A physics-informed impact model refined by multi-fidelity transfer learning

Abstract

Impact performance is a key consideration when designing objects to be encountered in everyday life. Unfortunately, how a structure absorbs energy during an impact event is difficult to predict using traditional methods, such as finite element analysis, because of the complex interactions during high strain-rate compression. Here, we employ a physics-based model to predict impact performance of structures using a single quasistatic experiment and refine that model using intermediate strain rate and impact testing to account for strain-rate dependent strengthening. This model is trained and evaluated using experiments on additively manufactured generalized cylindrical shells. Using transfer learning, the trained model can predict the performance of a new design using data from a single quasistatic test. To validate the transfer learning model, we extrapolate to new impactor masses, new designs, and a new material. The accuracy of this model allows researchers to quickly screen new designs or leverage pre-existing databases of quasistatic test data. Furthermore, when impact tests are necessary to validate design selection, fewer impact tests are necessary to identify optimal performance.