Success of our ongoing energy transition largely depends on subsurface exploitation. The subsurface can act as a “battery” to store energy dense fluids such as hydrogen, or a “host” to sequester unwanted substances such as carbon dioxide or radioactive waste. On the other hand, these operations cause the subsurface pressure and/or temperature to change and induce various (or cyclical) loadings to the surrounding formations. Their operational safety crucially hinges upon the subsurface integrity. The most imminent risk is nucleation of cracks that can lead to loss of mechanical integrity. Unlike hydraulic fracturing in geoenergy applications where one deliberately initiates cracks at certain targets, we normally design a system to avoid fracturing. At the designing stage, we thus have no prior knowledge of crack initiation locations or propagation paths. And, the computational designing tools should be able to assess the fracturing risk without such prior knowledge. In this study, we compared three computational approaches that do not require prescribed crack geometries—the discrete element method, the lattice element method, and the variational phase-field approach—against percolation experiments on rock salt. The experimental results show different fracture propagation paths depending on the boundary loads. The fracture geometries were reasonably matched by all approaches despite some differences in path irregularities. While the variational phase-field approach predicts relatively regular fracture paths, the paths predicted by the discrete and the lattice element methods are more irregular. These irregularities may seem more comparable to intergrain failure in real rocks, but they are also necessary triggers for fracture initiation in the discrete and the lattice element methods. In contrast, the fracture initiation in the variational phase-field approach is a realization of the energy minimization in the system, and the grain level descriptions are absent in the current formulation. These findings highlight their predictive capabilities and gaps to be bridged between the grain and continuum scales for field-scale applications.