Evolutionary ensembles that learn spectroscopic characteristics of scintillation and CZT detectors

A method is described to automatically generate spectrum reference data for radioisotope identification devices respecting a detectors physical individuality. It extracts the peak shape and non-proportionality characteristics of scintillation and CZT detectors. The representation of these quantities is done with evolutionary ensembles, groups of N-dimensional autonomous points, which are propagated within a constrained space. Each ensemble member is used as parametrical input for describing peak shape and position within a simulation framework based on Geant4. Each subsequent generation of the ensemble iteratively converges the simulation result towards an optimised match with the measurement. Examples for the scintillator show that the shape convergence is straightforward due to the gaussianity of the peak, while the correction of the non-proportionality is within the quantity of up to 10%. Contrarily, our CZT example yielded nearly no non-proportionality along the energy scale, but required a complex, multi-parametrical shape definition with learning curves for kurtosis, skewness and resolution to establish an adequately peak reproduction. A metric is presented to calculate the distance between the experimental data and the calculated result. The described system is suited to establish a production line with a fully automatised acquisition of spectral characteristics to support the deployment of detector individual reference data for nuclide identification instrumentation.