Sequential Experimental Design and Model Calibration for Targeted Events

Model calibration is often performed with a limited number of data points due to the significant cost of high fidelity simulations or experiments. To capture specific events, such as failure, optimal data collection methods can aid in achieving globally accurate models that can also predict targeted events. In this research, the expected information gain criterion for experimental design is used to determine the most informative designs for sequential calibration of uncertain model parameters. For accurate prediction of events of interest, the Targeted Information Gain for Error Reduction (TIGER) method is introduced to balance the placement of exploration points in the design space based on model accuracy and capturing the event of interest. This approach was compared to using sequential and all-at-once random data collection methods. The comparison of global and local prediction errors indicated that this is a feasible approach based on an analytical two-dimensional example. The method was also successful in a classification problem for flutter and critical limit cycle oscillation amplitude for a panel in hypersonic flow.