基于MCMC方法的虚拟试验贝叶斯校准研究

虚拟试验利用计算模型研究复杂物理过程,并预测其性能.计算模型输入参数中通常包含部分固定但未知参数,可利用计算模型结果和少量有限的试验数据校准未知参数,并研究存在建模不确定性时虚拟试验的预测问题.提出了一种贝叶斯统计方法,采用高斯过程为仿真计算模型以及模型不确定性建模,利用Markov chain蒙特卡罗抽样方法计算校准参数和仿真模型预测后验分布.设计测试算例演示所提出方法的高效性.

Bayesian Calibrated Prediction Approach in Virtual Experiment Based on MCMC

Computational models are increasingly developed to investigate and predict complex physical processes. Various uncertainties inevitably exist in model inputs, in which some are random variables, and some are supposed to take fixed but unknown. The unknown context-specific model parameters should be calibrated by computational model outputs and limited experiments observations before predicting physical process using model. A Bayesian calibration approach was proposed relying on Gaussian process to model unknown input parameters and computational model. The estimation was carried out using Markov chain Monte Carlo. An example is demonstrated that the methodology is efficient.