Surrogate modeling and model calibration under lack of data, and conservative RBDO with model uncertainty

Abstract

Model-based design is a concept that contrasts with design through tests. The model refers to any simulations or calculations using a computer instead of physical values obtained from actual measurements or tests. The model-based design follows the process of 1) construction of surrogate models to reduce the computational cost of the simulations, 2) calibration of the models using the actual test data, and 3) running an optimization using the surrogate models as the performance functions. In this study, four methods are presented to overcome the difficulties of existing methodologies in the series of the engineering design framework in each research topic. The first difficulty is that the model is often too heavy and of high dimension, which causes difficulty in surrogate modeling. Therefore, this study proposes an efficient method to construct a surrogate model and variable selection simultaneously with very few samples. The second difficulty is the lack of the number of test data for statistical model calibration. Therefore, a method is proposed to overcome the insufficient number of test data in model calibration using multiple responses. As the third research, this research proposes a method to perform conservative reliability-based design optimization (CRBDO) considering model uncertainty not to result in the unreliable optimum for model-based optimization. Finally, this research proposes a process to reduce the experimental error, which is the most problematic when applying all of the above engineering design procedures using experimental data.