Layer 1 was process mapping (boxes and arrows), and now we move into Excel: assumptions go into
cells, and you do arithmetic on them.
Start from the stoichiometry of the core “reaction” — what’s going in, what comes out, and in what
ratio. Sketch it on paper if that’s faster, then put it into cells. Build it left→right, so inputs
land on one side and products on the other.
We also use ‘stoichiometry’ and ‘reaction’ loosely — any process can be mapped on paper, whether it’s a
chemical reaction or making a pizza; the focus is to walk through each node step by step.
Set the basis & point the flow
Think in things per unit time — usually per hour or per day: how much you’re delivering, and how
fast each step is running.
Pick one basis and hold to it — per tonne of product, per hour, per 100 mol of feed. Mixing
molar with mass, or per-pass with per-tonne, gives numbers that each look sensible while the totals
are off
Anchor on the input; let the product float out the bottom — set the feedstock you can
actually buy, then let production be the number that falls out. Avoid back-solving the whole plant
from a “10,000 t/yr” target, because it’s harder to run the process flow logic in reverse.
Keep what you put in separate from what the model gives back — hard-code only what you
measured or sourced; let the model derive everything downstream
Parameterize each box
Write each unit op as input → mechanism → output: a reactor is a conversion and a selectivity,
a separator is split fractions, a compressor is a pressure ratio and an efficiency.
Find the one parameter that drives each box (the reactor’s conversion, the compressor’s
pressure ratio) and spend your sourcing there — a proxy is fine for the rest (ignore the chemical
engineering terms if it’s irrelevant and adapt the unit operation to what’s relevant for your technology)
Use the real at-scale number, not the lab or literature best case — if the paper says 97%
and plants run ~85%, model 85%. Also don’t take numbers directly from an AI tool; use it to surface a
source, then go find the datapoint from the the source itself
🧭 Coach’s Read
Your parameter values are the analysis, more or less — which is why the
too-clean ones (100% conversion, zero losses) are the first thing a diligent reader checks. Get a rough number on every box
before drilling into any one (get the model running end-to-end!). Find lots of datapoints for each assumption if you can.
Keep asking: what’s the next-most-important parameter I haven’t pinned down yet?