Experimental measurement, mathematical description, and machine learning prediction of the mechanical properties of waste rocks for mine haul roads

Mining operations can generate large amounts of waste rocks, which are often stored on the surface, in piles, close to production sites. However, the management and reclamation of waste rock piles can be challenging because of their large size and the risks for geochemical and geotechnical instabilities. The valorization of waste rocks for the construction of mine haul roads is an attractive alternative to surface disposal, both economically and environmentally. The mine haul road network is, indeed, a critical and vital component of the production process, which directly and significantly impacts mine productivity and costs. Waste rock mechanical properties are critical parameters to improve mine haul roads design, but the relevant research on the mechanical properties of waste rocks from hard rock mine are relatively limited. Also, design tools and measurement methods were usually developed for civil applications, and their applicability to mine haul roads (which are subjected to significantly different traffic loading compared to highways) remains uncertain. The main objective of this project was therefore to evaluate the mechanical properties of waste rocks under mining truck loading to improve the design of mine haul roads. The applicability of laboratory test standards and methods, and of mathematical regression models to waste rocks was assessed. The effect of basic properties and climatic factors on the mechanical properties was quantified. Machine learning models for predicting the mechanical properties of waste rocks were also developed using different algorithms. A large number of CBR and triaxial tests, of various dimensions and with different load types, were carried out in this study. The stress levels proposed in the existing laboratory tests standards (e.g., AASHTO T307 and EN 13286) were significantly lower than the stress state within mine haul roads since these standards were initially developed for the highway design. High stress levels were therefore proposed and applied in this study to represent the typical stress conditions caused by extra heavy mining trucks. Repeated load CBR tests appeared to be an effective alternative to repeated load triaxial tests to estimate resilient and permanent deformation behavior of waste rocks. Results showed that the impact of loading frequency and waveform in repeated load CBR tests was limited, while the measured equivalent modulus increased significantly with contact stress. The developed equation and model for repeated load CBR tests were reliable to describe the stiffness and permanent deformation of waste rocks, respectively. The optimum gravel-to-sand ratio of crushed waste rocks was between 1 and 1.5 and contributed to provide higher resilient modulus and shear strength, and lower permanent strain. An increase in fines content could, to the contrary, result in the decrease of resilient modulus and permanent strain but also to significant increase of shear strength. The optimum gravel-to-sand ratio of uncrushed waste rocks with 60 mm of maximum particle size was around 5 and the permanent strain rate increased significantly when the maximum particle size was decreased to 25 mm. An increasing of compaction effort resulted in a significant decrease in permanent strain of waste rocks; waste rocks with 60 mm of maximum particle size, a gravel-to-sand ratio of 5, and 5% of fines showed higher shear strength. Freeze-thaw cycles resulted in a reduction of the resilient modulus and shear strength, and an increase of the permanent deformation. A greater water content during freeze-thaw cycles also resulted in a significant increase of the permanent deformation. Results also showed that wetting-drying cycles could contribute to increase the stiffness, CBR, and permanent deformation of crushed waste rocks. MR-θ model and MEPDG model appeared reliable to describe the resilient modulus behavior of waste rocks under different stress conditions. Rahman and Erlingsson model (extended using time hardening approach) was well adapted to fit the accumulation of permanent strain with number of loading cycles caused by numerous stress paths. Coefficient prediction models (k1, k2, a, and b), expressed as a function of basic physical properties and climatic factors, were also developed to predict the resilient modulus and permanent strain of waste rocks. The proposed models could be used to predict the mechanical properties of waste rocks in the field, but would need to be improved using more test results to extend their generalization capacity. Different machine learning algorithms, including multiple linear regression (MLR), support vector machine (SVM), k-nearest neighbors (KNN), decision tree (DT), random forest (RF), backpropagation neural network (BPNN), and neuroevolution of augmenting topologies (NEAT), were therefore used to develop prediction models for CBR, resilient modulus, and permanent deformation of crushed waste rocks based on experimental data. Eight properties (dry density, compaction energy, fines content, maximum particle size, CC, CU, D10, and D60) were selected as input variables for CBR models, and number of loading cycles, confining pressure, and deviator stress were chosen as the input variables for the resilient modulus and permanent deformation models. The optimum architecture of machine learning models was determined through hyperparameters study. A comparison study showed that the DT, RF, and BPNN models provided better results for CBR, resilient modulus, and permanent deformation, respectively. The results presented in this thesis should contribute to give mine operators general indications to select and prepare waste rocks to be used in haul road construction. The proposed recommendations could improve the short and long-term performance of haul roads and reduce the need for maintenance, and therefore decrease fuel consumption, improve vehicle component lifetime (especially tires), increase productivity and reduce operation costs.