Computationally efficient black-box modeling for feasibility analysis

Computational cost is a major issue in modern large-scale simulations used across different disciplines of science and engineering. Computationally efficient surrogate models that can represent the original model with desired accuracy have been explored in the recent past. However, with the exception of few efforts, most of these techniques rely on a reduced order representation of the original complex model, resulting in a loss of information. In this paper we demonstrate the applicability of high dimensional model representation (HDMR) technique in addressing this issue while preserving the original model dimension. We will discuss the applicability of this surrogate modeling technique in the field of feasibility analysis drawing examples from process systems and materials design. It will be shown that the original physical models can be essentially considered as a black box, and same methodology can be applied across all the examples studied. It is found that the accuracy of the surrogate models depends on the order of the approximation and number of sampling points employed. While first-order approximation is largely inadequate, second-order approximation is sufficient for the model systems studied. Sampling requirement is also dramatically low for the construction of these surrogate models.