with   the  shorter  distance  from  the  target  are  in  green  and  the  other  one  are  in  red.  I  don't  really  know  how  to  change  the  color  of  the  points.  We  don't  see  them  very  well.

You  can  have  this  type  of  mode.  That  is  really  interesting  for  the  client.  You  can  see  there  is  different  spots  in  green  and  different  spots  in  the in  red.  In  fact,  we  know  that  it's  not  interesting  for  the  client  to  playing  with  the   mixing plan  in  this  area.  It  doesn't  look  like  historical  wine.  It's  the  same  for  this  area.

But there  is  two  other  green  area.  This  one  there's  in  fact,  only  one  sample  really  close  from  the  target.  If  you  look  the  point  around,  then  doesn't  really  look  like  for  the  historical  wine, so  it's  not  really  interesting  to  play  in  this  area.

In  this  one,  it's  really  more  interesting  because  you  have  three  sample  really  close  from  the  target  and  two   other  one  a  bit far  away,  but  still  close.  W e  can  say  to  the  client  that,  okay,  if  you  want  to  keep  the  same  type  of  wine,  you  can  add  between  4  and  10 %  of  your  ingredient  one  and  between  20 %  and  60 %  of  your  ingredients  three.  The  most  interesting  is  to  keep  in  this  area  above  50 %  of  your  ingredient  3  and  around  7 %  of  your  ingredient  one.

With  this  information  our  clients  in  relation  with  the  age  is  volume  tank,  is  what  aging  you  want  to  do.  It  can  play  a  bit,  but  in  the  way  to  be  sure  to  keep  the  same  quality  and  the  same  type  of  wine.  This  helps  really  a  lot  the  clients.

We  can  do  another  treatment.  I  will  explain  you  quickly  because  it's  a  long  treatment  to  do.  But  in  this  case,  I  only  use  the  group.  It  doesn't  matter  if  it's  called  group  one  or  if  it's  called  Fruity,  Woody,  it  doesn't  matter.  It's  just  the  group.

But  I  asked  my  panel  to  describe  the  group.  In  this  case,  I  do  another  treatment.  From  the  data  table  result,  this  one,  I'm  doing  a  text  explorer  with  a  classic   JMP Pro .  I  can  have  this  type  of  data  table  with  my  samples  and  descriptor.  In  fact,  I  ask  him  to  count  how  many  times  each  descriptor  are  written  for  each  sample.  Like  that,  I  can  do  another  type  of  map  with  a  multi word  method,  but  this  one,  a  multiple  correspondence  analysis.

In  this  case,  I  will  put  in  response  the  descriptor,  and  in  factor,  the  modalities.  I  will  add  in  the  count  in  the  frequency.  Just  after  running  that,  I  just  will  show  you  with  the  script  because  we'll  see  it's  well,  there  is  a  better  presentation  in  this  way.

Okay.   You  can  have  this  type  of  map  with  in  blue  all  the  modalities,  all  the  samples,  and  in  red,  all   descriptor  used.   It's  a  complementary  map  from  the  first  one,  from  this  one.  From  this  one  because  in  this  one,  you  have  the  sample  close  or  far  away  from  each  other,  but  you  don't  know  why.  You  don't  know  why  this  sample  are  together,  or  why  this  sample  are  on  the  right  of  the  map,  and  why  this  one  are  on  the  left,  why  they  are  separated.

With  this  process,  we  try  to  explain  a  bit  why  the  sample  are  separated.  It's  not  always  exactly  the  same  map  because  it's  not  the  same  treatment.   This  one  needs  a  process  a  bit  longer  than  the  first  one  because  when  sometimes  you  don't  have  exactly  the  same  way  to  write  a  word,  whereas  in  French  we  have  accent,  so  sometimes  you  have  to  check  the  result  of  the  data  table before  to  do  the  process.

Some  words  are  more  or  less  the  same  sense,  so  you  have  to  put  them  together.  So  it's  a  bit  longer,  so  I  can't  do  it  right  after  the  testing,  but  I  do  it  after.  We  can  explain  a  bit  better  why  the  sample  are  located  on  this  way  on  the  map.

In  this  case,  you  can  see  some  sample  are  really  high,  coconut,  some  vanilla  nuts,  other  one  more  toasty,  spicy on  this  one.   Unfortunately,  some  are  really  negative descriptor,  so  you  can  explain  a  bit  better,  always  working  all  the  samples.  That  is  really  good  complementary  information  of  the   mixing plan  to  explain.  If  you  choose  to  go  on  that  side  of  the   mixing plan,  all   your  wine  will  be  described.   That's  it.

I  just  now  conclude.  I  hope  it  was  not  too  speedy.   For  us,  the  testing,  it's  a  difficult  exercise  to  modelize  and  to  represent  with  the  panel  we  used  because  it's  not  trend,  it's  not  a  big  one,  and  we  don't  have  the  same  target  when  we  begin  testing.

It's  why  we  decided  to  use  a  descriptive  testing,  not  quantitative,  the  free  sorting.   This  type  of  testing  can  be  only  thanks  to  JMP,  thanks  to  the  script.  I  can  do  all  the  process  really  quickly,  really  show  about  the  significance  of  the  results,  and  I  can  show  it  right  after  the  testing.

Like  that,  we  can  talk  with  all  the  taster  about  the  results.   When  we  leave  the  testing,  we  are  all  clear  with  the  wine  we  have  tasted  and  the  result.  It's  really  more  powerful  than  just  testing  with  some  weight.  W e  can  use  that  type  of  testing  with  a  small  and  untrained  panel.

Just  to  finish,  in  this  trial,  the  client  was  really  happy  with  this   mixing plan  and  it  can  adjust  the  recipe. I  know  the  recipe  is  working  since  two  years  with  the  four  ingredients  and  it  can  play  a  bit  each  year,  but  the  recipe  is  fixed  and  he's  really  happy  with  that.  Thank  you  very  much  for  your  attention.  Have  a  good  day.

Article Tags