Subsurface heat islands induce an underground climate change in urban areas, which can threaten public comfort and health, subsurface ecosystems, transportation infrastructure, and civil infrastructure. Meanwhile subsurface heat islands harbor a marked energy recovery potential. Despite increasing investigations, the understanding of subsurface heat islands remains limited and suffers from the lack of expedient and accurate simulation approaches. Here we explore the use of machine learning to accurately and expediently simulate subsurface heat islands in terms of ground temperature and deformation anomalies. Using the Chicago Loop district as a case study, we identify a series of physical features to establish a relationship between central drivers and effects of subsurface heat islands. We incorporate these features into a random forest model to simulate underground climate change with variable training datasets. The results indicate that ground temperature and deformation anomalies across an entire city district can be predicted based on data extracted solely from a handful of buildings. The proposed approach achieves comparable accuracy to current simulation methods but boasts a calculation speed that can be over a hundred times faster, promising to advance fundamental science while effectively informing engineering and decision-making in the mitigation of underground climate change.