Migration matrix performance

To test the performance of the MM, it was trained on the IBD prediction of one reactor cycle and applied to the prediction of a second cycle. The number of events generated per reactor cycle is roughly 2.5 to 3.5 million, including the scaling factor of 100 to smoothen the prediction.

Figure 4.8 shows the original distribution of reconstructed events for re-actor cycle 3-2018, based on the full simulation chain, compared to the result of the MM trained on reactor cycle 1-2018 and applied to the same set of events in true space, i.e. before they are processed with ROsim and the re-construction and IBD selection algorithms. For visualisation purposes, the distributions are projected in L_reco andE_reco separately.

The MM reconstruction inL_recomatches the expected distribution within statistical errors. Even for the lowest statistics bin, the relative difference is only 3%. The result in E_recoshows some more variations from bin to bin, but again stays within statistical errors and a 3% difference at maximum over the full range. We will see in chapter 7 that these differences are well within the statistical errors of the current data set.

Since the MM in principle only represents the detector and reconstruction effects, it should not depend on the shape of the input model on which it is trained. Therefore, the MM performance was also tested by training it on events generated with a flat energy spectrum and looking at the result of its application to a standard set of IBD events.

The resulting IBD distributions in reconstructed position and energy are shown in figure 4.9. A first problem with this MM was that the reconstruction efficiency - i.e. how many IBD events end up in the reconstruction compared

4.6. MIGRATION MATRIX 93

Figure 4.8: Performance of the migration matrix, trained on the events of reactor cycle 1-2018, in terms of the reproduction of the reconstructed distribution in energy (top) and position (bottom) for IBD simulation events of reactor cycle 3-2018.

94 CHAPTER 4. DETECTOR RESPONSE AND EVENT RECONSTRUCTION to the number of simulated interaction - was not well reproduced. This results in a different normalisation of the MM result compared to the original distri-bution. In figure 4.8, this difference was corrected for by rescaling the MM distribution to match the normalisation of the original one. For the scaled distributions, we see that the energy reconstruction is reproduced very well, but the length reconstruction shows an unexpected outlier of 4% in the first bin that can not be attributed to a statistical fluctuation.

To retrieve the origin of this issue, additional studies with other inputs for the MM training were performed. One is based on events generated with a simplified energy reconstruction model, using a simple shift of 1.8 MeV that represents the loss of the annihilation gammas and a Gaussian smearing of the energy. Further on, a point-like reactor model was used. This study shows similar, but more profound problems, as illustrated in figure 4.10. First of all the reconstruction efficiency does not match the expected one. After scaling, the reconstruction E_reco-shape is reconstructed very well, but the length distribution is poorly reproduced.

The studies shown here and additional ones that use different reactor geometries have so far not helped to find a clear solution for the problems described above. Some results indicate that the model has problems training on a data set where the L_true simulation is too simplistic. Further tests sim-ulating a finite reactor core with a shifted position will be performed in the near future, to estimate the effect of position reconstruction uncertainties on the MM model. In a second step, a different framework for the construction of the MM, called ReMU [126], will be implemented. This will allow a cross-check of the generated migration matrices and could help to understand the unexpected dependence of the MM performance on the input models.

4.7 Summary

To be able to perform a precise oscillation study, we need to understand all detector effects and how they affect the predicted signal. All steps in the detector response, from the particle transport and scintillation processes (SoLidSim, section 4.1) to the electronics performance (ROsim, section 4.2) are therefore simulated with dedicated Monte Carlo codes. These simulations are based on, and compared to, experimental data to realistically emulate all processes in detail.

Both for real and simulated data, the whole detection process shifts, smears and breaks up the recorded signals. As a consequence, algorithms that con-serve as much information as possible and correctly reconstruct the events of interest need to be applied. We have seen that for the SoLid experiment,

4.7. SUMMARY 95

Figure 4.9:Performance of the migration matrix when trained on a flat energy spec-trum, in terms of the reconstructed energy (top) and position (bottom).

96 CHAPTER 4. DETECTOR RESPONSE AND EVENT RECONSTRUCTION

Figure 4.10:Performance of the migration matrix when trained on events with a sim-plified energy reconstruction, in terms of the reconstructed energy (top) and position (bottom).

4.7. SUMMARY 97 these algorithms are implemented in the Saffron2 code. The four main recon-struction steps - selection, clustering, identification and reconrecon-struction - were briefly described in section 4.3. In section 4.4, the importance of the choice of energy estimator in the reconstruction process was discussed.

To highlight the actual sequence of steps, we have already briefly ad-dressed the next step of the data processing chain in section 4.5, which is the IBD signal selection. An in depth discussion on the selection methods is reserved for the following chapter.

Finally, section 4.6 of this chapter gave a description of the migration matrix, an object that is designed to map all relevant detector, reconstruction and selection effects at once, thereby simplifying the translation of the signal prediction from a model intrue spaceto an expected signal inreco space.

Signal selection 5

As was briefly mentioned in the previous chapter, the event reconstruction is followed by the process of signal selection, where the reconstructed event parameters are exploited to discriminate an IBD signal from the different types of experimental background.

We will first describe the properties expected for the IBD events, based on the study of Monte Carlo simulations and define the most useful parameters for their selection. Next, the background conditions of the SoLid experiment will be discussed. The two main sources of correlated background, induced by natural radioactivity and fast neutrons from atmospheric radiation, will be treated in more detail.

Using the knowledge of the signal and background properties and the general data taking conditions, a set of selection criteria or cuts can be de-fined to reduce the contribution of background noise as much as possible. Of course, these selection criteria also have a significant impact on the signal and it is thus a question of finding the optimal set of cuts to obtain the highest possible IBD detection efficiency and signal-to-background ratio.

Finally, the resulting IBD signal selection for the current data set will be presented.