Observations show that diversity of deciduousness contributes to successful alternative strategies for tropical forest response to water stress (Williams et al., 2008). For example, during the severe 1997 El Niño drought, brevi-deciduous trees and deciduous stem succulents within a tropical dry site in Guanacaste, Costa Rica, retained leaves during the extreme wet-season drought, behaving differently than during normal dry seasons (Borchert et al., 2002). Both models here predict that neither seasonal deciduousness nor drought-deciduous phenology at the seasonally dry tropical forest, Palo Verde (which consists of trees with different leaf phenological strategies), acts to buffer the forest from a large drop in LAI during UCEs (Fig. S1a–b). Even with this large decrease in LAI, ED2 predicted a very weak biomass loss at the time of UCEs (Fig. 2a), suggesting large-scale leaf loss is not a direct mechanism of plant mortality in ED2. Leaf loss is one component of total carbon turnover flux equations in terrestrial models, in addition to woody loss, fine roots, and reproductive tissues. Having a better understanding of when extreme levels of phenological turnover contribute to stand-level mortality could be improved. Among other turnover hypotheses explored, Pugh et al. (2020) found that phenological turnover fluxes where just as important as mortality fluxes in driving forest turnover time in the VDMs: LPJ-GUESS, CABLE-POP, and ORCHIDEE but not the LSM JULES. At the EucFACE site prior to the simulated extreme drought, LPJ-GUESS displayed strong inter-annual variability in LAI (Fig. S1a–b). This capability of large swings in LAI (5.8 to 0.8) by LPJ-GUESS could contribute to model uncertainty and the considerable mortality response at EucFACE.  Modeled LAI was the largest source of variability in another ecosystem model, CABLE, when evaluating the simulated response to CO2 fertilization (Li et al., 2018). VDMs could be improved by better capturing different plant phenological responses to UCEs by better representing a range of leaf-level morphological and physiological characteristics relevant to plant-water relations such as leaf age, retention of young leaves even during extreme droughts (Borchert et al., 2002), and variation in hydraulic traits as a function of leaf habit (Vargas et al., 2021) (Table 3). Two such examples are seen in the FATES model, where the possibility for “trimming” the lowest leaf layer can occur when leaves are in negative carbon balance due to light limitation, thus optimizing maintenance costs and carbon gain, as well as leaf age classifications, providing variations in leaf productivity and turnover.

Susceptibility of plants to hydraulic stress is one of the strongest determinants of vulnerability to drought, with loss of hydraulic conductivity being a major predictor of drought mortality in temperate (McDowell et al., 2013; Anderegg et al., 2015; Sperry and Love, 2015; Venturas et al., 2021) and tropical forests (Rowland et al., 2015; Adams et al., 2017b), as well as a tractable mortality mechanism to represent in process-based models (Choat et al., 2018; Kennedy et al., 2019). Both LPJ-GUESS and ED2 exhibited a wide range in amount and pattern of plant-available water prior to drought (Fig. S1c–d), contributing to large differences in UCE response. LPJ-GUESS, which does not simulate hydrodynamics, predicted lower total plant-available water at both sites compared to ED2 and subsequently simulated greater mortality and a greater increase in plant-available water right after the UCEs as a result of less water demand. Due to ED2 using a static mortality threshold from conductivity loss (88 %), it likely does not accurately reproduce the wide range of observations of drought-induced mortality. In ED2, large trees with longer distances to transport water were at higher risk and suffered higher mortality (Fig. 4), demonstrating how stand demography, size structure, and tapering of xylem conduits can play an important role in ecosystem models (Petit et al., 2008; Fisher et al., 2018). Of the VDMs that are beginning to incorporate a continuum of hydrodynamics, e.g., ED2 (described in Methods, Sect. 2.1) and FATES-HYDRO (Fang et al., 2022, based on Christoffersen et al., 2016), they are able to solve for transient water from soils to roots, through the plant, and connect with transpiration demands. Therefore, instead of the plant-water stress function being based on soil water potentials, it is replaced with more realistic connections to leaf water potentials. Mortality is then caused by hydraulic failure via embolism controlled by the critical water potential (P50) that leads to 50 % loss of hydraulic conductivity. For advancements in tree-level hydrodynamic modeling see the FETCH3 model (Silva et al., 2022); for justification for plant hydrodynamics in conjunction with multi-layer vertical canopy profiles see Bonan et al. (2021). There are strong interdependencies and related mechanisms connecting both hydraulic failure (e.g., low soil moisture availability) and C limitation (e.g., stomatal closure) during drought (McDowell et al., 2008; Adams et al., 2017b), and these interactions should be incorporated into ecosystem modeling and further explored (Table 3).

Plants have a variety of strategies to buffer vulnerability to water and nutrient stress caused by extreme droughts, such as allocating more C to deep roots (Joslin et al., 2000; Schenk and Jackson, 2005), investing in mycorrhizal fungi (Rapparini and Peñuelas, 2014), or reducing leaf area without shifting leaf nutrient content (Pilon et al., 1996). Alternatively, the presence of deep roots does not necessarily lead to deep soil moisture utilization, as seen in a 6-year Amazonian throughfall exclusion experiment where deep root water uptake was still limited, even with high volumetric water content (Markewitz et al., 2010). Elevated CO2 alone will enhance growth and water-use efficiency (Keenan et al., 2013), reducing susceptibility to drought. However, such increased productivity within a forest stand, and associated structural overshoot during favorable climate windows, can also be reversed by increased competition for light, nutrients, and water during unfavorable UCEs – potentially leading to mortality overshoot (Fig. 1d) and higher C loss. Mortality overshoot, as a result of structural overshoot, could be an explanation for the negative severity-climate index (i.e., C loss) in the majority of eCO2-only simulations (18 out of 24 scenarios; Table 2).

Effects of CO2 fertilization on plant C allocation strategies are uncertain. As a result, ecosystem models differ in their assumptions on controls of C allocation in response to eCO2, leading to divergent plant C use efficiencies (Fleischer et al., 2019). Global-scale terrestrial models are beginning to include optimal dynamic C allocation schemes, over fixed ratios, that account for concurrent environmental constraints on plants, such as water, and adjust allocation based on resource availability such as in LM3-PPA (Weng et al., 2015), but the representation of C allocation is still debated and progressing (De Kauwe et al., 2014; Montané et al., 2017; Reyes et al., 2017). Options for carbon allocation strategies can be based on the allometric partitioning theory (i.e., allocation follows a power allometry function between plant size and organs, which is insensitive to environmental conditions; Niklas, 1993) as an alternative to ratio-based optimal partitioning theory (i.e., allocation to plant organs based on the most limiting resources) (McCarthy and Enquist, 2007) or fixed ratios (Table 3), and the strategies should be further investigated, particularly due to VDMs' substantial use of allometric relationships. A meta-analysis of 164 studies found that allometric partitioning theory outperformed optimal partitioning theory in explaining drought-induced changes in C allocation (Eziz et al., 2017). Further eco-evolutionarily based approaches such as optimal response or game-theoretic optimization, as well as entropy-based approaches, are useful when wanting to simulate higher levels of complexity (reviewed in Franklin et al., 2012). With more frequent UCEs and the need for plants to reduce water consumption, a shift in the optimal strategy of allocation between leaves and fine roots should change. The goal functions (e.g., fitness proxy) used in optimal response modeling can account for these shifts in costs and benefits of allocation between all organs (Franklin et al., 2009, 2012).

Studies of neotropical and temperate seedlings show that pre-drought storage of non-structural carbohydrates (NSCs) provides the resources needed for growth, respiration osmoregulation, and phloem transport when stomata close during subsequent periods of water stress (Myers and Kitajima, 2007; Dietze and Matthes, 2014; O'Brien et al., 2014). Furthermore, direct correlations have been shown between NSC depletion and embolism accumulation and between the degree of pre-stress reserves and the utilization of soluble sugars (Tomasella et al., 2020). The amount of NSC storage required to mitigate plant mortality during C starvation and interactions with hydraulic failure from severe drought is difficult to quantify due to the many roles of NSCs in plant function and metabolism (Dietze and Matthes, 2014). For example, NSCs were not depleted after 13 years of experimental drought in the Brazilian Amazon (Rowland et al., 2015). As atmospheric CO2 increases with climate change, NSC concentrations may increase, as seen in manipulation experiments (Coley et al., 2002), but interactions with heat, water stress, enhanced leaf shedding, and nutrient limitation complicate this relationship and need to be further explored. Despite the recognition of the critical role that plant hydraulic functioning and NSCs play in tree resilience to extremes, knowledge gaps and uncertainties preclude fully incorporating these processes into ecosystem models.

Compared to ED2, LPJ-GUESS predicted low plant carbon storage (a model proxy for NSCs) prior to and during drought and at times became negative, thereby creating C costs (Fig. S2a–b), leading to C starvation and potentially explaining the larger biomass loss in LPJ-GUESS at both sites. Alternatively, ED2 maintained higher levels of NSCs, providing a buffer to stress and mitigating the negative effects of drought. Maintenance of NSCs in ED2 even during prolonged drought (at EucFACE) is due to (1) trees resorbing a fraction of leaf C during leaf shedding, (2) no maintenance costs for NSC storage in the current version, and (3) no allocation of NSCs to structural growth until NSC storage surpasses a threshold (the amount of C needed to build a full canopy of leaves and associated fine roots), allowing a buffer to accumulate. In LPJ-GUESS, accumulation and depletion of NSC are recorded as a C debt being paid back in later years. The contrasting responses of the two models to drought, and the likely role of NSCs in explaining differences in model behavior, highlight the need to better understand NSC dynamics and to accurately represent the relevant processes in models (Richardson et al., 2013; Dietze and Matthes, 2014). More observations of C accumulation patterns and how/where NSCs drive growth, respiration, transport, and cellular water relations would enable a more realistic implementation of NSC dynamics in models (Table 3).

Currently, LPJ-GUESS simulates the Palo Verde community using three PFTs, while ED2 uses four PFTs that differ in photosynthetic and hydraulic traits. The community composition simulated by ED2 is shown to be more resistant to UCEs compared to LPJ-GUESS (Fig. 5), perhaps due to relatively higher functional diversity (via more PFTs with additional phenological and hydraulic diversity). This additional diversity helps to buffer ecosystem response to drought by allowing more tolerant PFTs to benefit from reductions in less tolerant PFTs, thus buffering reductions in ecosystem function (Anderegg et al., 2018). Higher-diversity ecosystems were found to protect individual species from the negative effects of drought (Aguirre et al., 2021) and enhance productivity resilience following wildfire (Spasojevic et al., 2016); thus, functionally diverse communities may be key to enhancing tolerance to rising environmental stress.

Recent efforts to consolidate information on plant traits (Reich et al., 2007; Kattge et al., 2011) have contributed to identifying relationships that can impact community-level drought responses (Skelton et al., 2015; Anderegg et al., 2016a; Uriarte et al., 2016; Greenwood et al., 2017), such as life-history characteristics, and strategies of resource acquisition and conservation as predictors of ecosystem resistance (MacGillivray and Grime, 1995; Ruppert et al., 2015). While adding plant trait complexity in ESMs may be required to accurately simulate key vegetation dynamics, it necessitates more detailed parameterizations of processes that are not explicitly resolved (Luo et al., 2012). Further investigation of how VDMs represent interactions leading to functional diversity shifts is crucial to this issue. Enquist and Enquist (2011), as an example, show that long-term patterns of drought (20 years) have led to increases in drought-tolerant dry forest species, which could modulate resistance to future droughts. Higher diversity of plant physiological traits and drought-resistance strategies is expected to enhance community resistance to drought, and models should account for shifts in diverse functionality (Table 3).

Our simulation results generally demonstrated a fast recovery of plant-available water and LAI at both sites (Fig. S1). Annual plant-available water substantially increased right after a drought by an average of 163 mm at Palo Verde and 213 mm at EucFACE in the LPJ-GUESS simulations compared to much lower increases in ED2 (50 and 12 mm at Palo Verde and EucFACE). This increase in available water post-drought can be attributed to reduced stand density and water competition (Fig. S2c–d; diamonds vs. circles), alleviating the demand for soil resources (water) and subsequent stress, which has also been shown in observations (McDowell et al., 2006; D'Amato et al., 2013). After large canopy tree mortality events, there can be relatively rapid recovery of forest biogeochemical and hydrological fluxes (Biederman et al., 2015; Anderegg et al., 2016b; Biederman et al., 2016). These crucial fluxes strongly influence plant regeneration and regrowth, which can buffer ecosystem vulnerability to future extreme droughts. However, this enhanced productivity has a limit. In a scenario where UCEs continue to intensify, causing greater reductions in soil water and reduced ecosystem recovery potential, the SO growth that typically occurs after UCEs may be dampened (Fig. 1d). In water-limited locations, similar to the dry forest sites used here, initial forest recovery from droughts was faster due to thinning-induced competitive release of the surviving trees and shallow roots not having to compete with neighboring trees for water, allowing for more effective water usage (Tague and Moritz, 2019), stressing the importance of root competition and distribution in models (Goulden and Bales, 2019). Tague and Moritz (2019) also reported that this increased water-use efficiency and SO ultimately led to water stress and related declines in productivity, similar to the MO concept (Jump et al., 2017; McDowell et al., 2006). Since a core strength of VDMs is predicting stand demography during recovery, improved quantification of density-dependent competition following stand dieback would be beneficial for model benchmarking (Table 3).

Time lags in forest compositional response to and survival of drought could indicate community resistance or shifts to more competitive species and competitive exclusion. During a 15-year recovery period from extreme drought at Palo Verde, LPJ-GUESS predicted an increase in stem density (stems m2 yr-1) (Fig. S2c) compared to ED2, which predicted almost no impact in stem recovery. The mortality “spike” in ED2 due to drought was muted and slightly delayed, contributing to ED2's lower biomass loss and more stable behavior of plant processes over time at Palo Verde. At EucFACE, both models exhibited a pronounced lag effect in stem turnover response, i.e., ∼ 8–12 years after drought (Fig. S2d). After about a decade, strong recoveries and increased stem density occurred, which in ED2 was followed by delayed mortality/thinning of stems. Delayed tree mortality after droughts is common due to optimizing carbon allocation and growth (Trugman et al., 2018) but typically only up to several years post-drought, not a decade or more as seen in the model.

The versions of the VDMs used here do not directly consider post-drought secondary stressors such as infestation by insects or pathogens and the subsequent repair costs due to stress damage, which could substantially slow the recovery of surviving trees. Forest ecologists have long recognized the susceptibility of trees under stress, particularly drought, to insect attacks and pathogens (Anderegg et al., 2015). Tight connections between drought conditions and increased mountain pine beetle activity have been observed (Chapman et al., 2012; Creeden et al., 2014) and can ultimately lead to increased tree mortality (Hubbard et al., 2013). Leaf defoliation is a major concern from insect outbreaks following droughts and can have large impacts on C cycling, plant productivity, and C sequestration (Amiro et al., 2010; Clark et al., 2010; Medvigy et al., 2012). Implementing these secondary stressors in models could slow the rate of post-UCE recovery and lead to increased post-UCEs tree mortality. Additional background on secondary disturbances, lag effects, and repeated extremes can be found in Sect. S2 in the Supplement.

Change in stand structure is an important model process to capture because large trees have important effects on C storage, community resource competition, and hydrology (Wullschleger et al., 2001) (Table 3), and maintaining a positive carbohydrate balance is beneficial to sustaining (or repairing) hydraulic viability (McDowell et al., 2011). There is increasing evidence, both theoretical (McDowell and Allen, 2015) and empirical (Bennett et al., 2015; Rowland et al., 2015; Stovall et al., 2019), that large trees (particularly tall trees with high leaf area) contribute to the dominant fraction of dead biomass after drought events. Under rising temperatures (and decreasing precipitation), vapor pressure deficit (VPD) will increase, leading to a higher likelihood of large-tree death (Eamus et al., 2013; Stovall et al., 2019), driving MO events as hypothesized in Fig. 1d. Consistent with this expectation, ED2 predicted that the largest trees (>100 cm) experienced the largest decreases in basal area compared to all other size classes (Fig. 4). This drought-induced partial dieback and mortality of large dominant trees have substantial impacts on community-level C dynamics, as long-term sequestered C is liberated during the decay of new deadwood (Palace et al., 2008; Potter et al., 2011). In ED2, the intermediate size class (60–80 cm) increased in basal area following large-tree death, taking advantage of the newly open canopy space. However, small size classes do not necessarily benefit from canopy dieback. For example, in a dry tropical forest, prolonged drought led to a decrease in understory species and small-sized stems (Enquist and Enquist, 2011).

Due to VDMs being able to exhibit dynamic biogeography, they are more useful at predicting shifts in community composition beyond LSMs capabilities. Further areas of advancement (described in Franklin et al., 2020) include models of natural selection, self-organization, and entropy maximization, which can substantially improve community dynamic responses in varying environments such as UCEs. Eco-evolutionary optimality (EEO) theory can also help improve functional trait representation in global process-based models (reviewed in Harrison et al., 2021) through hypotheses in plant trait trade-offs and mechanistic links between processes such as resource demand, acquisition, and a plant's competitiveness and survival, traits associated with high degrees of sensitivity in models. The power of prognostic VDMs to predict shifts in demography and community migration with climate change is large but is rarely being constrained with plant-level EEO theory and thus will likely need to use stand-level competition and coexistence principles of how plants self-organize (Franklin et al., 2020).

In field experiments, higher disturbance rates have shifted the recovery trajectory and competition of the plant community towards one that is composed of opportunistic fast-growing pioneer tree species, grasses (Shiels et al., 2010; Carreño-Rocabado et al., 2012), and/or deciduous species, as also seen in the model results (Hickler et al., 2004). In the treatments presented here, deciduous PFTs were also the strongest to recover after 15 years in both models, surpassing pre-drought values (Fig. 5). It should be noted that ED2 exhibited a strong recovery in the evergreen PFT as well, inconsistent with the above literature (Fig. 5b). PFTs in ED2 respond to drought conditions via stomatal closure and leaf shedding, buffering stem water potentials from falling below a set mortality threshold (i.e., 88 % of loss in conductivity). This conductivity threshold may need to be reconsidered if further examination reveals an unrealistic advantage under drought conditions for evergreen trees, which exhibited a lower impact from droughts (compared to deciduous and brevi-deciduous PFTs) in ED2. Nitrogen cycling feedbacks were not investigated here but could also be an explanation for a strong evergreen PFT recovery.

Recovery of surviving trees could be hindered by the high cost of replacing damaged xylem associated with cavitation (McDowell et al., 2008; Brodribb et al., 2010). Many studies have identified “drought legacy” effects of delayed growth or gross primary productivity following a drought (Anderegg et al., 2015; Schwalm et al., 2017), and the magnitude of these legacies across species correlates with the hydraulic risks taken during a drought itself (Anderegg et al., 2015). The conditions under which xylem can be refilled remain controversial, but it seems likely that many species, particularly gymnosperms, may need to entirely replace damaged xylem (Sperry et al., 2002), and trees worldwide operate within narrow hydraulic safety margins, suggesting that trees in all biomes are vulnerable to drought (Choat et al., 2012). The amount of damaged xylem from a given drought event and recovery rates also vary across trees of different sizes (Anderegg et al., 2018).

Plasticity in nutrient acquisition traits, intraspecific variation in plant hydraulic traits (Anderegg et al., 2015), and changes in allometry (e.g., Huber values) can have large effects on acclimation to extreme droughts. This suggests some capacity for physiological adaptation to extreme drought, as seen by short-term negative effects from drought and heat extremes being compensated for in the longer term (Dreesen et al., 2014). Still, given the shift towards more extreme droughts with climate change, vegetation mortality thresholds are likely to be exceeded, as reported in Amazonian long-term plots where mortality of wet-affiliated genera has increased, while simultaneously new recruits of dry-affiliated genera are also increasing (Esquivel-Muelbert et al., 2019). Increasing occurrences of heat events, water stress, and high VPD will lead to extended closure of stomata to avoid cavitation, progressively reducing CO2 enrichment benefits (Allen et al., 2015). Where CO2 fertilization has been seen to partially offset the risk of increasing temperatures, the risk response was mediated by plant hydraulic traits (Liu et al., 2017) using a soil–plant–atmosphere continuum (SPAC) model, yet interactions with novel extreme droughts were not considered. The VDM simulations suggest that the combination of elevated warming and potential structural overshoot from eCO2 (or inaccurate representation in NSC allocation/usage priority) will exacerbate consequences of UCEs by reductions in both C stocks and post-drought biomass recovery speeds (Fig. 3). Therefore, future UCE recovery may not be easily predicted from observations of historical post-disturbance recovery. An associated area for further investigation is to better understand the hypothesized interplay between amplified mortality from hotter UCEs followed by structural overshoot regrowth during wetter periods (Fig. 1d), which could potentially lead to continual large swings in MO and SO and vulnerable net ecosystem C fluxes through time (Table 3).