<p>Microstructural characteristics have been demonstrated to dominate the damage process in dual-phase (DP) steels, and the role of each microstructural feature in the damage process has not been well characterized, since the microstructures in DP steels are exceedingly complex and these microstructural features are mutually coupled. In this study, a machine learning-based systematic framework was proposed to rank the microstructure features critical to ductile damage of DP steels and interpret their physical meanings. Finite element modeling was performed to generate the dataset that correlated the artificial microstructures with damage strain, and experiments of DP590 tensile testing were also conducted for calibration and validation purposes. 2-point correlation and principal component analysis were used to represent the diverse microstructures in a uniform format. Mutual information and random forest-based Boruta algorithm were chosen to filter and sort the microstructural features critical to the damage process. Microstructure reconstruction using Monte Carlo algorithm and individual conditional expectation (ICE) plots were employed to qualitatively and quantitatively interpret the physical meaning of these selected microstructural features.</p>