Article ID | Journal | Published Year | Pages | File Type |
---|---|---|---|---|
1181500 | Chemometrics and Intelligent Laboratory Systems | 2012 | 6 Pages |
A new method is presented to model mixture data which simultaneously regresses the fractions of the materials used in a series of blends, and the matrix of the physical properties of the materials used in such blends to the properties measured from the resulting blend. The Weighted Scores Projection to Latent Structures (WSPLS) method combines the fractions of the used materials and their physical properties by first transforming the physical properties with a Principal Component Analysis (PCA) model and then estimating a matrix of weighted average scores using the fractions of the materials used and the corresponding scores for each material from the PCA models. This matrix of weighted scores is the regressor in the PLS model against the measured properties of the mixture. The new method is contrasted with other alternatives and shown to provide robust models with strong predictive components across all latent variables. A data set from blends of pharmaceutical powders is used to illustrate the features of the method proposed.
► Mixture data in a product development is different from a manufacturing one. ► Mixture models are keys to accelerate the development of new products. ► The WSPLS method is presented to model mixture data in product development. ► WSPLS is robust to the presence of missing data.