Speaker
Description
Pulse-shape analysis (PSA) techniques are widely used to measure quantities such as energy, time as well as other valuable information including particle identification and position of interaction. In the case of large segmented high-purity germanium (HPGe) detector arrays, such as AGATA, analysing the pulse-shapes of the primary hit segment and its first neighbours is performed to determine all the interaction positions of a single $\gamma$-ray event. These positions are an indispensable ingredient for $\gamma$-ray tracking, used for Compton-background reduction and Doppler correction. Typical PSA on such arrays generally requires substantial computing resources to handle advanced sample-by-sample minimisation algorithms of large number of samples ($\sim$100 per pulse) in order to determine the interaction position(s).
In this work we present an alternative lightweight PSA approach using a reduced number of global pulse features ($\lesssim$15), easily extracted from the hit segment and its first neighbours pulses. The collected features are thenceforth injected in an ensemble of machine learning models based on a \emph{gradient boosted regression trees} algorithm to determine the hit position.Source data, measured with an AGATA crystal at the IPHC scanning table, provided the training and validation data sets. The first application of this algorithm demonstrated a position resolution comparable to those obtained with computationally expensive \emph{classic} PSA. This new approach requiring modest computational resources is typically, but not exclusively, useful for applications such as an ambulant Compton imaging device.
An overview of the construction and training of the machine learning models will be presented, as well as the achieved position measurement performances in the context of a novel HPGe Compton imaging device.