A Machine Learning Based Transfer to Predict Warhead In-Flight Behavior from Static Arena Test Data
Warhead fragmentation predictions are based on either numerical simulations or static arena tests where detonations occur in unrealistic conditions (not flying). The first methodology presents many shortcomings: there is no agreement on the state of the art for simulations, and many tools ignore important aspects such as gravity, aerodynamic forces and moments, and rigid body motion of different shape fragments. Numerical simulations are also lengthy and cannot be used as on-line/on the battlefield tools. The experimental approach is also extremely limited, as it does not reproduce the real world conditions of a moving warhead. The objective of this work is to combine high fidelity numerical models with unique/ad-hoc experimental activities to strengthen basic science underpinning the test and evaluation framework for warhead fragmentation and fragments fly-out. In particular, we will aim at combining the most advanced simulation capabilities with static experimental data, to obtain a transfer function predicting lethality and collateral damage of a given warhead in real life conditions. Artificial neural networks and/or other machine learning tools (e.g., Random Forests) will be used to capture the underlying physics governing fragments dispersion under dynamic conditions, coming from NAVAIR’s Spidy software, and eventually combine this knowledge with a real warhead characteristics, coming from the static test. This proposal is of high impact because of the existing gap in analytical tools to define and validate warhead fragmentation testing. The broader impact (long term) of this work may be a software tool that the warfighter can use on the field to rapidly assess the effects of the arsenal at his disposal. This tool will be equally beneficial to designers and testers within the Air Force and the DoD in general.