| Abstract: |
| The estimation of spatially-dependent model parameters based on limited observations of related dynamic outputs is computationally difficult. When treated as a dynamics-constrained least squares problem, parameter identification suffers from the inherent ill-posedness of the inverse mapping, and in the Bayesian setting, the potential high-dimensionality of the parameter space limits its efficient traversal by a sampling scheme such as the one generated by a Markov Chain Monte Carlo (MCMC) sequence. Machine learning offers the potential of incorporating additional auxiliary information of the dynamical system to design regularization schemes, constrain the parameter space, or further condition the estimation problem. In this talk we investigate various machine learning tools and architectures that can be used to condition parameter estimation problems related to a simple model predicting the spread of pollutants in a slowly moving river. |
|