This study explored a strategy for predicting the proportion of martensite-austenite (M-A) constituents and impact toughness of stir zone (SZ) on X80 pipeline steel joints welded by friction stir welding (FSW). It is found that the welding forces, including the traverse force (F x ), the lateral force (F y ) and the plunge force (F z ), are the key variables related to the change of welding parameters and influence remarkably the characteristics of M-A constituents and impact toughness of SZ. The impact toughness of SZ is commonly lower than that of the base material due to the formation of lath bainite and coarsening of austenite. The characteristics of M-A constituents in SZ are sensitive to the variation of welding parameters and respond well to the change of welding forces. The proportion of small island M-A constituents increases with the decrease in rotational speed and the increase in F z . The increase in the amount of island M-A constituents is beneficial to improve the impact toughness of SZ. Based on the above findings, a machine learning (ML) model for predicting the M-A constituents and impact toughness is constructed using the force features as the input data set. The force data-driven ML model can predict the M-A constituents and impact toughness precisely and exhibits higher accuracy than ML built with welding parameters. It is believed that the high accuracy is achieved because the force features include more details of FSW process, such as the heat generation, material flow, plastic deformation, and so on, which govern the microstructural evolution of SZ during FSW.