An analysis of effects based on a statistical model was performed with a set of 40 data of yield of protein isolate (Y) as a function of five different factors from an industry of production of soybean protein isolate (SPI). At a probability level higher than 0.925, only the temperature in the desolventizer (T), and pH of the first tank of protein extraction (pH1) had a significant influence on Y. Based on this evidence, an artificial neural network model was suggested to better correlate the yield of protein isolate with the revealed significant variables (i.e., T and pH1). Multilayer perceptron networks with three to seven neurons in the one hidden layer were tested, but the best architecture involved five neurons. Over‐fitting was observed when a number of neurons larger than five was considered. The relative average deviation between measured and calculated yields was reduced from ≈3.9 to ≈1.7% when the neural model instead of the statistical one was used. A second set of 25 data of Y from laboratory experiments mimicking commercial SPI production at the designated plant confirmed the importance of T and pH1 on the examined response.
Practical Applications
The yield of protein isolate in industrial plants is dependent on many factors, which makes difficult to set and keep the process under optimal operating conditions. However, the current examination shows that the dynamic control of only the desolventizing temperature of defatted soy flakes and pH of protein extraction, based on a trained artificial neural network model, is a valid real‐time optimization strategy.