This study proposes a fast simulation method that employs machine learning-based parameter calibration for three-dimensional (3D) rigorous simulation of defective masks in extreme ultraviolet lithography. The parameters of the structure-decomposed fast simulation model for defective mask diffraction are calibrated using machine learning methods, such as random forest and K-nearest neighbors, to improve the simulation accuracy and adaptivity. Herein, rigorous simulation is used as a benchmark standard for the calibration of model parameters. Simulation results of 50 validation contact masks set randomly reveal that the average simulation accuracy of aerial images is increased by 45% after calibration; both calibrated and uncalibrated fast models display better simulation accuracy (improved by 4.3 and 8.7 times, respectively) compared with an advanced single-surface approximation model. By applying defect-compensation simulation to a mask of 44-nm period, the simulation speed of single diffraction of the corrected fast model is ~97 times faster than that of the rigorous simulation when the simulation results are consistent (error is 0.8 nm).