Deep and machine learning architecture for high accuracy prediction of interior ballistics for modular charges

Abstract

AbstractInterior ballistics modelling intends to calculate physical quantities by solving equations requiring a set of input like thermodynamical properties linked to propellant or mechanical properties like friction coefficients, linked to the system, that are often difficult to assess experimentally. Using a consequent and high‐quality dataset of thousands of 155 mm modular charge experiments, we investigated how deep and machine learning approaches could build an efficient model for predicting muzzle velocity, pressure, and ignition time. After identifying the main variables influencing the physics and several potentially relevant to the machine learning and deep learning algorithms, a stacking ensemble learning model (SEL) was built. It is composed of eight machine learning algorithms and one deep learning algorithm, all as weak learner, reaching a consensus for predicting experimental outputs through a strong learner. The performance assessment, on a 900 experiments test data set that were not used in training, shows a prediction accuracy of 97 % at 1 % accuracy for muzzle velocity, 93 % at 3 % for pressure and 72 % at 5 ms for ignition time. The prediction ability for muzzle velocity within 2 m/s reaches 73 % that can be considered as sufficient. Furthermore, the SEL model even having zero knowledge about physics can discriminate gun barrel erosion effect at high and low pressure. This approach allows to imagine a way of modelling physical phenomena poorly described by physical equations or requiring unmeasurable figures.