Statistical Approach to Establishing Tuned Simulation Parameters

The data needed to fully benchmark a simulation is frequently inadequate. The analyst is then challenged to make assumptions or otherwise determine how the simulation can be appropriately used to predict future outcomes. A thermal analysis will be presented where temperature data is available at several locations throughout the structure, but information on boundary conditions and thermal properties of the structure is limited. The unknown variables are parameterized and bounded based on the underlying physics and available data for ‘typical’ values. Simulations are then run and a fitting score is used to determine the ranges of the inputs which gave the best overall fit to the test data for the entire structure, avoiding bias of focusing on the single location of interest. In this process, multiple combinations are found that yield similar fit scores, but the variables were confounded and the available data could not support one set of values over another. The resulting ranges of input values were then used in the follow-on analysis work, allowing results to be found as a range of likely values in terms of the input uncertainties. Statistical methods were also applied to the results, allowing determination of which inputs had the greatest impact on the results, and thus identifying where future efforts should be focused to distinguish the variables and provide greater accuracy in the method.