Feasibility of event-by-event primary mass discrimination using radio observables and supervised machine learning
In this work, we investigate the feasibility of event-by-event primary mass discrimination using radio observables only. Although the analysis does not require an explicit reconstruction of the shower maximum ($X_{max}$), the discrimination power still arises from the sensitivity of the radio observables to the longitudinal development of the extensive air shower (EAS). Such radio-based approaches could be particularly relevant for radio-only experiments, such as GRAND. To assess this feasibility, we obtained conservative upper limits for the discrimination accuracy using a supervised machine-learning (ML) algorithm, namely a random forest (RF). The input features used were the peak electric fields and the spectral slopes, which have complementary discrimination power, along with the antenna distances to the shower axis. The RF was trained and tested using large event sets generated by the fast radio emission simulation and simplified detector response implemented in the RDSim framework. We obtained discrimination accuracies between 81\% and 96\% over the studied zenith range, even after normalizing each shower by its own electromagnetic energy. Since the analysis includes deliberately conservative choices, such as a large 10\% uncertainty on the reconstructed EM energy, these quoted values should be interpreted as conservative upper limits suitable for a feasibility assessment. Our results demonstrate that event-by-event primary mass discrimination using radio observables is, in principle, feasible.