Solvent treatment is critical to improving the stability of halide perovskite materials that suffer from notorious issues that inhibit their industrial deployment; however, the complicated perovskite virtual design space with different types of solvent modifiers is inaccessible to traditional trial-and-error methods. In this study, machine learning is employed to predict stable multiple solvent-modified perovskite films under hostile conditions, and a complicated quinary solvent system "DMSO + DMF + toluene + NMP + GBL" is effectively identified to significantly improve the optoelectronic stability of CH 3 NH 3 PbI 3 in water. The "combinatorial solvent design" approach is realized by an extra tree machine learning model, which leads to a prediction dataset containing aqueous stability labels of 6720 new quinary solvent/perovskite systems. Importantly, the accuracy of the machine learning model is verified via photoelectrochemical experiments, achieving an experimental accuracy of 80%. A machine learning-predicted quinary solvent system offers significantly enhanced aqueous stability and 1000 times larger aqueous photocurrents, compared with the control CH 3 NH 3 PbI 3 film under the same hostile conditions. This study demonstrates the efficacy of machine learning for solvent design toward stable halide perovskite materials under hostile conditions.