Pulse shape analysis algorithms for decay spectroscopy of short-lived nuclei
The decay spectroscopy of exotic nuclei is a central pillar of the Nuclear Physics Group’s research programme. The experimental method employed at laboratories such as Jyväskylä and GSI/FAIR is to separate the nuclear reaction products of interest using an electromagnetic separator and implant them into a double-sided Si strip detector (DSSD). The DSSDs have an orthogonal strip architecture that effectively provides >5000 pixels and the data analysis involves reconstructing the histories of all ion implantation and decay events occurring independently in each pixel. The rare nuclear decays of interest are then extracted from these histories, allowing their properties (decay energies, half-lives and branching ratios) to be measured.
In the drive to study ever more exotic nuclei, it has become necessary to instrument the DSSDs with digital readout electronics. This allows lifetimes down to ~150 ns to be accessed compared with the limit of a few ms with conventional analogue electronics. In these measurements, the preamplifier output waveforms are digitised by typically sampling 14-bit ADCs at 100MHz and stored for offline analysis. The event rates in the DSSDs can be as high as ~10 kHz, so over the course of a typical 2-week long experiment many billions of events are recorded and the total volume of data comfortably exceeds 10 TB.
One problem with this experimental method is that the DSSDs suffer severe radiation damage from the implantation of the ions. This leads to charge trapping and consequently the preamplifier pulse shapes change constantly during the experiment. Since the implantation profile of the ions is not uniform, the pulse shapes change at different rates in different pixels. In order to obtain the most precise timing and pulse height (energy) information from the digitised pulses it is necessary to have the best possible knowledge of the pulse shape response for each DSSD pixel at every point in time.
Several data sets of this type have been analysed to date. Different analysis methodologies have been employed, but owing to the complexity of the challenge and the data volumes involved, even the most sophisticated algorithms have only attempted to analyse pulse shapes at the per strip level, rather than the per pixel level. The analyses took students many months (often much more than a year) to implement, so developing the capability of analysing the pulse shapes from such data sets using machine-learnt algorithms offers the prospect of significant gains in efficiency and performance that will directly benefit the Group’s future programme and give the student high-level skills in data science as well as experimental nuclear physics. The aim of the thesis will be to develop the machine-learning algorithm and to demonstrate enhanced performance.