Defensive multifidelity importance sampling for reliability-based design: source code and simulation data

Description

This dataset contains the Python source code and the simulation data behind the manuscript "Defensive multifidelity importance sampling for reliability-based design with uncertain source models" (Reliability Engineering & System Safety). The method addresses reliability-based design of filament-wound composite pressure shells, where every constraint evaluation has to estimate a rare-event failure probability under a small budget of high-fidelity model calls. Source models of the same shell family are available, but their reported failure probabilities differ from the target by orders of magnitude, so they cannot be fused into the answer. In the method a source model contributes only the ordering of candidate samples: a rank-preserving cross-entropy construction builds one proposal per source from a source-specific quantile, a selection-free defensive multiple-importance-sampling estimator pools those proposals with the nominal density, and the reported failure probability comes from target-model evaluations alone. The estimator's internal stratified variance estimate then serves as the uncertainty scale of a calibrated reliability margin in the optimization loop.

Steps to reproduce

the figures can be rebuilt from the included JSON files with python make_figures_v2.py. Required package versions are pinned in requirements.txt (Python 3.9, numpy 1.26.4, scipy 1.11.2, matplotlib 3.8.4); note that numpy must stay below 1.28 for this scipy build. Re-running the experiments from scratch takes considerably longer, because each design point needs 150 MIS repetitions at 800 target-model calls plus a 2×10⁶-sample Monte Carlo reference.

Institutions

• Huazhong University of Science and Technology

  Hubei, Wuhan

Categories

Licence