Four ways I work with polymer manufacturers.
All four start from data you already have. None of them require you to buy a platform, replace your LIMS, or run a digital transformation programme.
Experiment planning and trial reduction
The situation
You have a specification to hit, a formulation space with more dimensions than you can factorially explore, and a lab schedule measured in weeks per round.
How I approach it
I build a surrogate model of your formulation space from whatever historical trials exist, then use Bayesian optimisation to choose each next experiment — the one expected to tell you the most about where the specification is met. After each round the model updates and proposes again.
What you get
- A ranked list of the next trials to run, refreshed after each round
- The model and code, yours to keep
- An honest uncertainty estimate on every prediction
- Typical timeline
- 6–10 weeks for a first programme, then ongoing per round
- What I need from you
- Historical trial data — even messy spreadsheets. Compositions, process conditions and the measured responses.
Characterisation into property prediction
The situation
You run FTIR, DSC, TGA, GPC and rheology as a matter of routine, and the reports reduce each one to a single number. The rest of the curve is archived and never looked at again.
How I approach it
I treat the full trace as the input rather than a summary statistic, and model the relationship between what the instrument sees and the mechanical performance you specify against. This is where incoming-material variability usually hides.
What you get
- A model taking instrument output to predicted properties
- A clear statement of which inputs carry the signal and which do not
- Confidence bands, and the conditions under which the model should not be trusted
- Typical timeline
- 6–12 weeks depending on data volume
- What I need from you
- Raw instrument exports, not PDF reports, paired with the corresponding mechanical test results.
Scrap and failure root-cause analysis
The situation
A grade that ran fine for two years starts failing impact. Or one line produces twice the scrap of the identical line beside it, and nobody can say why.
How I approach it
Root cause is usually already recorded across process logs, incoming QC and maintenance records — just never joined up. I assemble those sources, look for what actually separates good runs from bad, and test the candidate explanations against held-out data rather than accepting the first plausible story.
What you get
- A ranked set of candidate causes with the evidence for each
- The analysis, reproducible, so your team can rerun it next time
- Where the data cannot answer it: what to start measuring
- Typical timeline
- 3–6 weeks
- What I need from you
- Process data covering both the good period and the bad, plus whatever QC and material traceability exists.
Data readiness audit
The situation
You suspect there is value in your data but do not want to commission a modelling project to find out there was not.
How I approach it
A short, fixed-fee assessment. I look at what you record, how it is stored, how it is joined, and what could realistically be predicted from it. This is deliberately the cheapest way to work with me, and it exists so neither of us starts a project that should not run.
What you get
- A written assessment of what is and is not modellable today
- The specific gaps worth closing, in priority order
- A recommendation — including, where warranted, "not yet"
- Typical timeline
- 2 weeks, fixed fee
- What I need from you
- A conversation with whoever owns the data, and read access to a representative sample.
Not sure which one fits?
Start with the audit. It is short, fixed-fee, and it is designed to tell you whether the other three are worth doing at all.