Identifying and Analyzing Outlying Futures in Integrated Assessment Models

Climate models have grown increasingly complex as they aim to capture interactions between environmental, social, and economic systems. These models are now routinely used to generate large ensembles of scenarios, requiring robust and scalable methods to extract meaningful insights. Our research demonstrates the application of Outlier Set Two-step Identification (OSTI) to systematically extract, evaluate and interpret outlying ensembles of futures from Integrated Assessment Model (IAM) outputs. OSTI is a novel technique that combines Gaussian Mixture Models for probabilistic clustering with Inter-cluster Mahalanobis distance measurement and hypothesis testing to identify clusters of scenarios that deviate significantly from typical patterns. 

Here, we analyze irrigation withdrawal patterns across 27 major river basins using outputs from the Global Change Analysis Model (GCAM). GCAM integrates climate, economic, and human systems to explore future pathways through 2100 at five-year intervals. We apply OSTI to 3,000 scenarios of agricultural water demands through 2100, generated by varying seven key GCAM parameters including socioeconomic development, agricultural practices, groundwater availability, reservoir storage capacity, climate trajectories, and carbon tax policies.

We then systematically extract these OSTI-identified outlying futures to identify distinct patterns that appear repeatedly across multiple basins, focusing on scenarios that share unique combinations of socioeconomic and agricultural parameters. The extraction process highlights outlier sets against their input parameters to understand what combinations of model inputs lead to outlying behavior. In these outlying sets, water supply parameters have minimal influence on outlying future determination, while demand-related parameters dominate. We speculate this reflects GCAM's recursive economic equilibrium mechanism, which interprets  physical water scarcity constraints in terms of economic cost but does not make them binding. The spatiotemporal analysis shows distinct irrigation withdrawal patterns across two time periods (2015-2050 and 2050-2100). Most basins exhibit increasing irrigation withdrawals until mid-century, followed by significant declines or stabilization, particularly for winter crops like wheat. This pattern strongly correlates with groundwater dynamics, where peak extraction occurs around 2050, followed by declining usage due to increasing pumping costs and declining water tables. However, high-value crops like cotton maintain relatively stable withdrawal patterns throughout the century, while sugarcane shows continued growth in some scenarios, reflecting adaptation to changing water availability and economic priorities.

These results establish OSTI as a diagnostic tool for systematic identification of limitations and potential artifacts in complex models like IAMs. As IAMs like GCAM become increasingly pivotal in understanding multi-sectoral dynamics under deep uncertainty, OSTI offers a robust and scalable tool for scenario discovery. Beyond, our approach is applicable to the exploration of large scenario ensembles in other contexts. It provides a scalable way to identify and analyze potentially outlying scenarios requiring special attention in adaptation planning.