Leveraging Machine Learning to Reveal Transparency in Integrated Assessment Model Ensembles

Integrated assessment and energy system models have long been criticized for their limited interpretability of results, particularly by policymakers, who find it difficult to use the evaluation results for informed policy-making due to the challenge of understanding the underlying model drivers and their assumptions. This underscores the urgent need to systematically attribute model outcomes to their underlying drivers for effective decision support. Here, we present a novel post-attribution framework combining error diagnostics, machine learning, and econometric analysis to disentangle the impacts of model inputs, structural inertia, and implicit assumptions. This framework is applied to post-evaluate global energy system and demand sector scenarios across mainstream Integrated Assessment Models (IAMs), identifying the root causes of discrepancies between models regarding the pace of energy transition. We find that the largest discrepancies in model inputs stem from energy demand variables, while errors in economic and energy supply variables are relatively minor, although the latter's input errors can have a delayed impact on long-term emissions forecasts. Significant differences in decarbonization pathways across models, largely driven by model preferences and technological assumptions such as technological inertia, cost, and maturity timelines, underscore the importance of considering modeling preferences in IAMs when attributing long-term emission differences. Our study paves the way for interpreting IAM ensembles results through machine learning, identifying the deep drivers of result discrepancies, and supporting model development and policy decision-making.