Error performance analysis of artificial neural networks applied to Rutherford backscattering

We have developed a code based on artificial neural networks (ANN) to analy se Rutherford backscattering data. In particular, we have applied the code to the analysis of germanium implants in silicon substrates. Here, we study the reliability and accuracy of the quantitative results obtained. We firs t constructed three different training data sets. The first data set was fu lly general. On the second one, we restricted the experimental conditions t o well-defined values, and on the third we also restricted the implantation parameters (depth and dose of implant) to a narrower range. We then studie d the trade-off between generality and accuracy of the ANNs obtained. Furth ermore, for a given architecture we applied two different training processe s. The first was backpropagation on the whole data set. In the second we ex cluded, after an initial training phase, all the training cases with errors double the average and then continued training. Each of the processes was applied to the three different data sets. We report the performance of the ANNs so obtained when applied to real experimental data. Copyright (C) 2001 John Wiley & Sons, Ltd.