Comparison of Two Bayesian-MCMC Inversion Methods for Laboratory Infiltration and Field Irrigation Experiments

Area of Science:

• Environmental science
• Geophysics
• Computational modeling

Background:

• Bayesian parameter inversion relies on forward models linking subsurface properties to data.
• These forward models are often computationally intensive, requiring numerous iterations.
• Fast alternative systems are crucial for making stochastic inversion tractable.

Purpose of the Study:

• To compare the performance of original HYDRUS-1D forward models with Artificial Neural Network (ANN) and Gaussian Process (GP) surrogate models.
• To evaluate the reliability and efficiency of ANN and GP surrogates within a Bayesian inversion framework.
• To assess the impact of model error quantification on surrogate model performance.

Main Methods:

• Implemented ANN and GP surrogate models as alternatives to the original HYDRUS-1D forward model.
• Quantified ANN model error using principal component analysis.
• Measured GP model error using its inherent variance.
• Utilized measured pressure head data from laboratory soil column infiltration and field irrigation experiments for parameter inversion.

Main Results:

• Strong correlations between simulated and observed pressure head data confirmed the reliability of surrogate models in Bayesian inversion.
• Approximate forward models (ANN and GP) significantly enhanced inversion efficiency.
• Similar accuracy was achieved between original and optimized results using surrogate models.

Conclusions:

• Surrogate models, including ANN and GP, are reliable for Bayesian inversion when forward models are computationally prohibitive.
• These surrogates offer a significant improvement in computational efficiency without sacrificing accuracy for nonlinear subsurface flow problems.