We propose a machine learning method to investigate the propagation of cosmicrays based on the precisely measured spectra of the primary and secondary cosmicray nuclei of Li, Be, B, C, and O from AMS-02, ACE, and Voyager-1. We train twoconvolutional neural networks. One network learns how to infer propagation andsource parameters from the energy spectra of cosmic rays, and the other network,which is similar to the former, has the flexibility to learn from the data withadded artificial fluctuations. Together with the simulated data generated byGALPROP, we find that both networks can properly invert the propagationprocess and infer the propagation and source parameters reasonably well.This approach can be much more efficient than the traditional Markov chainMonte Carlo fitting method for deriving the propagation parameters if userschoose to update confidence intervals with new experimental data.Both of the trained networks are available at (https://github.com/alan200276/CR_ML).