In numerical simulation studies utilizing the Particle Flow Code （PFC），the direct acquisition of microscopic parameters of discrete particles through experimental means presents a significant challenge.Traditionally，manual trial-and-error techniques are utilized，involving continuous adjustments of microscopic parameters to observe the corresponding effects on macroscopic mechanical parameters within the simulation.This iterative process is characterized by its randomness and lack of systematic approach，heavily reliant on the expertise of the calibrator，ultimately leading to diminished calibration efficiency and reproducibility.This challenge is particularly evident when calibrating microscopic parameters for extensive quantities of rock samples，requiring substantial manual labor and repetitive tasks.In order to mitigate these issues，a novel approach for calibrating microscopic parameters is essential，one that is not reliant on the calibrator’s skill level and ensures consistent reproducibility in the calibration of rock specimen parameters on a large scale.Utilizing the parallel bond model as a case study，an orthogonal experimental design table was constructed to investigate microscopic parameters，followed by numerical simulations to generate a comprehensive small-sample dataset.Support vector regression （SVR） and back propagation（BP） neural network models were separately trained on this dataset.This approach involves utilizing macroscopic parameters derived from PFC numerical simulations as the forward process，with machine learning techniques employed to predict microscopic parameters as the inverse process.If the prediction error for the macroscopic parameters measured in the laboratory is deemed inadequate，the corresponding microscopic parameters and their resultant macroscopic parameters are incorporated into the dataset for additional training until the desired outcome is attained.Studies have shown that utilizing machine learning with small-sample data，in conjunction with forward and inverse modeling，can effectively calibrate parameters.In particular，the BP model required 7 iterations，whereas the SVR model only needed 3 iterations to attain satisfactory outcomes，showcasing superior calibration efficiency.In scenarios involving numerous and highly nonlinear macro-micro parameters，the utilization of machine learning-assisted calibration presents notable benefits over traditional manual trial-and-error approaches，including enhanced efficiency，increased reproducibility，and improved generalizability.

http://www.goldsci.ac.cn/article/2024/1005-2518/1005-2518-2024-32-4-675.shtml