*3.1. Important Predictors of Very-Large Fires*

The diversity of predictors used in the PETs was high, and the important meteorological variables varied by region, the summary statistic, and the type of fire probability. Temperature metrics, in particular seasonality, are a commonly utilized weather predictor in LF forests, and in 10 of the 16 LF forests, seasonality is present in 90 or more of the PETs. On the other hand, while temperature metrics are frequently utilized when constructing PETs, they are not always the optimal splitting criterion. For instance, in the Savanna, Prairie, and Hot Continental Regime Mountains, precipitation metrics overwhelmingly replace temperature based metrics as the optimal discriminant of large and no fire months, and in other regions such as the Subtropical and Hot Continental Division, this designation is highly uncertain.

The importance of temperature metrics also varied by the type of fire probability considered, with temperature metrics more commonly identified as the optimal split criterion in LF forests compared to VLF forests. This sensitivity of PET structure to the type of fire probability could also arise in other ways. For example, in the Hot Continental Regime Mountains, the LF forest overwhelmingly relies on precipitation metrics for prediction, while the corresponding VLF forest utilizes no predictors and reports a constant conditional VLF probability. Similarly, in the Tropical/Subtropical Regime Mountains and Prairie divisions, conditional VLF forests tended to identify wind metrics as the optimal split criterion much more frequently than in the LF forests. Additionally, PET complexity tended to be lower in VLF forests than in the LF forests. The average number of variables used per PET, size, and the number of leaves were inflated in the latter, and weather invariant null models were only ever observed in the VLF forests. The variability in the optimal splitting criterion was also higher in the VLF forests, suggesting a relative lack of certainty regarding the optimal discriminant in conditional VLF probabilities compared to LF probabilities. Weighting did not appear to drastically influence the relative contribution of the weather predictors within each forest (Figure 3).

**Figure 3.** Summary statistics of each forest of probability estimation trees. The top panels show the percentage of probability estimation trees (PETs) in each forest that uses a particular weather predictor for at least one split. The bottom panels show the percentage of PETs for which a weather predictor is selected as the optimal splitting criterion. The relative contribution of each first-split variable under the unweighted (left) and weighted (right) averaging methods are displayed side-by-side.

### *3.2. Climate Change and Very-Large Fire Occurrence*

In most divisions, the expected number of VLFs is predicted to increase in 2050–2099 compared to 1956–2005. The Marine Regime Mountains Redwood Forest division is predicted to have the largest absolute increase with about 13 additional fires per decade under the RCP 4.5 scenario and about 18 additional fires per decade under the RCP 8.5 scenario. For most of the regions under consideration, the average predicted increase ranges from near-zero to several additional VLF per decade relative to historical predictions. In some regions, like Mediterranean and Savanna divisions, the multi-model average predicts slight decreases in VLF activity. The largest absolute decrease occurred in Mediterranean California, which is predicted to have about one less VLF per decade relative to historical predictions under both RCP scenarios. In general, increases in VLF frequency are more severe under the RCP 8.5 scenario than under the RCP 4.5 scenario, although the sensitivity to RCP scenario varies by division (Figure 4). The largest absolute difference in average VLFs per decade between the RCP 8.5 and RCP 4.5 scenarios was in the Marine Regime Redwood Forest division, which had about 5 additional VLFs in the RCP 8.5 scenario. Although the Hot Continental Regime Mountains predicts a larger VLF count per decade under the RCP 4.5 scenario than the RCP 8.5, the difference is negligibly small. The median difference between the RCP 8.5 and RCP 4.5 scenario across the 16 ecoregions was 0.9 additional VLFs per decade under the RCP 8.5 scenario.

**Figure 4.** Kernel density estimates of the posterior and mean of the number of additional very-large fires per year relative to the 1956–2005 reference period per year by ecoregion under the representative concentration pathway (RCP) 4.5 (blue) and RCP 8.5 (red) scenarios arranged by magnitude of change. The excess very-large fire frequency is calculated by randomly sampling (*n* = 106) from the posterior of historical (1956–2005) and future (2050–2099) multi-model averages and calculating the difference.

Future changes in VLF frequency may or may not be uniformly distributed throughout the year. The largest absolute monthly changes in VLF frequency are observed in the Marine Regime Mountain Redwood Forest division during the summer months, while the shoulder months are not predicted to drastically differ from present day VLF frequency. In contrast to the predictions in the Marine Regime

Mountain Redwood Forest division, semi-uniform changes in VLF frequency are also predicted in some regions. For instance, the Subtropical division is predicted to have about 6-8 additional VLF events during the last half of the 21st century relative to the 1956–2005 reference period, but shows no strong preference as to what month these events will occur. For nearly all regions and months, VLF frequency is predicted to increase or show no change compared to historical reference conditions, with the Prairie division being an example of the former and the Hot Continental Regime Mountains the latter. The Mediterranean division is an exception to this pattern, as reductions in future very-large fire frequency are predicted from October to May (Figure 5).

**Figure 5.** Predicted intra-annual changes in very-large fire frequency across sixteen biogeographical regions within the Continental United States under the RCP 4.5 (blue) and RCP 8.5 (red). The central 90th percentile and mean of the excess very-large fires are based on 1,000,000 random samples of the posterior multi-model average very-large fire probabilities from the historical (1956–2005) and future (2050–2099) scenarios.

The changes in overall VLF frequency are predicted to be a result of changes in both model components: the LF and conditional VLF occurrence probabilities. For divisions like Marine Regime Mountains Redwood Forest, both probabilities increase, implying that the LF months will become increasingly frequent and a larger proportion of the months classified as LF will become VLF months. Other divisions showed increases in only one of the model components. In the Temperate Steppe Regime Mountain division, only conditional VLF probabilities are predicted to increase, and in the Tropical/Subtropical Steppe division, only LF probabilities are anticipated to increase. Significant decreases in the model components are only predicted in the Mediterranean and Tropical Subtropical Regime Mountains divisions, which respectively have decreases in the conditional VLF and LF probability components in 2050–2099 compared to 1956–2005 climate model forcings. The Mediterranean LF probability components are predicted to increase, while the conditional VLF probability component is expected to remain the same in the Tropical Subtropical Regime Mountains division. In general, the changes in model components are greater in the RCP 8.5 scenario compared to the RCP 4.5 scenario, although the differences between the two future scenarios were nearly imperceptible in some regions (Figure 6).

**Figure 6.** Simulated change in monthly multi-model large-fire and conditional very-large fire probability estimates across biogeographical divisions of the continental United States. The point cloud is a sample of 100,000 differences in average posterior probability components under the historical (1956–2005) and future scenarios (2050–2099); with the RCP 8.5 scenario colored red and RCP 4.5 colored blue. The solid black lines represent the central 90th percentile and the dashed lines are horizontal and vertical lines passing through the origin.
