An ensemble learning-based surrogate model and a process parameter optimization method were proposed to improve the efficiency of solution mining design for salt cavern gas storage and reduce dependence on manual experience. A dataset containing 5324 samples was constructed using the self-developed single-well salt cavern storage cavity leaching simulation (SSCLS) software. The input variables included stratigraphic data and six-stage process parameters, while the output variables consisted of cavern volume and radius at key depth. Four types of tree-based models were used as base learners, and ridge regression was adopted as the meta-learner to establish the surrogate model. A design optimization workflow consisting of parameter generation, model prediction, deviation evaluation, and software verification was further developed. The test results showed that the mean absolute percentage error of volume prediction was 2.85%, and the mean absolute error of radius prediction for Jintan Well 52 was 2.17 m. A single prediction was completed within milliseconds, improving the computational efficiency by six orders of magnitude compared with conventional simulation. The proposed optimization method screened 830000 candidate parameter sets within 23 min and obtained an optimized scheme with a morphology deviation of 4.1%. The method can be used for preliminary scheme selection and rapid comparison of process parameters in salt cavern gas storage construction.