Direct Forecasting of Subsurface Flow Response from Non-Linear Dynamic Data By Linear Least-Squares in Canonical Functional Principal Component Space.

Hydrogeological forecasting problems, like many subsurface forecasting problems, often suffer from the scarcity of reliable data yet complex prior information about the underlying earth system. Assimilating and integrating this information into an earth model requires using iterative parameter space exploration techniques or Monte Carlo Markov Chain techniques. Since such an earth model needs to account for many large and small scale features of the underlying system, as the system gets larger, iterative modeling can become computationally prohibitive, in particular when the forward model would allow for only a few hundred model evaluations. In addition, most modeling methods do not include the purpose for which inverse method are built, namely, the actual forecast and usually focus only on data and model.

In this study, we present a technique to extract features of the earth system informed by time-varying dynamic data (data features) and those that inform a time-varying forecasting variable (forecast features) using Functional Principal Component Analysis. Canonical Coefficient Analysis is then used to examine the relationship between these features using a linear model. When this relationship suggests that the available data informs the required forecast, a simple linear regression can be used on the linear model to directly estimate the posterior of the forecasting problem, without any iterative inversion of model parameters. This idea and method is illustrated using an example of contaminant flow in an aquifer with complex prior, large dimension and non-linear flow & transport model.