Mining and beneficiation of underground solid potassium salt deposits generate large quantities of solid tailings and old brine solutions. Environmental protection policies prohibit the discharge of solid and liquid waste. Therefore, the general practice is to mix solid and liquid tailings for underground backfilling, along with specific cementitious materials. Understanding the compressive strength of magnesium-based tailings cemented bodies under different proportions of filling materials is crucial for improving ore recovery and ensuring mining safety. However, there is currently no suitable method to predict the strength of these cemented bodies. Moreover, research on potassium salt mines is minimal, with limited available experimental data. In recent years, machine learning has emerged as an effective numerical prediction approach, demonstrating significant promise in material strength prediction. Compared with empirical formulas, machine-learning models offer greater generalizability and accuracy, and they can achieve satisfactory prediction accuracy even when available data are limited.

To investigate the strength of magnesium-based solid-liquid tailings cemented bodies under different mixture conditions, laboratory experiments, qualitative analysis, grey relational analysis, and intelligent algorithm coupling modeling were conducted. Uniaxial compression tests were performed on cemented specimens composed of solid tailings, old brine, and composite cementitious binders. A qualitative analysis of the factors influencing the uniaxial compressive strength of magnesium-based tailings cemented bodies was performed, and the grey correlation among these factors was analyzed. By comparing the mean absolute error (MAE) and mean square error (MSE) of predictions generated by neural networks with varying numbers of hidden layer neurons, the optimal structure for the back propagation (BP) neural network was determined. As a result, a 3-11-1 neural network structure was established, and a genetic algorithm (GA)-BP coupled prediction model was developed. This model was then applied for intelligent strength prediction of magnesium-based tailings cemented materials.

The analysis revealed grey relational degrees for three key factors: the mass ratio of the composite cementitious binder to MgCl₂ (0.690), the mass ratio of old brine to tailings (0.639), and the mass ratio of fly ash in the composite binder (0.596). All three factors exhibited significant correlations with compressive strength, with the binder to MgCl₂ mass ratio identified as the most influential parameter. The GA-BP coupled model achieved an R value of 0.9775, MAE of 0.1625, root mean square error (RMSE) of 0.1771, and mean absolute percentage error (MAPE) of 5.24%. By contrast, the traditional BP model yielded an R value of 0.9038, an MAE of 0.2935, an RMSE of 0.3595, and a MAPE of 12.81%. The GA-BP model outperformed the traditional BP model, improving the four evaluation metrics by 8.15%, 44.63%, 50.74%, and 59.09%, respectively.

The findings indicate that the GA-BP model exhibits superior effectiveness and accuracy in predicting the strength of magnesium-based tailings cemented materials compared with the traditional BP model. This study presents a new approach for predicting the strength of magnesium-based solid-liquid tailings cemented bodies, providing a reliable and intelligent method applicable to analyzing and designing backfill materials in potassium salt mines.