A variety of machine-learning methods has been applied to problems for which physics-based solutions are either nonexistent or computationally expensive. Based on such methods, surrogate models, i.e., empirical models that are trained on outputs of the more computationally intensive methods, can provide acceptable accuracy while dramatically reducing execution time and expense. This work describes the application of an artificial neural network ensemble (ANNE) approach, originally developed over the past fifteen years for cheminformatics studies in the pharmaceutical industry, to train surrogate models that predict missile aerodynamic coefficients.

This modeling engine has been consistently demonstrated to provide best-in-class predictive models for a variety of molecular properties from only a molecule’s structure, and is in widespread use in the pharmaceutical industry today. The surrogate models developed to predict aerodynamic coefficients for arbitrarily shaped missiles at arbitrary Mach numbers and angles of attack have resulted in highly accurate predictions that execute in milliseconds on a modern laptop computer. The ability for rapid predictions can be integral in the design process for missiles and other aerodynamic bodies. Building on previous work, we show how descriptor sensitivity analysis identifies the key descriptors driving the model performance, and relates inputs to outputs to help meet critical design/mission objectives.

• Next Post
• Prev Post