With the construction of RNO-G and plans for IceCube-Gen2, neutrino astronomy at EeV energies is at the horizon for the next years. Here, we determine the neutrino pointing capabilities and explore the sensitivity to the neutrino flavor for an array of shallow radio detector stations. The usage of deep learning for event reconstruction is enabled through recent advances in simulation codes that allow the simulation of realistic training data sets. A large data set of expected radio signals for a broad range of neutrino energies between 100 PeV and 10 EeV is simulated using NuRadioMC. A deep neural network is trained on this low-level data and we find a direction resolution of a few degrees for all triggered events. We present the model architecture, how we optimized the model, and how robust the model is against systematic uncertainties. Furthermore, we explore the capabilities of a radio neutrino detector to determine the flavor id.

We present an end-to-end reconstruction of the neutrino energy, direction and flavor from shallow in-ice radio detector data using deep neural networks (DNNs). For the first time, we were able to determine the neutrino direction with a few degrees resolution also for the complicated event class of electron neutrino charged-current interactions where the shower development is impacted by the LPM effect. This result highlights the advantages of DNNs to model the complex correlations in radio detector data. We will present an outlook of extending the model to predict the complex probability distribution of the neutrino direction using Normalizing Flows. Furthermore, we discuss how this work can be used for real-time alerts and an end-to-end detector optimization of, e.g., IceCube-Gen2 radio.