A novel form of calibration of online analyzers was implemented. Rather than simple equations, a neural network was used to model the relationship between the scintillation counts (Am and Cs) of an analyzer and the measured ash for improved online analysis of run-ofmine coal. Also, a new approach was followed to better evaluate neural network performance. Samples were first divided into various statistically different groups using a Kohonen network. Data were then selected for the training, calibration and prediction subsets using criteria developed in this paper for sparse data, with representation from each group. Back propagation-based neural network architecture was used in conjunction with quick-stop training. The predictions were very good on average, but due to noise in the data, the predictions were not good individually.
