0 Comments
As consumers recede from social eating to rely on satellite cooking and delivery, optimal food preparation has become increasingly vital. Using steamed dumplings - a simple, nutritious, yet tasty dish, multidisciplinary DSD DOE was utilized to prepare the most efficient recipe. Although a controlled dumpling experiment was initially designed using DOE, three runs had to be substituted with different meat/vegetable types after a shortage of ingredients. The design structure of the original DSD was found to be problematic as the categorical variable meat type was confounded with other main effects, and there were severe problems in Resolution IV (interaction terms were confounded with other interaction terms). To preserve the DSD structure, the original design was modified to better account for the substitution by changing meat type to a continuous variable. Before assessing model fit, the DSD was checked for orthogonality by looking at a variety of tools such as the Power Analysis, Color Map on Correlations (for interaction), Scatter Plot matrix with Nonparametric density (for uniformity), Prediction Variance Profile, Fraction of Design Space Plot, and Design Diagnostics. Since there were no major problems, the Response Surface Methodology (RSM) results would be attributed to scientific aspects of dumpling physics instead of problems in the data structure. Next, the optimal RSM model was selected using mixed Stepwise regression, and model robustness was probed using t-standardized residuals and global/local sensitivity (through various Profile tools). The relatively moderate r-square was attributed to the Run #6 outlier (which had a very long rising time due to its extreme factor settings - lowest water temp, highest dumpling weight, and highest batch size) as reflected in its high local sensitivity in the Prediction Profiler. The interaction plot also showed an interaction between temperature and dumpling weight, and the science behind convection and conduction was studied to explain and understand this interaction. This presentation demonstrates how a problematic DSD structure can be "saved" by carefully modifying important factor settings and later assessed by studying the design evaluation and the model results. For future studies, the DSD setting can be changed to study a problematic data structure, and more advanced modeling capabilities such as Neural Network, Bootstrap Forest, and Boosted Tree can be used to further improve and optimize the model.
|