A multifidelity method for the nonlinear propagation of uncertainties in the presence of stochastic accelerations is presented. The proposed algorithm treats the uncertainty propagation (UP) problem by separating the propagation of the initial uncertainty from that of the process noise. The initial uncertainty is propagated using an adaptive Gaussian mixture model (GMM) method which exploits a low-fidelity dynamical model to minimize the computational costs. The effects of process noise are instead computed using the PoLynomial Algebra Stochastic Moments Analysis (PLASMA) technique, which considers a high-fidelity model of the stochastic dynamics. The main focus of the paper is on the latter and on the key idea to approximate the probability density function (pdf) of the solution by a polynomial representation of its moments, which are efficiently computed using differential algebra (DA) techniques. The two estimates are finally combined to restore the accuracy of the low-fidelity surrogate and account for both sources of uncertainty. The proposed approach is applied to the problem of nonlinear orbit UP and its performance compared to that of Monte Carlo (MC) simulations.
Budget-limited distribution learning in multifidelity problems
在计算科学中,多保真度方法被广泛用于通过采用不同成本和精度的数值模拟来估计感兴趣的量(QoI)。许多方法近似于仅表示有限信息的数值统计,例如,标量统计,关于生活质量。不确定性的进一步量化,例如,对于风险评估、失效概率或置信区间,需要估计全部分布。在本文中,我们推广了[Xu et al.,Siam J. Sci. Comput. 44.1(2022),A150-A175]以开发近似标量值QoI的全分布的多保真度方法。与其他方法相比,我们的方法的主要优点是,我们不需要高保真度和低保真度模型之间的特定关系(例如模型层次结构),并且我们在程序开始之前不假设任何模型统计信息,包括相关性和其他跨模型统计信息。在适当的假设下,在上述框架中,我们实现了可证明的1-Wasserstein度量收敛的算法构造的分布式仿真器通过探索-开发策略。我们还证明了我们的算法所采取的关键政策行动是渐进最优的。数值实验支持我们的理论分析。
A non-intrusive reduced-order modelling for uncertainty propagation of time-dependent problems using a B-splines Bézier elements-based method and Proper Orthogonal Decomposition: application to dam-break flows