Having computed a new best limit on the axion-electron and chameleon couplings, as well as presented a limit on the axion-photon coupling, let us think about the next steps that should be taken from here.

The first obvious step would be computing a combined limit of the dataset from this thesis with the GridPix data from 2014/15 (Christoph Krieger 2018). As the old detector used the same readout system, the entire analysis framework used in this thesis is directly compatible with the old data. This would be straightforward and should lead to a decent improvement on the limit, in particular due to the good expected limits for the MLP classifier at high software efficiencies, as there are no vetoes available for the old detector.

Secondly, a combined limit of all GridPix data with all Micromegas data should be computed. At least at the likelihood level – meaning to multiply the posterior likelihood functions for each detector and dataset – would be little work and could lead to another limit improvement. While this would imply exclusion of systematic uncertainties for non GridPix data, in theory one could even attempt to compute limits with the MCMC code written for this thesis. GridPix and Micromegas detector differ, but fundamentally at the level of the limit calculation both work with cluster center positions and energies. This would require the corresponding systematic uncertainties and detection efficiencies.

Focusing on a new detector, there are two major improvements on the horizon. First of all a future detector is likely going to be built from radiopure materials. This should reduce the inherent background seen in the detector significantly. In addition, such a detector will be built on top of the Timepix3 instead of the Timepix. This comes with massive improvements in terms of achievable background rates at higher signal efficiencies. The Timepix3 can be read out in a stream-based fashion. Therefore, there is no more dead time associated with a readout and long shutter times are a thing of the past. More importantly, this means for a 7-GridPix detector layout, the very high random coincidence rates seen for the current detector (see sec. 12.5.5) are completely removed. For a Timepix3 based detector the septem veto and line veto can be used without any efficiency penalty. Combined with an MLP classifier as used in this thesis and without considering radiopure materials, background rates in the \(\SIrange{5e-6}{7e-6}{keV^{-1}.cm^{-2}.s^{-1}}\) range are achievable at software efficiencies around the \(\SI{95}{\%}\) mark (see tab. 26 and tab. 28). This alone would lead to a significant improvement in limit calculations. Furthermore, the Timepix3 supports reading out the Time over Threshold values in addition to the Time of Arrival (ToA) values. At that point the FADC is not needed anymore to provide time information nor to act as a readout trigger. All noise related issues due to dealing with analogue signals in such an experiment are side-stepped. Further, the availability of pixel-wise ToA values allows to reconstruct events in three dimensions, making the equivalent of the rise time veto for the FADC much more accurate. Indeed, in a prototype Timepix3 based detector utilizing an upper cut on the ToA values of clusters already shows promising results (Schiffer 2024).

One positive aspect of a new Timepix3 based detector is that the full analysis procedure explained in the context of this thesis is already written with Timepix3 based detectors in mind. While some additions will likely be desired, the basis is already there and given a dataset for a Timepix3 based detector a limit calculation could be done in a matter of days. ¹

In terms of a possible setup of such a Timepix3 based detector at BabyIAXO in the future, a few things are important. Firstly, a much improved scintillator veto setup will greatly reduce the possible background due to muons or X-ray fluorescence. One thing to consider for this though, is to move away from a purely discriminator based system. The scintillators used in this thesis were much less useful than initially hoped, due to lack of interpretability of the data and imperfect calibrations. Ideally, the analogue signals of each scintillator are read out so that discriminator thresholds can be applied offline during data analysis (while this introduces analogue signals back into the system, the highly amplified nature makes this significantly easier to deal with). This makes the system much more flexible and less error prone in terms of choosing the wrong parameters for a data taking campaign.

Next, given that a Timepix3 based detector will surely need a cooling system as well, time should be invested in characterizing its performance and possibly even working towards a temperature stabilized system. Generally, more sensors should be installed, both for temperature measurements in different places, as well as pressures to make correlating detector behaviors with external parameters easier.

On the data calibration side, more importance should be placed on measurements behind an X-ray tube. Data quality in the CAST Detector Lab (CDL) data for this detector was unfortunately not ideal and generally more statistics should be available. While probably unreasonable, BabyIAXO should have a calibration source installed on the opposite side of the magnet, which can produce different X-ray energies. This would allow calibrations both through the X-ray telescope for more frequent alignment measurements as well as X-ray calibration data taken at the same conditions as the background and tracking data. The discrepancy between CAST data detector behavior and CDL data is larger than I would like.

Moreover, simulations of a future detector should be higher priority in the future. Both in terms of general background studies and electric field simulations, as well as event simulations. Having more realistic synthetic datasets would be invaluable for better developments of classifiers without the need to rely on real datasets. This was already very successful with the comparatively simple synthetic X-rays in this thesis, but could be extended for background data in the future for pure synthetically trained machine learning classifiers. In this vein more powerful approaches would become feasible, especially if more calibration data were available as well. Convolutional Neural Networks are a promising candidate to investigate, especially also for the 3D cluster data of a Timepix3 detector.

And hey, maybe soon we can also compute a limit on the Axio-Chameleon (Brax, Burgess, and Quevedo 2023)?