Polymer informatics · London, United Kingdom

Fewer trials.
Less scrap.
The same lab.

I'm Ravikumar Mohan — a polymer engineer who builds machine-learning models on formulation, processing and characterisation data. I help compounders and moulders reach specification in a fraction of the experiments.

The problem

Most formulation programmes spend their budget confirming what you already suspected.

01

Trial-and-error scales badly

Four additives at five levels each is 625 combinations. A full factorial is impossible, so you guess a subset — and the interaction that mattered was usually in the part you skipped.

02

Your instrument data is under-read

A rheology sweep gets reduced to one viscosity number in the report. The shape of the curve — the part that tells you about degradation and molecular weight distribution — is thrown away.

03

The knowledge leaves with the people

Twenty years of "we tried that in 2013 and it embrittled" lives in one engineer's head. When they retire, the next team runs the trial again.

What I do

Four ways I work with compounders and processors.

01

Experiment planning that stops at the answer

Sequential design of experiments and Bayesian optimisation applied to your formulation space, so each trial is chosen to be the most informative one left — not the next cell in a grid.

02

Characterisation into property prediction

FTIR, DSC, TGA, GPC and rheology traces already contain more than your reports pull out of them. I build models that map those curves onto the mechanical properties you actually specify against.

03

Scrap and failure root-cause analysis

When a grade starts failing impact or a line starts producing shorts, the cause is usually in data you already have. I connect process, material and QC records to find it.

04

Data readiness audit

A fixed-fee first engagement. I look at what you record, how it is stored and what could realistically be modelled — and tell you honestly if the answer is "not yet".

How an engagement runs
Step 1 · 2 weeks · fixed fee

Data readiness audit

I look at what you actually hold and tell you what can be modelled. You get a written assessment either way. If the honest answer is that you need to collect differently first, I will say so.

Step 2 · 6–10 weeks

Pilot on one real problem

One grade, one line, one failure mode. Narrow enough to finish, real enough to matter. You get the model, the code and a plain-English account of where it is unreliable.

Step 3 · ongoing or handover

Embed or hand over

Either I stay on retainer as the modelling capability you don't have in house, or I train your team to run and retrain it themselves. Both are fine. I have no interest in being permanently necessary.

Background

Ten years in engineering plastics before any of the modelling: engineering plastics compounding at Kingfa, advanced R&D atAvient, and senior materials engineering atBD, where I introduced Bayesian optimisation into high-throughput formulation work. Cambridge EnterpriseTECH fellow.

That order matters. The models are the easy part — knowing which polymer question is worth asking is what makes them useful.

More about my background →

Have a problem you have already thrown trials at?

That is usually the best kind. Tell me what you have tried and what data exists — I will tell you honestly whether modelling would help, before either of us commits to anything.