Pulsed disk and doughnut column (PDDC) is widely applied in liquid-liquid solvent extraction. Due to a nonlinear and complex mechanism in PDDC, existing single empirical models often fail to predict the performance of different PDDCs. In this work, machine learning (ML) models such as random forest (RF), support vector machine (SVM), and artificial neural network (ANN) are developed to predict the PDDC's performance including dispersed-phase holdup (xd), drop size (d32), axial diffusion coefficient (Ec) and the height of mass transfer unit (Hoc). ML models were trained based on a comprehensive dataset and the results showed that the prediction performances of the ML models are better than the empirical correlations. The best average absolute relative error (AARE) and correlation coefficient (R2) of d32, xd, Ec and Hoc were 3.97% and 0.99, 10.16% and 0.955, 12.71% and 0.973, 13.44% and 0.982, respectively. RF and SVM exhibited the highest predictive accuracy. Furthermore, the feature importance was determined, which indicated the most significant features for d32, xd, Ec and Hoc were pulse intensity, the velocity of dispersed phase, the velocity of continuous phase and the properties of continuous phase, respectively. This study provided a new perspective to model and design PDDC.