The waste rock produced by mining pollutes the environment. However, transforming waste rock into backfill material can not only reduce pollution but also alleviate surface subsidence. The mechanical properties of backfill materials are crucial for surface protection. Therefore, in this study, a large-scale dataset based on gangue and tailings as backfill materials was established through experiments and collection. An ensemble learning model was developed to assess the nonlinear effects of 43 dimensional factors on the mechanical properties. Different backfill materials, preparation methods, and measurement errors can lead to significant differences in mechanical properties, which can easily affect the accuracy of evaluations. Hence, we proposed a heuristic adversarial perturbation method to enhance the model on differentiated data Through an iterative approach, an ensemble robust support vector regression model (ERSVR) was established. The model's robustness was studied under different integration levels, disturbance patterns, disturbance levels, and defense levels. This model can adaptively evaluate the mechanical differences of backfill materials in both coal and non-coal mining contexts. Compared to single machine learning models and conventional ensemble models, ERSVR has a mean square error of 0.05 and a correlation coefficient of 0.95, demonstrating better robustness and accuracy. This study plays a promoting role in establishing large models in the field of mining waste.

Mining waste, Backfill material, Ensemble learning, Robustness, Mechanical properties