Model calibration with big data

Abstract

The aim of this paper is to address computational efficiency in model calibration in the presence of big data. Model calibration refers to the adjustment of model parameters so that the model output matches well with the observation data. Improvement of sensing technology makes large volume data available, which challenges the capability of current Bayesian calibration techniques, especially when model runs are expensive. Often researchers have used reduced-order models or surrogate models in calibration to save the computational cost; however, this leads to loss of accuracy. This paper applies the MapReduce technique to parallelize the model calibration, in order to make the computation more efficient without lowering the accuracy. This paper applies MapReduce in two steps of the Bayesian calibration process: (1) surrogate model construction, and (2) computation of the posterior distribution of the calibration parameters. In order to build a high quality surrogate model, many training runs should be performed. Therefore the collection of training data is first parallelized in this paper using MapReduce. After the surrogate model is obtained, one commonly used Bayesian updating methods is considered for computing the posterior distribution of the calibration parameters, namely, particle filtering (PF). This paper uses MapReduce to parallelize both steps, and implements them in the Spark platform. The methodology is illustrated using one numerical example. The numerical example is the calibration of the thermal conductivity of concrete, with field temperature observed from infrared thermography (IR) camera. In this example, the large volume of data occurs in the spatial domain.