Predictive Modeling of Material Properties Using GMDH-based Abductive Networks

Material properties are very important in most material science and engineering computations. A number of modeling and machine learning techniques have been used for the prediction of material properties, including Fuzzy Regression, Adaptive Fuzzy Neural Network, Extreme Learning Machine, and Sensitive Based Linear Learning Method. This paper proposes the application of Abductive Networks to the problem. We studied the performance of various Abductive Network architectures on a dataset used by earlier published work. A Root Means Square Error (RMSE) as low as 15.34MPa was achieved on the predicted tensile strength values, which represent about 50% improvement compared to the performance reported in the literature for other modeling techniques on the same dataset. Moreover, the technique achieves 20% reduction in the number of features required.