MFaRM: Machine Learning based Fast Running Model for Debris Prediction of Hardened Structures

The site planning of magazines must meet the explosive safety requirements. High fidelity physics-based modeling approaches are emerging for understanding structural damage and predicting fragmentation. However, based on non-standardized modeling and simulation, the reevaluation process of design changes in magazine structures often takes weeks or months for just one specific case. It is desired to develop a fast running model that provides reliable and accurate predictions in the scale of minutes for all the magazine and related hardened structures, to expedite the evaluation process and minimize operational discontinuity. In this study, we propose to develop a fast running model based machine learning techniques. The fast prediction relies on a database developed a priori, and compiled the batch simulation results based on high-fidelity Lattice Discrete Particle Method (LDPM) and existing experimental data. Multifidelity Gaussian Process Regression will be applied to produce the final estimation of the fragmentation characteristics based on the data combinations and given design changes. Moreover, the LDPM method will incorporate the comminution theory to capture very small fragments. To speed up the development process of the database, reduced order modeling (ROM) based on machine learning technique will also be investigated to shorten the single run time of the high-fidelity model.