Machine learning approach for full Bayesian parameter reconstruction (Conference Presentation)

The goal of Bayesian parameter reconstruction is to identify the probability distribution of parameter values that describe the shape of a sample. The probability density is based on a set of measurements and a corresponding simulation model of the measurement process. A prior distribution and a parametric model of the measurement error are also taken into account. For elaborate simulations, the reconstruction problem is computationally expensive. We present a machine learning (ML) approach aimed at minimizing the number of evaluations of the expensive simulation model. In each iteration the ML model is trained on data from all evaluations of the expensive simulation model, and then used to identify model parameters that have a high probability. Once the parameters with maximal probability density are found, one can use the ML model to analyze the full parameter distribution. We evaluate the approach for the example of grazing emission X-ray fluorescence (GEXRF).