Machine learning, which has been widely applied to scientific research in recent years, can be used to investigate the inherent correlations within a large number of complex data.

We evaluate the performances of two types of machine-learning algorithms for correcting nuclear mass models, reconstructing the impact parameter in heavy-ion collisions, and extracting the symmetry energy slope parameter. We also discuss the extrapolation and generalization ability of the machine-learning models.

For correcting the nuclear mass models, 10 characteristic quantities are fed into the LightGBM to mimic the residual between the experimental and the theoretical binding energies. For impact parameter or symmetry energy, two types of observables constructed based on the particle information simulated by using the UrQMD transport model for setting up the different impact parameters or symmetry energy slope parameters are used as inputs to a conventional neural network and the LightGBM to extract the original information.

Analysis of these nuclear physics problems reveals the potential applicability of machine-learning methods.

Machine-learning methods can be used to investigate new physical problems, thereby promoting the development of both theory and experiment.