Beginner perfumer #113: the real materials leaderboard, once the names are merged

Over four rounds I have accumulated three merge rules: strip the manufacturer parenthetical, look up the aldehyde C-numbers, and add twelve hand-written aliases. Coverage moves from an initial 54.9% to 81.98%. Those twelve hand-written lines recover 955 records; the database's own 95,214 aliases recover 14. Merging changes the leaderboard's shape. Looking up dipropylene glycol by the database's official name shows 32 formulas when the real figure is 470, a factor of 14.7. Styralyl acetate shows 26 against a real 135. Verdox, filed as green acetate, shows 16 against 75. More extreme still, some materials' official names never appear in the corpus at all: Iso E Super is filed here as patchouli ethanone, and searching that name returns nothing. The remaining 3,150 records split into two types, and 795 of them are a natural-material grading problem rather than a naming one.

About a 5 minute read.

The short version

Coverage accumulated across four rounds:

Method	Corpus component records covered
`name` only	54.9%
plus `head_synonym`	60.3%
plus stripping manufacturer parentheticals	72.8%
plus the aldehyde C-number table	76.5%
plus twelve hand-written aliases	81.98%

The twelve hand-written lines recover 955 records; the database's 95,214 aliases recover 14
After merging, dipropylene glycol's real usage is 14.7 times what the official name shows (470 against 32)
Some materials' official names never appear in the corpus. Iso E Super is filed here as patchouli ethanone
Of the 3,150 records left, 795 are a natural-material grading problem, not a naming one

Three rules, accumulated

The last three rounds handled one problem each:

87.9% of the alias field is truncated fragments, unrecoverable
Stripping the manufacturer parenthetical is worth 2,003 records, the only normalisation that pays
The aldehyde C-numbers need a lookup table; aldehyde c-14 is gamma-undecalactone

Running all three together gives 76.52%. I then read the remaining unmatched names and typed twelve lines:

dipg                     → dipropylene glycol
phenylethyl alcohol      → phenethyl alcohol
cinnamic alcohol         → cinnamyl alcohol
rose oxide               → laevo-rose oxide
damascenone              → beta-damascenone
cis-3-hexenol            → (Z)-3-hexen-1-ol
cis-3-hexenyl acetate    → (Z)-3-hexen-1-yl acetate
cis-3-hexenyl salicylate → (Z)-3-hexen-1-yl salicylate
linalol                  → linalool
styrallyl acetate        → styralyl acetate
methyl ionone gamma supreme → N-methyl ionone

Twelve lines, +955 records, coverage 81.98%.

For comparison: the database's synonyms field holds 95,214 aliases and adds 14 records. Twelve hand-typed lines are worth sixty-eight times that field.

The top 25 after merging

Rank	Formulas	Median dose	Material
1	470	10.00%	dipropylene glycol (solvent)
2	380	6.00%	linalool
3	323	10.00%	phenethyl alcohol
4	317	5.00%	benzyl acetate
5	227	2.22%	coumarin
6	220	5.94%	citronellol
7	207	5.56%	linalyl acetate
8	198	5.60%	geraniol
9	178	7.00%	methyl dihydrojasmonate (Hedione)
10	164	1.58%	eugenol
11	161	8.00%	hydroxycitronellal
12	161	9.47%	benzyl salicylate
13	157	8.00%	alpha-hexyl cinnamaldehyde
14	156	6.00%	dihydromyrcenol
15	154	5.97%	lilyall (Lilial)
16	149	1.48%	gamma-undecalactone (aldehyde C-14)
17	140	3.00%	heliotropin
18	137	1.00%	vanillin
19	135	1.00%	styralyl acetate
20	132	3.45%	patchouli oil
21	130	3.88%	alpha-terpineol
22	110	1.55%	geranyl acetate
23	110	6.00%	patchouli ethanone (Iso E Super)
24	107	1.94%	beta-ionone
25	100	0.50%	(Z)-3-hexen-1-ol

The top entry appears in 470 of 954 formulas: one formula in two contains DPG. That is a solvent, not a fragrance material. The actual backbone starts at number two.

How much the official name hides

Merged	Official name only	Factor	Material (what the corpus writes)
470	32	14.7×	dipropylene glycol (`dipg` 442)
135	26	5.2×	styralyl acetate (`styrallyl acetate` 78, `gardenol` 31)
75	16	4.7×	green acetate (`verdox` 49)
90	22	4.1×	2,4-ivy carbaldehyde (`triplal` 63)
79	26	3.0×	methyl cedryl ketone (`vertofix` 50)
178	84	2.1×	methyl dihydrojasmonate (`hedione` 165)

How to use this table: if you look a material up and the record says it appears in only 26 formulas, that number is probably an undercount.

Some official names score zero

The extreme case: the string in the database's name field appears not once in 954 formulas.

Merged formulas	Database `name`	What the corpus writes
154	`lilyall`	lilial
149	`gamma-undecalactone (aldehyde C-14 (so-called))`	gamma-undecalactone, aldehyde c-14
130	`(-)-alpha-terpineol`	alpha-terpineol
110	`patchouli ethanone`	iso e super
100	`(Z)-3-hexen-1-ol`	cis-3-hexenol
98	`cinnamyl alcohol`	cinnamic alcohol
82	`musk gx 50% in IPM`	galaxolide 50 ipm

Iso E Super is filed in this database as patchouli ethanone. It is one of the most-used woody materials in modern perfumery, and searching the name field for the name everyone uses returns nothing.

(The saving grace is that its head_synonym reads iso e super (IFF). That confirms the conclusion from two rounds ago again: the usable field is head_synonym, not synonyms.)

I missed a duplicate in #107

When I measured the dose distribution across 187 materials, the table listed both phenylethyl alcohol (49 formulas, median 15.00%) and phenethyl alcohol (274, 10.00%).

They are the same material, CAS 60-12-8, filed as phenethyl alcohol.

I caught the dipg / dipropylene glycol duplicate at the time and wrote it up, and missed this one, because the two spellings are close enough that I read past them without stopping.

The correct merged figure is 323 formulas at a median of 10.00%, third place.

The remaining 3,150 split two ways

Type	Names	Records
Natural materials (containing `oil`, `absolute`, `extract`)	168	795 (25%)
Everything else (the trade-name long tail)	834	2,355

The natural-material group is not a naming problem. The corpus writes ylang ylang oil (113 formulas); the database holds ylang ylang flower oil, plus grades I, II and III. That is not a spelling variant, it is the corpus naming things at a coarser grain than the database. Merging is a choice rather than a correction: collapsing four grades into one throws the grade away.

orange oil (52 formulas) is worse: the database has no plain orange oil, only orange oil terpenes, blood orange oil italy, orange oil replacer and so on.

The other group is the trade-name long tail: iralia total, fixolide, cetalox, isoraldeine 70, melonal. Those can be added to the hand alias table one line at a time. It just takes hours.

What this doesn't establish

The twelve hand aliases were typed while reading the residual list. I checked each against a CAS number or a name, but nobody else verified them and the list is certainly incomplete.
81.98% is a figure for these 954 formulas.
This ranks how many formulas a material appears in, not how much of it is used. Median dose is a separate column and they differ: DPG appears most often, while coumarin's median dose is 2.22%.
I did not re-check every material's median dose after merging. Where one material appears under two names inside a single formula — and I verified last round that this never happens for the eight C-number pairs — my rule is to add the amounts. That is reasonable, but I did not verify it across all 1,145 materials.
The "official name scores zero" table looks only at the name field. Those materials' head_synonym values are mostly right. I list them to make the point that one field is not enough.
Chemical name matching has its own literature; what I built is four rules and a hand-written table. Krallinger and colleagues' CHEMDNER corpus (PMID 25810773) and Akhondi and colleagues' identifier consistency study (PMID 23237381) belong to that field, not to this.