Artificial Neural Network Modeling of DDGS Flowability with Varying Process and Storage Parameters

Abstract

Neural Network (NN) modeling techniques were used to predict flowability behavior of distillers dried grains with solubles (DDGS) prepared with varying condensed distillers soluble (10, 15, and 20%, wb), drying temperature (100, 200, and 300C), cooling temperature (-12, 0, and 35C) and cooling time (0 and 1 month) levels. Response variables were selected based on our previous research results, and included aerated bulk density, Hausner Ratio, angle of repose, Total Flowability Index, and Jenike Flow Function. Various neural network models were developed using multiple input variables in order to predict single response variables or multiple response variables simultaneously. The NN models were compared based on R2, mean square error (MSE), and coefficient of variation (% CV) obtained. In order to achieve results with higher R2 and lower error, the number of neurons in each hidden layer, step size, momentum learning rate, and number of hidden layers were varied. Results indicate that for all the response variables, R2 >0.83 was obtained from NN modeling. Compared to our previous studies, NN modeling provided better results than PLS modeling procedures, and fit (R2 >0.63), indicating higher robustness in the proposed NN models. Based on the predicted values from the NN models, surface plots yielded process and storage conditions for favorable vs. cohesive flow behavior for DDGS. Modeling of DDGS flowability using NN has not been done before, so this work will be a step towards application of intelligent modeling procedures to this industrial challenge.