Abstract: To determine the feasibility of parent rock for aggregates using tunnel slag, a method is proposed for predicting the crushing index of coarse aggregate from tunnel slag using four machine learning models: gradient boosting (GB), extreme GB (XGB), support vector regression, and random forest, with the parent rock properties of the slag as inputs. Additionally, an analysis of the importance of the parent rock characteristics of the slag and their influence patterns on the crushing index is conducted. The results show that: (1) The XGB model exhibits the best prediction effect, with an average R²  of 0.90 and root mean square error of 3.51% on the test set. (2) The impact of parent-rock properties of tunnel slag on the crushing index ranges from highest to lowest as follows: saturated uniaxial compressive strength (importance value 0.33), bulk dry density (0.28), water absorption (0.05), softening coefficient (0.04), mica content (0.01), and sulfide/sulfate content (0). (3) Increasing the strength of parent rock boosts the anticrushing ability of coarse aggregate from tunnel slag. High bulk density signifies hard, low-porosity rock with a dense structure, enhancing this ability. High water absorption, due to more pores in the rock, reduces the anti-crushing property. A high softening coefficient means less strength loss in water, indicating better anti-crushing performance. (4) Although mica and sulfides/sulfates in parent rock have minor effects on the crushing index, they can weaken the overall performance of concrete.

Key words: tunnel slag, coarse aggregate, crushing index, extreme gradient boosting model