PLS
What Is PLS Analysis? Predicting Quality Before the Batch Ends
By J. de Vries · · 7 min read
Predicted vs. measured quality for a PLS model (illustrative data).
What is PLS?
Partial Least Squares (PLS) regression relates a block of process inputs (X — temperatures, flows, raw-material properties) to one or more quality outputs (Y — yield, purity, tablet hardness). Unlike ordinary regression, it is built for exactly the data a plant produces: many variables, strongly correlated, often more columns than batches.
Why not ordinary regression?
Multiple linear regression falls apart when predictors are correlated — coefficients flip sign, blow up, and stop meaning anything physical. Process data is almost nothing but correlated predictors. PLS handles this by first compressing X into a few latent variables (much like PCA does), then regressing Y on those — so the model stays stable and interpretable even with hundreds of tags.
Predicting quality before the batch ends
Once a PLS model is trained on historical batches, it can score a running batch from the measurements available so far. If the predicted purity is drifting out of specification at hour six of a twenty-hour batch, you still have fourteen hours to correct course — or to stop wasting raw material on a batch that can't be saved. That's the difference between a lab result that arrives after the fact and a prediction you can act on.
VIP scores: which variables actually matter
A PLS model also ranks its inputs. VIP scores measure how much each process variable contributes to predicting the quality outcome; a common rule of thumb is that variables scoring above 1 matter. If reactor temperature and feed rate carry the model while stirring speed contributes nothing, you've just learned where process control effort pays off — and which expensive sensor you could stop arguing about.
How do you know the model is any good?
By testing it on batches it hasn't seen. Cross-validation holds out part of the data during training and checks predictions against it — the honest version of the predicted-vs-measured plot above. ProcessLens runs cross-validation automatically and reports the validated R², so a model that merely memorized its training batches doesn't sneak into production use.