عنوان مقاله [English]
A pre-requisite for predicting the performance of any process is nothing but a study of the factors affecting the process and their interactions. Conventional methods, like the design of experiments, as well as unconventional methods, like artificial neural networks (ANN), are two major approaches for the discovery of such interrelations. Each of these approaches enjoys its own unique advantages in modeling a production process. A conventional method defines the variables based on, for example, statistical analysis, and, on the basis of the outcome, practitioners are well positioned to form their interpretations and draw their inferences about rocess erformance. Unconventional methods, in turn, have their own advantages. ANNs, for instance, have such advantages as; simplicity of application, high degree of reliability in discovering complicated interactions among variables, and last, but not least, being inexpensive as a practical method.There are reports in the literature which are devoted to comparison of the performance of unconventional models against conventional ones. This paper is dedicated to such a cause, in the sense that it attempts to model the complicated spray drying process by the logistic regression approach (a conventional method), as well as ANNs (an unconventional method), in order to compare the performance of these methods in predicting (by interpolation and extrapolation) process performance.Once the conceptual model of the spray drying process is developed, the model building process for the logistic and ANN is described through the following steps: a) designing the model architecture; b) data collection and processing; c) defining the model structure; d) selecting the right criteria for fitting the model; e) estimating the parameters of the model; f) verifying the model; g) selecting the right criteria for model reliability; and h) evaluating model reliability. The logistic and ANN models are fitted by a set of 100 data values, and are, subsequently, tested and evaluated by another set of 30 data values.Based on the results, in terms of the coefficient of determination and the percentage of correct predictions, it can be concluded that the ANN model demonstrates a relative edge over the logistic model in predicting process performance. It is obvious that there is room for sharpening this edge by increasing the number of test data. By establishing the superiority of one method over another in predicting process performance, one may define and investigate various scenarios, in order to arrive at conditions under which the input variables are so tuned that the quality of predicting the process performance is desirably enhanced.