Liquid FLiBe molten salt is utilized as a coolant and moderator in molten salt reactors. The theoretical study of neutron moderation behavior is highly correlated with molecular dynamics simulations of materials. Accurate dynamic information is vital for better describing the scattering effect between neutrons and materials inside the core of a molten salt reactor. In this work, an on-the-fly machine learning force field of liquid FLiBe is trained for the first time in the Ab initio molecular dynamics simulation and employed for classical molecular dynamics simulation, replacing traditional potential functions. The quality of the MLFF obtained through training is reflected by Bayesian error estimation and root mean square error. The Bayesian error estimation is around 0.03 eV/Å, and the root mean square error converges to 0.06 eV/Å, both of which are within a reasonable range. Subsequently, atomic trajectories obtained from classical molecular dynamics simulations based on machine learning force field are used to calculate radial distribution functions of specific atomic pairs and diffusion coefficients of various elements in liquid FLiBe. These calculated results are in good agreement with those obtained from first-principles calculations, further confirming the accuracy of the force field compared to Ab initio molecular dynamics simulations. In addition, classical molecular dynamics simulations based on on-the-fly machine learning force field have significantly reduced time consumption compared to Ab initio molecular dynamics simulations. The phonon density of states is finally derived from the Fourier transform of the velocity autocorrelation function which is calculated based on atomic velocity information.