In this research, we study four influence factors of the damping performance of ferrofluid dynamic vibration absorber, as well as predict and optimize the damping performance by machine learning method. The vibration absorber in our research is based on the second order buoyancy principle, which consists of a non-magnetic container, a small amount of ferrofluid and a permanent magnet. The effects of the initial amplitude, the cone angle of the cover, the thickness of the gasket and the mass of the ferrofluid on the damping performance are investigated by experiments. Based on the experiment data, we use BP neural network to establish a prediction model between the four influence factors and the damping performance. The prediction error of damping efficiency predicted by BP neural network is mainly within ± 0.4%. Meanwhile, the determination coefficient R2 of test data is 0.96242. The both indicate that BP neural network has a good performance in predicting the damping efficiency. Furthermore, we use the search algorithm to find the optimized values of each influence factor through the prediction model and the high damping efficiency is confirmed by experiments. Our work introduces machine learning into the field of vibration absorber designing, which provides an innovative method for the rapid design of high efficiency vibration absorber.