A waterborne protective coating needs better outdoor durability at a lower cost. Changing filler content can affect price, viscosity and film formation together. Acceptable initial adhesion may not persist after ageing. With limited preparation capacity, which formulations should receive the next tests?
This article works through how search software and experiments could be arranged, rather than reporting a completed customer trial. Start within the existing raw-material system: software proposes candidates worth measuring, the experimental team prepares and evaluates them, then the team decides whether to change the formulation, measure more or investigate preparation first.
Define the conditions under which performance matters
Before recommending candidates, bring formulation, preparation and evaluation conditions together. Results from different substrates, film thicknesses and ageing methods cannot go straight into a common ranking.
| Input | Information to retain |
|---|---|
| Raw materials | Identity, supplier grade, batch, available ranges and pricing basis |
| Formulation | Component amounts, wet or dry-solids basis, and mass units |
| Preparation | Addition order, dispersion conditions, drying process and actual deviations |
| Specimen | Identifier, substrate, application method, dry-film thickness and units |
| Evaluation | Adhesion method, ageing conditions, evaluation times and raw readings |
| Requirements | Mandatory thresholds and performance-cost trade-offs |
Application viscosity, material availability and adhesion can be screening constraints. Compare ageing performance and cost after those constraints are met. If the team has not decided how much extra durability is worth paying for, retain the trade-offs instead of rushing to combine them into a score.
Cost also needs a common basis of use. A lower price per kilogram may not mean a lower cost to coat the same area. Solids content, required dry-film thickness and consumption all affect that comparison.
Choose the next batch from formulations that can be prepared
First check non-negative proportions, consistent totals and permitted component ranges programmatically. Adding one ingredient means reducing another, so search directly within the constrained mixture space. Arbitrarily changing proportions and normalizing afterwards may also change ingredients intended to stay fixed.
Begin with a fixed material list and preparation sequence, arranging experiments around the baseline. If repeated historical results for one formulation already vary widely, independently remake the baseline and identify process conditions that must be recorded or controlled. Expanding the material list at that point asks the model to guess both chemistry and process effects.
Check whether a simple response model explains the available data before choosing a surrogate appropriate to the sample size. The model estimates properties and uncertainty; Bayesian optimization uses those estimates to allocate experiments. A batch can include both promising candidates and candidates that distinguish important unknowns. Evaluate uncertainty against held-out data too: a displayed interval is not automatically trustworthy.
Track samples in preparation and results not yet returned to avoid repeatedly recommending near-identical formulations. A language model can organize material documents and records. Programs calculate proportions, cost and constraints, while property values identify whether they came from measurement, calculation or prediction.
Include repeat experiments in the current round
Retain a baseline and necessary repeats in each batch. Prepare baseline and candidate specimens under comparable conditions, with randomized order or appropriate blocking. Keep material batch, thickness, drying and ageing conditions with each specimen.
Repeated readings of one panel primarily assess measurement variation. Independently prepared panels expose variation in weighing, dispersion and film formation. Label the two kinds of repeat separately rather than treating them as equally independent evidence.
Quick tests can exclude clearly unsuitable candidates before committing to long ageing tests. But the relationship between a quick indicator and long-term performance must be checked; it cannot directly establish outdoor service life. Evaluate models by independent preparation batch, keeping one specimen's records out of both training and validation. Reserve additional batches or conditions for final candidates.
After ageing, compare the baseline alongside the candidate
Suppose a candidate has acceptable initial adhesion but deteriorates after ageing. First compare its film thickness, preparation notes, test conditions and repeat results with the baseline. One low reading does not identify a particular ingredient as the cause.
If the difference recurs across independent batches under comparable conditions while the baseline stays stable, it can support updating the property model, lowering that formulation region's priority and testing nearby compositions to distinguish the responsible change.
If the baseline and candidates vary together across batches, or anomalies coincide with thickness and drying differences, pause broader search. Remake the baseline, observe the process more closely, or include an established process variable in the experiment. Otherwise, software may mistake preparation problems for formulation failure and discard a useful direction.
Keep unexplained records open for investigation; do not turn missing values into low scores. Retain raw data and reasons for revisions. Recheck model applicability after process changes instead of carrying forward the old ranking.
Give the next experiment an explicit reason
Deliver editable formulations, candidate comparisons, preparation and test plans, raw records and revision history. Each candidate should trace back to the data, prediction basis, unknowns and reason it was selected or rejected at the time, rather than displaying only a ranking recalculated today.
State which formulations met which measurement conditions, whether differences exceed repeat-test variation, and which batches or use environments still need checking. GOPROMP handles the agreed software and calculation workflow; the experimental team handles preparation and measurement. Stability is supported by repeats under those conditions, which determine whether to continue selection, investigate the process or independently test a candidate.