Speaker
Description
Accurate modeling and prediction of flow and reactive transport in fractured and porous media under uncertainty requires characterization of unknown model parameters such as permeability, hydrodynamic dispersion coefficient, and/or reaction kinetics, with their estimation uncertainty. Sensitivity analysis of such parameters plays an important role before or during the uncertainty analysis by checking whether one has enough information to identify a parameter given uncertainty of other parameters and hence reducing the number of parameters to be estimated. In this work, we propose local and global sensitivity analysis methods using a deep learning-based approach. With advancements in computational power and open-source programming packages, deep learning can offer computationally efficient reduced-order models to the flow and reactive transport problem. With automatic differentiation, trained deep learning models can produce local sensitivity, i.e., the gradient of the forward model output with respect to the observation, in the order of seconds and the expensive computation of global sensitivity such as Moris and Sobol indices is computationally feasible. In specific, we use physics-informed neural network approaches to offer local and global sensitivity analysis. Since such analysis tells which parameters are informative for optimal experimental design with data-worth analysis, decision makers can allocate limited budgets optimally to collect field observations for future site characterization.
SNL is managed and operated by NTESS under DOE NNSA contract DE-NA0003525.
| Time Block Preference | Time Block C (18:00-21:00 CET) |
|---|---|
| Acceptance of Terms and Conditions | Click here to agree |
