Inductive learning of materials performance indices from a material properties database

Assessment of materials performance in engineering design and manufacturing is a complex process that requires significant knowledge of scientific and engineering principles. In a typical design application, for instance, determination of how a material's performance indices such as fatigue resistance, corrosion resistance, and formability are affected by its physical properties is a laborious task. Expert system technology has been able to partially address this issue by transferring domain specific expertise and heuristics from human experts into knowledge based systems that can precisely communicate materials knowledge to their users. Traditional expert systems, however, are limited from the standpoint that acquisition of their needed operational knowledge from human experts, who may not even exist in some particular applications, is a time consuming process which is at best subjective and prone to error. In this paper, we present a new approach where ID3-type decision tree classifiers extract the material selection knowledge from a properties database and present it to designers in the form of generalizations. One of the important features of decision trees is their ability to deal with unknown and/or imprecise training information. We also show how receiver operating characteristic (ROC) curves can be used for describing the performance of a parametric family of decision tree classifiers.