Prediction of baking results from farinograph measurements by using stepwise linear regression and artificial neuronal networks

In an effort to extract additional data from farinograph experiments a model was developed to simulate the measurements and correlate the parameters of the model with results from baking tests. This additional information can be used in bakeries to predict the baking properties of the flours and adjust the recipes to maintain a constant product quality. For this eight different flours were characterized with a farinograph and 13 different results from baking experiments. An approach with five nonlinear differential equations was able to model the farinograph measurements very well (average R2 = 0.995 ± 0.005). While a stepwise multilinear regression only showed weak correlations in cross validation between a single parameter of the model and the baking volume (R2 = 0.745) and the volume yield (R2 = 0.796) respectively, the artificial neuronal network was more successful. For the baking weight (R2 = 0.926), the dough yield gross (R2 = 0.909) and net (R2 = 0.913) strong correlations were found. A good correlation for the baking volume (R2 = 0.853) was also determined, while the volume yield showed comparable results to the linear regression (R2 = 0.792).