Data-driven prediction of growth rate for a shocked heavy gas layer

Abstract

Numerical investigation on the evolution of a heavy gas layer is performed over a wide range of parameters. Neural networks and curve fitting techniques are employed to predict the growth rate of downstream interface based on 2688 simulated cases. Significant amounts of observable data are generated by considering four primary variables: shock wave intensity, density difference between the inside and outside of the gas layer, gas layer thickness, and initial interface shape. The neural network model maps the growth rate directly to the initial parameters, while the curve fitting approach provides an explicit formula. The neural network model has high accuracy and a certain extrapolation capability. The explicit formula provides a more intuitive understanding compared to the neural network model and has a stronger extrapolation. Furthermore, to thoroughly examine the evolution of the gas layer, the numerical investigation is conducted on the shocked single interface. It is discovered that there is a range of parameters in which the growth rate of gas layer is lower than that of the single interface. Meanwhile, a modified model that includes an attenuation factor is proposed to replace the impulsive model of the single interface. In summary, these methods can significantly reduce simulation time by quickly identifying desirable cases.