Main content area

A new efficient Bayesian parameter inference strategy: Application to flow and pesticide transport through unsaturated porous media

Younes, Anis, Mara, Thierry A., Voltz, Marc, Guellouz, Lamia, Musa Baalousha, Husam, Fahs, Marwan
Journal of hydrology 2018 v.563 pp. 887-899
Markov chain, drainage, models, pesticides, porous media, samplers, uncertainty
Statistical calibration of flow and transport models in unsaturated porous media is often carried out with Markov Chain Monte Carlo (MCMC) methods. However, the practicality of these methods is limited by their computational requirement, particularly when large prior intervals are assigned to the model parameters. In this work, a new operational strategy is investigated to alleviate the computational burden of MCMC samplers using results from a preliminary calibration performed with the First-Order Approximation (FOA) method. With the new strategy, the posterior distribution is approximated using a high-order Polynomial Chaos Expansion (PCE) surrogate model constructed over reduced parameter ranges. The latter are obtained from the 99.9 FOA confidence intervals.Two challenging test cases are investigated to assess efficiency and accuracy of the new strategy. The first test case considers estimation of flow and pesticide transport parameters from a synthetic infiltration experiment. The second test case deals with the assessment of unsaturated hydraulic soil parameters from a real-word laboratory drainage experiment. The results of the proposed strategy are compared to those of FOA, of the standard MCMC method and of an improved MCMC method in which the sampler is preconditioned with draws from the FOA posterior distribution. For both test cases, the new strategy provides accurate mean estimated parameter values and uncertainty regions and is much more efficient than the other MCMC methods. It is up to 50 times more efficient than the standard MCMC method.