Real-time inverse-model analysis and control on data collection

Sophisticated numerical models are commonly used to simulate fluid and chemical flow in the subsurface. The science of flow in porous media is composed of general physical principles (transferable knowledge) and site-specific details. All sites are unique, so even if the physics is well understood, we need detailed, site-specific information to develop a model for each site (subsurface heterogeneity, initial and boundary conditions, etc.). In this respect, at each new site, we "start over". The most time- and resource-consuming step in reducing predictive uncertainty bounds in subsurface systems is the process of uncovering the site-specific details. The current paradigm is to perform a lengthy reconnaissance phase to understand the site, followed by additional data collection and modeling to synthesize the information. Model development methods are slow and labor-intensive for complex sites; therefore, model results generally lag behind the data collection by a considerable length of time. This delay limits the usefulness of the model as a tool to guide data collection: any given iteration of the model is out of date by the time it is completed. The whole process is unacceptably protracted in an era in which, for example, we may ultimately need hundreds of sites to implement CO2 geologic sequestration. Our technical capabilities for efficiently collecting and organizing subsurface data have progressed recently with the advent of modern data collection and transmission systems. However, our ability to process this information in the form of numerical models has lagged behind. We propose a new paradigm for the development of complex subsurface flow and transport models in which the inverse analysis is performed in real time, simultaneously with the data collection. Furthermore, we propose to use the inverse model to control the data collection or the operating conditions of an extraction system in real time. This is extremely important because post mortem examination of many field studies illustrates that much of the collected information is redundant and does not further reduce the model uncertainty. Some of these studies have also uncovered missed opportunities to collect information that might have substantially reduced model uncertainty. Often in hydrogeology, the repetition of the data collection is prohibitively expensive or even impossible (e.g. 100-year flood event, the movement of contaminant plumes). By integrating the model development and data collection processes, we believe we can radically reduce the cost and time required for site characterization. We demonstrate the applicability of this concept using simple synthetic analyses representing groundwater flow and contaminant transport. We use two separate models: (1) a forward model representing our synthetic reality, and (2) an inverse model that assimilates the accumulated data and guides the data collection in real time. Our initial results demonstrate the capabilities and challenges in the proposed approach.