Although the basic concepts of neural networks date from 1943, it is only in the last decade, with the discovery of new robust learning algorithms, that the research in the area of neurocomputing has boomed, resulting in a myriad of new and practical applications. Especially in the area of data analysis, with the ultimate goal to extract concentrated value from diffuse and intrinsically less valuable raw data, neural networks seem prone to success. In this context, neural networks have become mainstream technology in the domain of materials data.
Since about three years, research on the applicability of neural networks in the corrosion domain has been conducted at our department. Several networks were successfully trained to predict the corrosion risk of materials in various environments. They form part of the knowledge incorporated in an expert system for corrosion troubleshooting and intelligent risk determination.
In this paper, we will discuss some of the bottlenecks encountered when developing neural networks for data modeling. Tips will be given on how to find a good neural network architecture. However, as most of the failures of neural networks can be brought back to inappropriate data selection and representation, attention will mainly be focused on this topic. The problems that can be experienced will be illustrated by means of an example from the corrosion domain, and possible solutions will be given. Some basic rules to keep in mind during the development of a neural network will then be formulated.