<p>The compressive strength of sandstone is significantly affected by hydro-chemical corrosion, and current analytical and empirical models struggle to predict it accurately under varying conditions of hydro-chemical corrosion. While experimental methods are effective in clarifying the interactions between hydro-environmental factors, rock intrinsic properties, and rock strength, they tend to be both expensive and time-consuming. This research explores the potential of machine learning (ML) models to address these challenges. Six ML models were employed, including SVM, BPNN, RF, XGBoost, LightGBM, CatBoost. These models were applied to a dataset comprising 195 data points from previous studies involving the compression tests of sandstone pre- and post- hydro-chemical corrosion. The dataset encompasses nine input variables, including the content of feldspar, mica, clay minerals and calcite in the sandstone samples, porosity, solution pH and concentration, soaking time, and confining pressure, and one output variable, which is the ratio of compressive strength pre- and post- hydro-chemical corrosion. Four metrics, the coefficient of determination (R²), the root mean square error (RMSE), the mean absolute error (MAE), and the a20-index were used for the evaluation of model performance. The results showed that the CatBoost model demonstrated superior performance, achieving an R² of 0.957, an RMSE of 0.031, an MAE of 0.022 and an a20-index of 0.97 for the test dataset. Feature importance analysis indicated that pH value and soaking time were the most critical input variables for accurately predicting compressive strength, which indicated that ML models were highly dependent on hydro-chemical corrosion conditions for predicting compressive strength under hydro-chemical corrosion. Through SHAP analysis, it can be observed that the impact of the acidity/basicity of chemical solutions on the compressive strength of sandstone follows the order: acidic solutions &gt; alkaline solutions &gt; neutral solutions. Additionally, soaking time, feldspar content, porosity, solution concentration, clay mineral content, and calcite content are negatively correlated with the ratio of compressive strength, while mica content and confining pressure are positively correlated with this ratio. These findings align with physical laws, indicating that the machine learning model does not solely rely on data fitting but learns and predicts based on the underlying physical or chemical mechanisms inherent in the data.</p>