Ensemble-based global fire modeling as a tool to characterize extreme wildfire events

Abstract. Understanding the full range of possible extreme wildfire events is crucial for risk assessment and adaptation planning. While the historical record offers only one realization of climate, large ensemble simulations sample a broader range of physically plausible climate trajectories, enabling the assessment of rare but realistic extreme events beyond what the observational record alone can reveal. Here we drive the process-based dynamic vegetation-fire model LPJmL-SPITFIRE at the global scale with different climate inputs to produce three sets of simulations: a 40 member large ensemble (40 members × 36 years sample), a single member drawn from the same ensemble (36 years sample), and a reanalysis-driven simulation (36 years sample), with the latter two each representing only a single trajectory of the climate system. This design enables direct comparison of how these two single realizations (single member and reanalysis) versus large ensemble simulations sample the most extreme fire events. We demonstrate that the single realizations are not suited to study risks associated with the most extreme events in fire danger, burned area, and fire carbon emissions that would be possible under current climate conditions. As expected, the highest values in these runs are typically much lower than those of the large ensemble in most regions. The undersampling of extremes by single realizations is greater for fire impacts (burned area and carbon emissions) than for fire danger, highlighting that vegetation–fire feedbacks interact nonlinearly with internal climate variability. While large ensembles reveal more extreme possible events than those simulated with reanalysis and a single climate model ensemble member, they also enable a more robust analysis of the relationship between extreme fire danger and extreme impacts. In particular, the most extreme burned area and emissions do not always coincide with the most extreme fire danger, underscoring the role of non-climatic factors such as ignitions and fuels. Specifically, years with global maximum impacts may occur in years with global fire danger 4.6 % – 8.4 % lower than the maximum. The findings demonstrate that modeling a broader range of physically plausible wildfire events through large ensemble simulations can help identify the mechanisms leading to the most extreme and high-impact events.