A deconvolution procedure for ESR Spectra

Quantitative determination of spin contents from complex spectra requires deconvolution of each peak. There are several ways to perform this task, the more common being the use of a non-linear fit such as the Marquardt-Levenberg algorithm (M-L). However this technique although very powerful and largely diffused fails when the initial set of parameters of the peaks are far from the true ones, i.e. when the sample composition is not a-priori completely known. In this case a new run with different initial values is required. This isn't convenient for an automatic deconvolution procedure. Here a new procedure is presented, based on a genetic algorithm as a first step and a M-L method as the second step. The first step will always work regardless of the initial set of data but the accuracy obtainable is lower compared to the M-L technique. The accuracy is improved by using the parameter obtained in a subsequent algorithm employing the M-L method. However, due to the precision of the genetic algorithm presented here, the second step can sometimes be omitted. Results of some trials both on integral and differential spectra are reported and discussed here.