Abstract
This paper proposes a novel fixed inducing points online Bayesian calibration(FIPO-BC) algorithm to efficiently learn the model parameters using a benchmarkdatabase. The standard Bayesian calibration (STD-BC) algorithm provides astatistical method to calibrate the parameters of computationally expensivemodels. However, the STD-BC algorithm scales very badly with the number of datapoints and lacks online learning capability. The proposed FIPO-BC algorithmgreatly improves the computational efficiency and enables the onlinecalibration by executing the calibration on a set of predefined inducingpoints. To demonstrate the procedure of the FIPO-BC algorithm, two tests areperformed, finding the optimal value and exploring the posterior distributionof 1) the parameter in a simple function, and 2) the high-wave number dampingfactor in a scale-resolving turbulence model (SAS-SST). The results (such asthe calibrated model parameter and its posterior distribution) of FIPO-BC withdifferent inducing points are compared to those of STD-BC. It is found thatFIPO-BC and STD-BC can provide very similar results, once the predefined set ofinducing point in FIPO-BC is sufficiently fine. But, the FIPO-BC algorithm isat least ten times faster than the STD-BC algorithm. Meanwhile, the onlinefeature of the FIPO-BC allows continuous updating of the calibration outputsand potentially reduces the workload on generating the database.