AUTOMATIC FITTING OF GAUSSIAN PEAKS USING ABDUCTIVE MACHINE LEARNING

Analytical techniques have been used for many years for fitting Gaussi an peaks in nuclear spectroscopy, However, the complexity of the appro ach warrants looking for machine-learning alternatives where intensive computations are required only once (during training), while actual a nalysis on individual spectra is greatly simplified and quickened, Thi s should allow the use of simple portable systems-for fast and automat ed analysis of large numbers of spectra, particularly in situations wh ere accuracy may be traded for speed and simplicity, This paper propos es the use of abductive networks machine learning for this purpose, Th e Abductory Induction Mechanism (AIM)(1) tool was used to build models for analyzing both single and double Gaussian peaks in the presence o f noise depicting statistical uncertainties in collected spectra, AIM networks were synthesized by training on 1000 representative simulated spectra and evaluated on 500 new spectra, A classifier network determ ines the multiplicity of single/double peaks with an accuracy of 98%, With statistical uncertainties corresponding to a peak count of 100, a verage percentage absolute errors for the height, position, and width of single peaks are 4.9, 2.9, and 4.2%, respectively, For double peaks , these average errors are within 7.0, 3.1, and 5.9 %, respectively. M odels have been developed which account for the effect of a linear bac kground on a single peak, Performance is compared with a neural networ k application and with an analytical curve-fitting routine, and the ne w technique is applied to actual data of an alpha spectrum.