New Approach to Assess Uncertainty in Predictions from Reactive Transport Models
Researchers used Bayesian networks to develop a new method to measure and rank which components of numerical models contribute the most uncertainty to model outputs.
A multi-institutional team of scientists developed a new sensitivity analysis framework using Bayesian Networks to quantify which parameters and processes in complex multi-physics models are least understood. The method can guide continued development and refinement of predictive models of environmental systems by highlighting which components of complex systems require enhanced characterization data to reduce uncertainty.
Sensitivity analysis is a numerical tool used to identify important parameters and processes that contribute to the overall uncertainty in model outputs. This new research applies a Bayesian Network approach to sensitivity analysis frameworks. This approach increases the flexibility and power of the sensitivity analysis by quantifying the contribution of uncertainty from a variety of controlling factors and ranking them, which can better inform decisions on where to focus resources in order to improve the predictive capability of a variety of multi-physics models.
Numerical modeling is an important tool for predicting the future behavior of complex systems that impact the environment and for managing natural resources. For example, PNNL researchers are developing numerical models to study the factors that control the exchange of river and groundwater in the Hanford Reach, the last free-flowing stretch of the Columbia River that defines the north and east boundaries of the DOE Hanford Site.
Predictive uncertainty is inevitable in numerical models of systems such as the Hanford Reach because of the complexity of the hydrologic and biogeochemical properties of the natural system and the limited site characterization data available. To effectively and efficiently reduce predictive uncertainty with limited resources, researchers perform sensitivity analysis to rank the importance of different uncertainty sources that contribute to overall uncertainty in model predictions.
Current state-of-the-art sensitivity analysis frameworks are unable to describe the entire range of uncertainty sources involved in predictive models of complex systems. The integration of Bayesian network-based methods into these frameworks allows the full representation of uncertainty sources and the relationships between them, opening the door to perform sensitivity analysis on complex systems. For example, the networks allow researchers to computationally and graphically understand how uncertainty in one node of the network, or group of nodes, propagates through a network and impacts the overall predictive uncertainty of a model.
The authors implemented their method based on Bayesian networks on a real-world biogeochemical model of the groundwater-surface water interface within the Hanford Site’s 300 Area. They used the framework to run model simulations to predict how factors such as variation in river stage under future climate scenarios and the release or damming of water in upstream hydroelectric dams would contribute to variations in groundwater-surface water exchange, and impact biogeochemical processes that affect the rate of organic carbon consumption.
The team found that groundwater flow and reactive transport processes contribute most significantly to the predictive uncertainty in carbon consumption rate, and that the future states of the climate, which defines the driving forces of the system, were less significant. Further analysis of the uncertainty contributed by groundwater flow processes revealed that the geological structural information, such as the thickness of the confining layer between the river and groundwater, was more important than the within-formation permeability field in controlling the flow processes.
While the Bayesian-network based methodology in this research was implemented on a complex biogeochemical model of the Hanford Site 300 area, the authors say it is mathematically rigorous and generally applicable to reduce uncertainty in a wide range of Earth system models.
Xingyuan Chen, Pacific Northwest National Laboratory, Xingyuan.Chen@pnnl.gov
Funding for this research came from DOE Office of Science BER, PNNL Subsurface Biogeochemical Research SFA.