A novel machine learning framework for efficient calibration of complex DEM model: A case study of a conglomerate sample

The conglomerate reservoirs in the Mahu Sag of the Junggar Basin, northeastern China are featured high heterogeneity and complicated lithology. The reservoirs have experienced a decent production with the application of horizontal well drilling and multistage fracturing. However, the mechanical behavior of the conglomerate formation remains poorly understood. In this respect, the discrete element method (DEM) can be used to efficiently investigate the mechanical behavior of geomaterials with complex lithologies. The implementation of DEM requires a precise set of microscopic parameters, for which the conventional trial-and-error methods are time-consuming. In this regard, an end-to-end machine learning (ML) framework was proposed to achieve an efficient calibration of the complex DEM model. The framework contains two stages, with one predicting the strength of a single phase geomaterial and the other forecasting the overall strength of a complex geomaterial sample containing multiple phases. Different ML models were attempted to evaluate their performances. The results demonstrate that the random forest (RF) model performs with high accuracy at the first stage, and the support vector machine (SVM) model outperforms other ML models in terms of accuracy and convergence at the second stage. From deploying the ML framework, a fast and accurate calibration approach of conglom-erate DEM models was proposed and explored in the case study. Moreover, the sensitivity analysis examined the contribution of each parameter to the overall mechanical behavior and provides the suggested values of microscopic parameters. The validity of the application of the training models on the synthesized data was analyzed to address the data shortage issue. In summary, the ML framework not only enables users to fast and accurately calibrate the microscopic parameters of the complex DEM model, but also effectively assesses the contribution of microscopic parameters to help understand their effects on the overall mechanical behavior.

Automated summary

The study proposes a machine learning framework for efficient calibration of complex discrete element models and assessment of the effects of microscopic parameters on overall mechanical behavior.