To study the relationship between composition and compressive strength of ultrahigh performance fiber reinforced concrete (UHPFRC), an interpretable data-driven approach is proposed. This approach addresses the high cost, complexity, time consumption and difficulty in studying the multivariate dependence effects of design parameters with traditional experimental methods. To construct the prediction model, 400 experimental data points are leveraged to compose the dataset, and five GridSearchCV (GSCV)-tuned ensemble machine learning (ML) approaches are evaluated. The GSCV-CatBoost model, with its higher prediction accuracy and lower error, i.e., R2 = 0.945, RMSE = 8.567 and MAE = 5.084, is proposed as a robust tool for predicting the compressive strength. Furthermore, the results of the SHAP analysis reveal that concrete age and steel fiber content are the most important features for predicting the compressive strength and those of the multivariate dependence effects analysis show that an increase in FRC will reduce the positive effect of the L/B on the compressive strength when the L/B exceeds 1. The feature importance and size effect of the specimen are noteworthy for material designers in the selection of suitable raw materials to achieve the desired strength properties. Finally, a reverse design method based on the abovementioned approach is proposed for the development of new UHPFRC materials. To maximize the effect of the model, the application range of the model is limited to the target compressive strength of 50–232 MPa and includes basic material parameters such as cement strength and silica fume, sand, steel fiber, water, and superplasticizer content.