Details of Research Outputs

This paper proposes an innovative physics-informed and time-dependent surrogate method based on machine learning to calibrate the parameters of detonation products for cylinder test. Model calibration is a step of model validation, verification, and uncertainty quantification. A good calibration result will effectively enhance the credibility of a simulation, even model and software. This method extracts and quantifies the features of data, and corresponds them to the specific physical processes, such as the fluctuation caused by shock wave and the damping effect caused by energy dissipation. Different from the conventional surrogate models, our method gives a special consideration to the time variable and couples it with the detonation parameters properly through feature extraction and correlation analysis. The use of feature screening and variable selection enables this method to deal with high-dimensional and nonlinear situations. Models based on the Cramer-von Mises conditional statistic can reduce the complexity and improve the generalization performance by screening out the variables with strong correlation. And with the Oracle property of adaptive lasso, the convergence property of the method is guaranteed. Numerical examples of PBX9501 show, that the calibration results effectively improve the accuracy of simulation. With the relation between parameters and feature coefficients, we offer an instructive parameter adjusting strategy. Last but not least it can be generalized to other explosives. Model comparison results on 17 types of explosives show that our method has a better agreement with the cylinder test than the classical exponential form.