The feed to metallurgical reactors are often characterised by a variety of different ores, and in pyrometallurgical furnaces additionally by reductants. These feeds effect the operation of the reactors in various ways. In far example flotation plants the feed and other streams are characterised in terms of classes and sub-classes. This paper goes one step further in that it uses this class/sub-class classification to define empirical dimensionless feed numbers, which may subsequently be applied to establish optimal reactor operating regimes as a function of the feed material. Methods used to determine optimal operating regimes included neural nets, autoassociative neural nets and Sammon nets as well as distance weighted least squares smoothing of the data. This methodology was applied to metallurgical furnace data and the results clearly show that there exist functional relationships between the defined feed numbers and various other pertinent operational parameters. Therefore, this approach permits a very concise representation of industrial feed data to reactors at the same rime making furnace optimization possible. This paper also provides a practical comparison between the mentioned empirical data representation methodologies.;lt is also demonstrated where such an approach fits into a process control structure providing real-rime supervisory control.