<p>In the numerical simulation of particle flow code (PFC^2D), the values of micro-parameters directly influence the macroscopic mechanical parameters and overall performance of the numerical model. However, traditional methods for determining micro-parameters often require extensive manual trials and adjustments, leading to a highly blinded, time-consuming, labor-intensive process with limited accuracy. This paper employs the discrete element software PFC^2D in combination with four machine learning algorithms: support vector machine (SVM), random forest (RF), gradient boosting decision tree (GBDT), and xtreme gradient boosting (XGBoost), to analyze the sensitivity of PFC parameters. The machine learning algorithms use 6 particle flow parameters encompassing 156 sets of data, as input variables, with the model's peak stress and elastic modulus (E) as output variables. Simultaneously, three performance evaluation metrics used to assess the performance of the algorithms. The research results indicate that the RF algorithm outperforms other models in simulating the test set of mesoscopic parameters, with the highest trend evaluation index. The parameter pb_coh has the greatest positive impact on the model's peak stress, while the parameter deform emod has the greatest positive impact on the model's elastic modulus. The machine learning algorithms provide a better method for parameter calibration, aiding in a better understanding and prediction of micro-parameters for PFC^2D rock models.</p>