Title: Convolutional Neural Network-reconstructed velocity for kinetic SZ detection

Abstract: We report the detection of the kinetic Sunyaev-Zel'dovich (kSZ) effect in galaxy clusters with a 4.9 sigma significance using the latest 217 GHz Planck map from data release 4. For the detection, we stacked the Planck map at the positions of 30,431 galaxy clusters from the Wen-Han-Liu (WHL) catalog. To align the sign of the kSZ signals, the line-of-sight velocities of galaxy clusters were estimated with a machine-learning approach, in which the relation between the galaxy distribution around a cluster and its line-of-sight velocity was trained through a convolutional neural network. To train our network, we used the simulated galaxies and galaxy clusters in the Magneticum cosmological hydrodynamic simulations. The trained model was applied to the large-scale distribution of the Sloan Digital Sky Survey galaxies to derive the line-of-sight velocities of the WHL galaxy clusters. Assuming a standard beta-model for the intracluster medium, we obtained the gas mass fraction in R500 to be fgas,500 = 0.09 +- 0.02 within the galaxy clusters with the average mass of M500 ~ 1.0 x 10^14 Msun/h.