We present the partition of uncertainty into four sources for decadally averaged transient runs of AMax, AMed, and AMin for the different ensemble setups. We map the relative contributions to the total uncertainty over the entire domain at five time slices from 2010 to 2090 showing results for AC and PC (in SI).
Throughout the twenty-first century and using the full ensemble, the uncertainty in AMax (Fig. 2a) is explained mostly by the GCMs and GIMs, with the latter increasing their contribution to the total uncertainty from the west of the domain spreading to the northwest (around 2030) and to the upper Midwest and northeastern USA; this leads to a reduced importance of the GCMs over the twenty-first century, especially for the northern half of the domain. The share of IVar is large at the beginning of the century over the domain but declines steadily with the exception of the south and southeast, with a contribution to the total variance (15–30%) comparable to that of the GIMs. In contrast and as seen in Fig. 1, the contribution to the total uncertainty from RCPs is very limited compared to GIMs and GCMs, with shares ranging from 5 to 20% of the total variance. In particular, both RCPs and IVar show the smallest contributions in the northern part of the domain (see mean relative contributions per region in Fig. S5a).
AMed (Fig. 3a) shows results similar to AMax: GCMs and GIMs are by far the major contributors, with GIMs increasing their contribution to uncertainty over the domain during the twenty-first century. This is particularly true in the northern and eastern parts of the CONUS, even though shares are more balanced between the two sources, without the clear predominance of GIM contribution as seen, for instance, in the western USA for AMax. Also, GCMs show larger contributions to the total variance both at the beginning of the run and, in general, in the western and northwestern USA compared to AMax (Fig. S6). Interestingly, in the US west and southwest, IVar has high shares from as much as 40% in 2006 to 15–20% in 2095. This is due to the large fluctuations in the projections that make up large deviations from the smooth fit, thereby making the contribution of IVar to uncertainty substantial (as seen in Las Vegas in Fig. S2).
For AMin (Fig. 4a), the similar contribution to the uncertainty by GCMs and GIMs seen for AMax and AMed only holds at the beginning of the twenty-first century, in which IVar, especially through the Rocky Mountains, has a substantial contribution to uncertainty. GIMs increase steadily their share throughout the period. In particular, the GCMs lose their share to the advantage of GIMs in the east, while in the west, IVar shows considerable percentages of about 40%, which slowly reduce to 15–20% at the end of the run. The strong role that IVar plays in the southwest (in the area spanning from the lower Rocky Mountains to the arid areas of the south) can partly be attributed to the difficulty of both GCMs and GIMs in simulating runoff in mountaineous and arid areas, where indeed AMin time series generally suffer from poor simulations with anomalous erratic departures from zero or very low values. With the exception of the US southeast around 2030–2040 (~ 20%), RCPs contribute only marginally to the total variance (Fig. S7).
Results with PC in AMax (Fig. S9a) are very similar to those obtained with AC, although the western part of the domain is partly masked. There are slight increases (5 to 15%) in GCM and IVar uncertainty shares that are taken from the GIM, except for the eastern part of the domain (Fig. S5b). The heavy masking occurring with the use of the oE makes AMed and AMin not comparable.
The exclusion of two of the GIMs (MATSIRO and LPJmL) from the full ensemble has an effect on both relative contributions and total variance. For AMax in particular, patterns are markedly different: the GIM source ceases to dominate over most of the domain (Fig. 2b), and the strong contribution in the northern half is limited to the west and northeast of the domain in favor of the uncertainties in the GCMs. Compared to the oE, the lower variance of the cE is reflected in lower total variance in all regions, particularly the northern ones (Fig. S8a). When we focus on AMed (Fig. 3b), GCM and RCP contributions grow from 5 to 10% over the entire domain, with a reduction in GIM contribution. There are large fractions of GCM uncertainty at the beginning of the run over most of the domain that decline in favor of the GIM uncertainty, which explains most of the uncertainty at the end of the run. Similar to AMed, AMin (Fig. 4b) has close shares of fractional variance to the oE (Fig. 4a) all throughout the run.
The total variance for cE is close to and only slightly higher than that of oE (Fig. S8a), suggesting that MATSIRO and LPJmL projections lie within the range of the other GIMs and that the exclusion of less credible GIMs does not yield lower total variances.
Results for PC in all three indices (Fig. S9b-S14b) depict the same patterns seen for AC. For AMed, the increase in RCP contribution over time in the southeast is even clearer with PC.
If, in addition to MATSIRO and LPJmL, we exclude JULES, the resulting ensemble (cE-noB) depicts marked differences especially for AMax. Indeed, LPJmL and JULES account for vegetation dynamics and varying CO2 (absent in the other GIMs) and tend to project higher runoff depending on the region.
Consistent with what discussed for cE, when we focus on AMax, the cE-noB (Fig. S15a) depicts even lower fractions of GIM uncertainty, in favor of the GCM and IVar. This is especially true in the southern parts of the study region. Both cE and cE-noB have in common the exclusion of LPJmL, which is responsible for contributing the high variance in the northern half of the domain. Substantial changes in the proportion of the total variance result from the exclusion of the three GIMs: the high GIM uncertainty in the northwest and north seen with the oE is no longer present, and GCMs become the dominant source over the entire domain during the whole period. Except for the south and southeast where the GIM contribution was already small (~ 15–30%), GCM and IVar increase their shares. This shows that biome GIMs bring about greater variance to the ensemble for the peak flows, with the exception of the southern/southeastern CONUS (Fig. S8a). It is worth noting that, over the regions covering the southern half of the USA, the RCP contribution to uncertainty is similar to that of GIMs (~ 10–20%) and sometimes greater, reaching 30% in the south and southeast. This suggests that low credibility and biome GIMs can be considered as outliers in the GIMs’ spread; therefore, their exclusion facilitates an emerging signal of the RCP in these regions.
Changes in proportions from oE to cE and cE-noB for AMax are not as pronounced as those obtained for AMed (Fig. S16) and AMin (Fig. S17) for which we see a clear shift from the GIMs’ to the GCMs’ contribution. Namely, for AMed, there is a marked gain in GCM contribution, especially in the northern regions.
The influence of biome GIMs on AMin’s total variance (Fig. S8) is fairly weak from the west to the south of the study region, where these models tend to simulate zero runoff over extended periods of time, leading to lower variance when included in the oE. Elsewhere (Fig. S17), GCMs increase their share by about 10%. As seen for AMed, RCP’s contribution increases (especially in the eastern part of the domain) to the same level of the GCMs; this holds even though GIMs still remain the dominant source of uncertainty. The inclusion of biome models in the ensemble brings about larger shares of GIM uncertainty. In particular, when JULES is present, we see a larger and increasing GIM share in the southern half of the domain. Results for PC in AMax (Fig. S15b), AMed (Fig. S16b), and AMin (Fig. S17b) are in line with those seen for AC.
We also tested the impacts of excluding the intermediate RCPs (4.5 and 6.0) on the model spread. For AMax (Fig. S18a—versus all RCPs in Fig. 2b), differences are barely noticeable, while for AMed (Fig. S19a—versus all RCPs in Fig. 3b), the RCP fraction of total variance has higher values in the eastern USA when compared to the cE, especially in the south and southeast USA after 2050. Similarly, for AMin (Fig. S20a—versus all RCPs in Fig. 4b), the RCP proportions increase at the end of the run, especially in the east. In essence, all indices show proportions similar to the ensemble with four RCPs, although these proportion tend to systematically fluctuate throughout the period, perhaps owing to the use of fewer projections (i.e., decreased information). Overall, the use of fewer RCPs leads to the same results for PC and AC for all indices (Fig. S18b-S20b).