**5. Conclusions**

While the key conclusion from this research was that fires that were historically considered very-large and rare are likely to become increasingly frequent in most regions of the Continental United States at the end of the 21st century, there are also a number of other complexities in future wildfire activity that may be of further relevance to researchers and decision-makers. For instance, although temperature based metrics were often important for prediction, this analysis also found that the identification of important predictors could be highly uncertain across a number of factors, which should be ignored at one's peril. Moreover, even using the relatively simple probabilistic models we developed, rich details regarding future wildfire activities were constructed that reasonably matched observed fire frequencies and were dynamic in terms of intra-annual trends, fire frequency, simultaneous fire occurrence, and the readiness with which LFs become VLFs. Although overall increases are predicted, we also observed exceptions and regional variability. In the Northwestern United States, VLF frequencies were predicted to increase, with nearly two additional events per year, and increases close to one additional VLF per year were fairly commonly throughout much of the Continental United States as well. In rare instances, the potential for decreases in VLF activity was also reported, most surprisingly in Mediterranean California.

The cumulative impact of these changes are anticipated to affect decision-makers in various ways and the techniques described here have a number of benefits for addressing their needs. For instance, the presented Bayesian model averaging techniques avoids many of the risks of traditional model selection techniques that are especially dangerous when predicting complex phenomena such as wildfire. Moreover, this method simultaneously provides a natural method of calculating important event probabilities that are critical to informed decision-making. While uncertainty in climate models is well understood amongst climate impact researchers, these results highlight the hidden sources of structural uncertainty, and encourage the use of Bayesian model averaging to reconcile them into robust forecasts of future wildfire and other impacts resulting from climate change.

**Author Contributions:** Conceptualization, N.K.L.; Data curation, N.K.L.; Formal analysis, H.R.P., N.K.L. and E.A.S.; Funding acquisition, N.K.L.; Methodology, H.R.P., N.K.L., E.A.S., A.C. and E.A.; Project administration, N.K.L.; Resources, N.K.L.; Software, H.R.P.; Supervision, N.K.L. and E.A.S.; Visualization, H.R.P., N.K.L. and E.A.S.; Writing—original draft, H.R.P.; Writing—review & editing, H.R.P., N.K.L., E.A.S., A.C. and E.A.. H.R.P. contributed to the paper's methodology, software, formal analysis, writing—original draft preparation, writing—review and editing, and visualization. N.K.L. contributed heavily to the paper's conceptualization, methodology, formal analysis, resources, data curation, writing—review and editing, visualization, supervision, project administration, and funding acquisition. E.A.S. contributed to the methodology, formal analysis, writing—review and editing, visualization, and supervision. A.C. and E.A. contributed to the methodology, as well as the writing—review and editing.

**Funding:** This research was funded by Joint Fire Science Program project number 11-1-7-4 and by the Joint Venture Agreement 13-JV-11261987-094 between the USFS PNW Research Station and the University of Washington.

**Acknowledgments:** The authors would like to acknowledge John Abatzoglou and Renaud Barbero for their work that was done in parallel with this analysis. We would also like to thank the members of the AirFire team for their continued support and suggestions throughout this project's lifetime. We are also grateful to the contributions of three anonymous reviewers whose suggestions greatly enhanced the quality of this manuscript.

**Conflicts of Interest:** The authors declare no conflict of interest.
