Statistical Analysis of Large Pilot Plant Datasets

Dow AgroSciences has supported two pilot plant facilities for development and scale-up of new batch agrochemicals for several years. One of the pilot plant facilities consistently obtained higher yields and lower run-to-run variability. An improvement project using recently developed data mining techniques and Six Sigma methodology was initiated to resolve the operating differences between the two plants. The project goal was to decrease the coefficient of variation (CV) of the pilot plant B to 6% and improve the mean yields to within 95-100% of pilot plant A. Therefore, pilot plant datasets were acquired and analyzed using two multivariate statistical techniques, Principal Component Analysis (PCA) and Partial Least Squares (PLS). These techniques identified variability both between batches and across time for each batch. The most significant input variables for overall variability were three raw materials and one operating condition. Based on this information and operator experience, experiments were run at B resulting in lowering one feed rate and raising another. Because of these improvements, Dow AgroSciences discontinued the use of plant A, which resulted in significant savings per year. Even greater savings has been leveraged into manufacturing from less raw material use.