Multifidelity Monte Carlo for Efficient Estimation of Drag-Coefficient Statistics in Rarefied Satellite Aerodynamics

Abstract

Very-low-Earth-orbit analyses require reliable uncertainty estimates for the aerodynamic drag coefficient, but repeated direct simulation Monte Carlo (DSMC) evaluations are expensive. We investigate whether cheaper collisionless models can reduce the computational cost of estimating DSMC-based drag statistics across different geometries. We also test whether improvements in the mean carry over to a variance reconstructed from two separately estimated raw moments. In the multifidelity Monte Carlo (MFMC) estimator, PICLas DSMC provides the high-fidelity response, while PICLAS test-particle Monte Carlo (TPMC) and the free-molecular Sentman panel model provide low-fidelity control variates. The benchmark contains four canonical Cube cases and scaled GOCE-like and CHAMP-like surface models. Each paired evaluation uses the same physical-input sample in all three models, and the reference values use 450 physically distinct DSMC evaluations per case. At the same estimation cost, DSMC-TPMC-Sentman MFMC reduces the RMSE of the mean drag coefficient in every tested case and budget, with median gains of 1.50-3.03. Variance estimation is less consistent: the three-model estimator improves three cases, performs close to HF-only Monte Carlo in two cases, and fails for GOCE-like. In that case, Sentman is only weakly correlated with the DSMC variance-related responses. Removing Sentman eliminates the negative estimates and improves the variance RMSE, although the decision to remove it was made after inspecting the results. MFMC therefore improves mean estimation reliably at reduced computational cost, but each low-fidelity model must be assessed separately for variance estimation.

arXiv comment20 pages, 6 figures. Submitted to Advances in Space Research