I tried to compute the optimal starter kit from the official blending suggestions, and got 16 materials of a single odour type

About a 5 minute read.

The last few articles have all been about taking fields apart. This one sets out to build something: use the official blending suggestions to compute a starter kit whose members go together.

Building the network

The database holds 673,662 blending suggestions, split between odour and flavour. Taking the odour half (336,574 rows) and keeping only pairs where both sides have ten or more suppliers and the name resolves back to a material record, then dropping direction and duplicates:

2,271 materials, 51,138 undirected links.

The most connected: ethyl cinnamate (671), cinnamyl alcohol (593), amyris wood oil (584), siam benzoin resinoid (558), benzyl salicylate (529), dodecanal C-12 (508).

That is the hub list. And I assumed it was the answer — buy the most connected ones.

It is not.

The hub list is not a good kit

Because hubs overlap. Two materials with 600 links each add nothing to the number of pairings inside your kit if those links go to the same crowd.

The question to ask is: if you buy N materials, how many pairs among those N are officially recommended?

I ran a greedy search, each step adding whichever material most increases the number of links inside the kit:

Materials	Pairings inside the kit
4	6
8	28
12	63
16	108

Buying the top 16 by connectivity gives 84. Greedy beats it by 24 pairings, a factor of 1.3.

"Buy the famous ones" is not optimal.

But the first nine give something away

Look at that growth curve: 0, 1, 3, 6, 10, 15, 21, 28, 36.

Those are the triangular numbers. Which means every material added connects to every material already in the kit — the first nine form a complete graph, all 36 pairs mutually recommended. I checked; not one is missing.

The nine:

Suppliers	Material	Odour type
70	Ethyl cinnamate	balsamic
79	Cinnamyl alcohol	balsamic
51	Amyris wood oil	balsamic
69	Siam benzoin resinoid	balsamic
53	Dodecanal (C-12)	aldehydic
19	Fir balsam absolute	balsamic
54	Raspberry ketone methyl ether	berry
38	Floral pyranol	floral
78	Sweet fennel seed oil	anise

Five of the nine are balsamic.

And the largest clique is more extreme

I ran four thousand randomised greedy passes looking for larger complete graphs. The largest found has 16 members:

maltol (107 suppliers), strawberry furanone (101), ethyl maltol (75), cyclotene (72), maple furanone (67), shoyu furanone (46), strawberry furanone acetate (41), ethyl cyclopentenolone (40), caramel furanone solution (37), immortelle absolute (34), fenugreek absolute (34), cyclotene hydrate (27), caramel pentadione (26), toffee furanone (25), ethyl furaneol (23), maltyl propionate (15).

Sixteen materials, every one caramellic.

This is not a formulation kit. It is one odour type recommending itself. Sixteen materials all pointing at each other, and broadly sixteen ways of writing the same thing — buy the set and you get one smell.

(Even the two naturals — immortelle and fenugreek absolute — are classed caramellic and fall into the same clique.)

So is the whole field an echo chamber?

No. Comparing the odour type at both ends of every link:

Links where both ends have an odour type	46,830
Same-type links	8,302 = 17.7%
Expected after shuffling the odour type labels	5.8%
Enrichment	3.1×

The same-type preference is real (3.1×), but 82.3% of links still cross odour types.

And the tendency to self-recommend varies enormously by type:

Same-type share	Odour type	Links
39.8%	citrus	8,032
30.7%	sulfurous	1,936
27.2%	floral	20,003
24.1%	fruity	7,806
22.4%	caramellic	2,707
15.0%	balsamic	9,469
13.6%	nutty	854
0.0%	powdery	206

Citrus recommends citrus most (39.8%). And not one of powdery's 206 links connects to another powdery material — powdery is a finishing quality, not a family you stack.

So how should this field be used

1. Do not pick materials by connectivity. Hubs overlap, and the densest regions are homogeneous.
2. Do not take a complete clique as a kit. That structure is biased toward a single odour type by construction.
3. The useful part is the 82.3% that crosses types. A citrus recommended alongside a balsamic carries more information than two citruses do.
4. If you are choosing a kit, fix odour-type diversity first and look at connectivity within that. My greedy search had no such constraint, which is how it fell into the caramellic hole — that is a fault in my method, not in the data.

I set out to compute a starter kit and computed the shape of the field instead. The shape is useful; it is just not what I asked for.

What this doesn't establish

• I used only the kind='odor' half; the flavour half is unrun.
• Names had to resolve back to a material record, and I measured how often that fails — so 51,138 is a floor, and the misses may be systematically biased.
• The ten-supplier threshold is mine. Loosening it grows the network and could change the conclusions.
• Greedy search guarantees no optimum, and the 16-material clique is not guaranteed to be the database's largest.
• The 17.7% baseline shuffles odour type labels, which assumes degree distribution is independent of odour type. That assumption may not hold.
• I have smelled none of these sets. This is network structure, not sensory verification — sixteen mutually recommended caramellics may be entirely sensible in a flavour context (they genuinely do get used together). What I am criticising is treating it as a perfumery starter kit.
• Each blending list is capped at 60 entries per kind, which I wrote about separately; that cap compresses the true connectivity of the most-connected materials.