In order to determine the reasonable spacing of underground interchanges, a series of traffic simulation experiments with varying spacing cases was conducted in VISSIM using the Wenhui Street underground interchange of the Two Lakes Tunnel in Wuhan City as a practical case. The effects of interchange spacing, mainline and ramp traffic volumes and design speeds on crash risk at the diverging and merging areas of the underground interchanges were analyzed. Then, the crash risk variation at the diverging and merging areas of the underground interchanges was predicted using the four typical machine learning algorithms, i. e., extreme gradient boosting(XGBoost), support vector machine (SVM), random forest (RF), and multilayer-perceptron (MLP). Results show that when the spacing increases from 1.5 km to 2.5 km, travel time, average delay, average queue length, and traffic conflict rate decrease significantly, and the collision risk index of time-to-collision(TTC) increase significantly. When the spacing is above 2.5 km, the decreases in travel time, average delay, average queue length, and traffic conflict rate start to slow down, and so does the increase in TTC. When the spacing is 2.5 km and above, the overall traffic operation efficiency and safety of the underground interchange increase significantly. The XGBoost model can predict the crash risk variation reaching a precision of 88.3%. This study can be a theoretical support and practical case for the setting of underground interchange spacing.