Energy reconstruction of cosmic rays at large zenith angles using a combined neural network

In this work, we develop a combined neural network, CNN+MLP, to reconstruct cosmic-ray energy of events at large zenith angles detected by Square Kilometer Array of Large High Altitude Air Shower Observatory (LHAASO-KM2A). We use two sets of input features for neural network training, both of which are reconstruction parameters from LHAASO-KM2A. There are two steps: first, a CNN identifies the cosmic-ray composition; second, the results are passed to an MLP for energy reconstruction. The results from the neutral network demonstrate that the method achieves good performance in both energy resolution and bias; in the zenith angle range of 50°-60°, the overall energy resolution is better than 18% at 10 PeV, and the bias is limited within 5% for individual mass groups. We also carry out simulation test of the method by comparing the input true spectrum and the reconstructed one.