Antarctic Peninsula ice shelf collapse triggered by föhn wind-induced melt 5

Ice shelf collapse reduces buttressing and enables glaciers to contribute more rapidly to sea-level rise in a warming climate. The abrupt collapses of the Larsen A and B ice shelves on the Antarctic Peninsula (AP) have been attributed to increased surface melt. However, no studies examine the timing, magnitude, and location of surface melt processes immediately preceding these disintegrations. Here we use a regional climate model and Machine Learning 15 analyses to evaluate the influence of föhn wind events on the surface liquid water budget for collapsed and extant ice shelves. We find föhn winds caused 25% of the total annual melt in just 9 days on Larsen A which helped melt lakes surpass a critical stability depth that, we suggest, ultimately triggered collapse. By contrast, föhns appear to pre-condition, not trigger, Larsen B's collapse. AP extant ice shelves will remain less vulnerable to surface-melt-driven instability due to weaker föhn-driven melt so long as surface temperatures and föhn occurrence remain within historical bounds.

melt to the climatological surface liquid water budget comparing collapsed and extant ice shelves. By constructing a timeline of melt and melt mechanisms and comparing melt metrics with collapsed and extant ice shelves, we can identify the contributing factors that caused collapse.

Regional Climate Model Data (RACMO2)
We base our analysis on 3-hourly output from simulations by the Regional Atmospheric Climate Model 2 (RACMO2), version 2.3p2, with a horizontal resolution of 5.5km (0.05°) focused on the AP from 1979-2018. RACMO2 uses the physics package CY33r1 of the ECMWF Integrated Forecast System (IFS) 80 ( https://www.ecmwf.int/en/elibrary/9227-part-iv-physical-processes\textit{{ECMWF-IFS ,} 2008}) in combination with atmospheric dynamics of the High-Resolution Limited Area Model (HIRLAM). RACMO2 has been evaluated against numerous surface observations (AWS) in locations all over the AP and has trouble simulating very high and low-temperature extremes in the region but is considered a good representation of surface conditions (Leeson et al., 2017;Laffin et al., 2021).

Föhn wind detection
85 We developed a Föhn Detection Algorithm (FöhnDA) that identifies föhn winds that cause melt using 12 Automatic Weather Stations (AWS) on the AP as detailed in Laffin et al., (2021). FöhnDA identifies föhn-induced melt events using binary classification when air temperature (T) is greater than 0°C, which ensures it captures föhn events that cause surface melt.
Thresholds for relative humidity (RH) and wind speed (WS) are more dynamic because high wind speeds and low relative humidity do not guarantee temperatures above freezing, they only aid to identify föhn. FöhnDA uses quantile regression to 90 identify these variable thresholds that take into account the climatology and seasonality at each weather station site. FöhnDA uses two empirically determined thresholds: the 60th percentile wind speed and 30th percentile relative humidity which are 2.85 m/s and 79% averaged at all AWS locations. We co-locate AWS with the nearest model grid cell and use FöhnDA results to train an ML model that detects föhn winds in RACMO2 output. Our ML model improves the accuracy of föhn detection by over 23% when compared to the simple binary classification method applied to RACMO2 output as described 95 above. This method is the most accurate detection method compared to previous work and allows us to use in situ observations from AWS and expand föhn detection with RACMO2 output to regions and times when AWS observations are not available ( Figure S2).
Föhn jet locations were identified using wind direction and strength during föhn events ( Figure 2a) and by the surface melt pattern during föhn (Figure 3b). The RACMO2 topography pixel size is 5.5 km which is sufficient to produce 100 the föhn jets identified on the LCIS (Elvidge et al., 2015), and allows for new föhn jet identification on the LAIS and LBIS despite lack of direct observation.. However, small-scale föhn winds funneled through local canyons and mountain gaps 4 https://doi.org/10.5194/tc-2021-301 Preprint. Discussion started: 25 October 2021 c Author(s) 2021. CC BY 4.0 License. smaller than 5.5 km are not directly simulated. Therefore, we consider RACMO2 simulated estimates of surface melt caused by föhn winds to be conservative and likely higher in regions where föhn winds are funneled and concentrated.

Ice shelf intercomparison analysis
105 We split each of the five ice shelves shown in Figure 1a and take the average of all model grid cells annually to create a climatology of surface melt, melt rate, melt hours, surface temperature. We use a two-tailed t-test statistic to identify if the mean of both ice shelves is statistically different from one another at the 95% confidence interval. We compare all ice shelves to the LBIS because it was the most recent collapse event and is adjacent to collapsed and existing ice shelves.
Qualitatively similar results are obtained when comparing all ice shelves to the LAIS.

Föhn jets and melt
Using RACMO2 historical simulations, informed by a Machine Learning algorithm that is trained with Automatic Weather Station (AWS) observations (Laffin et al., 2021), we identify seven recurring föhn jets or gap winds that lead to high surface melt rates on the eastern AP ice shelves ( Figure 2a). Four of these jets (CI, MI, WI, MOI) have been studied using airborne 115 observations and model simulations (Grosvenor et al., 2014;Elvidge et al., 2016). The remaining three jets (LA, LB, and JP) are, to our knowledge, identified here for the first time. Föhn winds form when moist air is forced over a mountain barrier, often leading to precipitation on the windward side of the barrier that dries the air mass (Elvidge et al., 2016). As the now drier air descends the leeward slope it warms adiabatically and promotes melt directly through sensible heat exchange, and indirectly by the associated clear skies that allow additional shortwave radiation to reach the surface in non-winter months 120 (Elvidge et al., 2020;Laffin et al., 2021). These positive energy balance components increase surface melt rates up to 54% relative to non-föhn induced melt ( Figure 2b). Additionally, AP winds from the west and northwest (föhn influence) produce including föhn-induced melt rate, lessens the influence of all the other factors that contributed to these collapses. A clearer picture of the role of föhns emerges after we examine föhn-induced melt extent and timing.
The spatial distribution and extent of surface melt influence ice shelf stability. Surface melt and melt lakes near the ice shelf terminus can lead to calving front collapse and structural instability for the remaining portion of the ice shelves 145 (Depoorter et al.,2013;Pollard et al., 2015). Consistent with this mechanism, the LA and LB föhn jets impact a large spatial area of the LAIS and LBIS, and reach the ice shelf calving fronts (Figure 3b). SCAR Inlet lacks a strong föhn jet/influence and does not regularly experience largescale melt lakes even during high melt years (Figure 1b-f). This helps explain why SCAR Inlet is still intact, despite major structural changes observed after the collapse of the LBIS (Borstad et al., 2016;Qiao et al., 2020). LCIS on the other hand is impacted by four major jets and regularly experiences föhn-induced melt lakes, 150 particularly in Cabinet inlet. However, the vast size of the LCIS does not allow the föhn-induced melt to reach the terminus.
The föhn melt mechanism breaks down by mixing with cold air which reduces the intensity of the föhn jets from their peak at the base of the AP mountains to the calving front (Figure 3b). Having established that föhn winds significantly enhanced surface melt overall and at the crucial calving front of LAIS and LBIS, we now examine the timing of föhn-induced melt events relative to the collapses.

Föhn melt and the surface liquid water budget
To better understand the role that föhn winds have played in AP ice shelf surface melt and stability we intercompare melt 200 climatologies of all major ice shelves. Comparing collapsed with intact ice shelves yields a clearer picture of the effects föhn winds have on ice shelf stability. We identify whether annual surface melt production, melt rate, melt hours, and surface  Table S1). We compare to LBIS because it was centered between other ice shelves and was the most recent to collapse. Total surface melt production on every ice shelf except LAIS differs significantly from LBIS melt (Mean annual 205 melt; LAIS-476 mm w.e., LBIS-479 mm w.e., SCAR-353 mm w.e., Larsen(north)-336 mm w.e., LCIS-238 mm w.e.) ( Figure   5a), which is expected when we consider the latitudinal location and mean annual air temperature (Figure 5d) (Table S1).
However, when föhn-induced melt is subtracted from total melt, the mean annual surface melt production on SCAR inlet and Larsen C (north) are not statistically different from the LBIS (LAIS-337 mm w.e., LBIS-321 mm w.e., SCAR-286 mm w.e., Larsen(north)-278 mm w.e., LCIS-203 mm w.e.) (Figure 5b). In other words, with the exception of föhn-induced melt 210 (Figure 5c), melt production on SCAR Inlet and LCIS are statistically indistinguishable at the 95% confidence interval from LBIS melt production. Föhn wind-induced surface melt impacted the collapse significantly more than extant ice shelves further implicates föhn melt as an important contributor to LAIS and possibly LBIS collapse. The relatively small role that föhns play in the liquid water production and variability on the remaining ice shelves bodes well for their continued resilience.

215
The nominal amount of föhn-induced melt on the LBIS in the 2001/02 melt season nevertheless played a role in ice shelf stability through firn densification. Firn densification occurs when the liquid water fills the pore space between snow/ice crystals decreasing the air content in the firn, which forms refrozen ice layers that promote melt lake formation (Kuipers Munneke et al., 2012;Polashenski et al., 2017). A liquid-to-solid ratio (LSR) is a crude proxy for available firn air content and can be estimated as,

T otal liquid water (snowmelt + liquid precipitation)
where areas with LSR < 1 represent an ice shelf that receives more solid precipitation than liquid water and is therefore less likely to saturate with liquid water and form melt lakes than areas with LSR > 1 (Figure 6). Extant ice shelves (SCAR inlet,  It is reasonable to expect differences in ice shelf melt regime, particularly with the north/south temperature gradient present on the eastern AP ice shelves. The annual surface temperature difference between ice shelves could explain ice shelf disintegration through long-term thinning and retreat (Scambos et al., 2003;Morris et al., 2003), however, temperature gradient alone cannot explain the substantial increase in surface melt on the LAIS and LBIS. Only with the addition of 245 föhn-induced surface melt do the LAIS and LBIS stand out significantly from the other eastern AP ice shelves. With that in mind, we have examined liquid water processes on the spatio-temporal scales pertinent to AP ice shelf stability. For instance, the structural flow discontinuities or suture zones, where tributary glaciers merge together to form an ice shelf, are mechanically weak points that impact stability (Glasser et al., 2008). These suture zones are further weakened through lateral shear depending on the difference in tributary glacier flow. All ice shelves in the region are comprised of numerous outflow 250 glaciers sutured together, and while some studies suggest this is a major contributor to ice shelf instability, only two of the ice shelves have collapsed (Borstad et al., 2016;Glasser et al., 2008;Glasser et al., 2021). Further research suggests that marine accretion of ice on the bottom of the ice shelves, specifically LCIS, may stabilize these suture zones, which may be why SCAR inlet has remained intact despite major rift formation (

Conclusions
The converging lines of evidence in these results show, for the first time, that observed and inferred föhn-driven melt is present in sufficient amounts, and at the right locations and times, to explain the disintegration of Larsen A in 1995 but not Larsen B in 2002. The fact that the LAIS and LBIS collapsed catastrophically within weeks and not through long-term 290 thinning and retreat like other ice shelves (Prince Gustav, Wordie, George VI) suggests sudden disintegration is anomalous and requires forcings to match vulnerabilities (Scambos et al., 2003). We conclude that föhn wind-induced surface melt was a trigger for the collapse of the LAIS but not the LBIS. The remaining AP ice shelves may be more stable, at least from melt-driven instability, than previously thought. We have come to these conclusions with the following forms of evidence:

295
• First, both the LAIS and LBIS are impacted by powerful melt-inducing föhn jets that affect a large spatial portion of each ice shelf that reach the ice shelf terminus. Surface melt and melt lakes near the ice shelf terminus can lead to calving front collapse and structural instability for the remaining portion of the ice shelves (Pollard et al., 2015;Depoorter et al., 2013). Extant ice shelves are either not directly affected by a föhn jet, or are too vast to have any significant effect near the terminus.
• Second, strong föhn winds were present prior to and at the time of collapse for the LAIS. This series of 3 föhn events lasted nine days total and produced over 25% of the total annual melt for the 1994/95 melt season. The enhanced melt, filled new and existing melt ponds above the critical (1 meter) melt lake depth of stability which ultimately triggered large-scale hydrofracture cascades and the LAIS collapse. A föhn event was also present at the onset of the LBIS collapse, however, melt rates were nominal and likely did not produce a trigger effect.

305
• Third, in the absence of föhn wind and concurrent melt, the surface liquid budgets of collapsed and intact ice shelves are climatically similar, which points to föhn winds as a driver of increased surface melt and possibly rapid collapse. The additional melt on the LAIS and LBIS compared to intact ice shelves created impermeable ice layers that support melt lake production, particularly when annual surface melt exceeds annual precipitation.