Open-stope mining, often employed in methods such as sublevel stoping, cut-and-fill stoping, and shrinkage stoping, plays a crucial role in the underground extraction of mineral reserves. In fact, 9% of the metallic ore reserves in the United States are mined using the sublevel stoping method in open stopes. The stability of the open stope and consequently the stability of the hanging wall are of paramount importance in this method. Stability directly influences worker and equipment safety, mining cycle efficiency, and excessive dilution of the ore. Despite the inherent difficulty in accurately predicting hanging wall stability, various methodologies have been proposed to forecast it based on influential variables. Currently, the most widely used methods include empirical methods, numerical modeling, analytical design, in-situ measurements, and the Mathews stability graph. However, significant challenges remain in predicting hanging wall stability, as these methods come with numerous limitations. For instance, numerical modeling requires back analysis to obtain input parameters, which are often not readily available. The application of analytical design is limited due to the simplifications inherent in its use. In-situ measurements are costly and time-consuming, and the Mathews stability graph is difficult to generalize to new situations. Therefore, this research leverages a database comprising 135 records from mining activities in Australia and employs intelligent methods to develop new empirical models for assessing and predicting the stability of underground open stopes. In this thesis utilizes intelligent algorithms based on decision trees, such as the classification and regression tree (CART), classification model J48, and logistic model tree (LMT) algorithms, as well as regression methods, including logistic regression (LR). In developing these models, 90% of the available data in the database were used for training and the remaining 10% for testing. The results indicated that the model developed using the LMT algorithm exhibited the best performance among the developed models, for this model Area under the ROC Curve is equal to 0.940 and Performance Index (PI) is equal to 4.357. In this research, the developed models were prioritized based on the performance eva‎luation index, which after the LMT model, the J48 model performed better than other models.It is also noteworthy that all developed models are of the white-box type, making them easily interpretable and applicable.

ابراهيم قاسمي ورنوسفادراني

هادي حسيني , سعيد مهدوي