Reconstruction of the 1941 GLOF process chain at Lake Palcacocha ( Cordillera Blanca , Perú )

19 The Cordillera Blanca in Perú has been the scene of rapid deglaciation for many decades. One of numer20 ous lakes formed in the front of the retreating glaciers is the moraine-dammed Lake Palcacocha, which 21 drained suddenly due to an unknown cause in 1941. The resulting Glacial Lake Outburst Flood (GLOF) 22 led to dam failure and complete drainage of Lake Jircacocha downstream, and to major destruction and 23 thousands of fatalities in the city of Huaráz at a distance of 23 km. We chose an integrated approach to 24 revisit the 1941 event in terms of topographic reconstruction and numerical back-calculation with the 25 GIS-based open source mass flow/process chain simulation framework r.avaflow, which builds on an 26 enhanced version of the Pudasaini (2012) two-phase flow model. Thereby we consider four scenarios: 27 (A) and (AX) breach of the moraine dam of Lake Palcacocha due to retrogressive erosion, assuming two 28 different fluid characteristics; (B) failure of the moraine dam caused by the impact of a landslide onto the 29 lake; and (C) geomechanical failure and collapse of the moraine dam. The simulations largely yield em30 pirically adequate results with physically plausible parameters, taking the documentation of the 1941 31 event and previous calculations of future scenarios as reference. Most simulation scenarios indicate trav32 el times between 36 and 70 minutes to reach Huaráz, accompanied with peak discharges above 10,000 33 m3/s. The results of the scenarios indicate that the most likely initiation mechanism would be retrogres34 sive erosion, possibly triggered by a minor impact wave and/or facilitated by a weak stability condition 35

parameters, but also to the material involved (release mass, entrainment). A larger number of backcalculated process chains will be necessary to derive guiding parameter sets which could facilitate predictive simulations, and so will an appropriate consideration of model uncertainties and possible threshold effects (Mergili et al., 2018b). Earlier studies, considering the 2010 event at Laguna 513 (Schneider et al., 2014) and three future scenarios for Lake Palcacocha (Somos-Valenzuela et al., 2016) have followed a different strategy, using model cascades instead on integrated simulations, so that a comparison with studies based on r.avaflow is only possible to a limited extent.
Few minor comments: 1. Line 42-45-I will suggest to include the latest literature here. Several GLOF impact modeling studies have been carried out in the Himalaya recently .
Thank you very much for this remark -the following references have been included in the revised manuscript: 2. The abstract is too general and does not reflect the specific quantitative results. Text in the abstract (line 23-24) can be shortened and instead information about the results can be included.
We have shortened the general part of the abstract (L23-26 in the original manuscript) and included some fundamental information about the results (L28-29): Most simulation scenarios indicate travel times between 36 and 70 minutes to reach Huaráz, accompanied with peak discharges above 10,000 m³/s. and also about the implications (L32-34): Predictive simulations of possible future events have to be based on a larger set of back-calculated GLOF process chains, taking into account the expected parameter uncertainties and appropriate strategies to deal with critical threshold effects.
Figures: Figure 1-The number of lat/long labels can be reduced We have increased the interval between the tick marks and labels from two to four minutes.
Figure 2 (f)-The impact area ends very abruptly. This is surprising. The inundation zone can be rechecked.
The reason for this abrupt ending is that (i) the valley is bounded by a steep slope at this side, and (ii) part of the lowermost portion of the inundation zone is hidden behind the hillslope in the left foreground of the photo. We have tried to better indicate this in the figure and also indicated it in the figure caption in the revised manuscript. Yes, the reason for this pattern is that a small part of the simulated flow has proceeded downstream the Río Santa Valley near the edge of the area of interest, instead of leaving the area of interest. This is an "edge effect" not considered a significant result of the study. Therefore, we have masked out this area in the revised Fig. 12, and also in the revised Fig. 9, where a similar effect is visible. minutes to reach Huaráz, accompanied with peak discharges above 10,000 m³/s. The results of the scenarios indicate 29 that the most likely initiation mechanism would be retrogressive erosion, possibly triggered by a minor impact wave 30 and/or facilitated by a weak stability condition of the moraine dam. However, the involvement of Lake Jircacocha 31 disguises part of the signal of process initiation farther downstream. Predictive simulations of possible future events 32 1 Introduction 37 Glacial retreat in high-mountain areas often leads, after some lag time (Harrison et al., 2018), to the formation of pro-38 glacial lakes, which are impounded by moraine dams or bedrock swells. Such lakes may drain suddenly, releasing a 39 large amount of water which may result in complex and potentially catastrophic process chains downstream. Glacial 40 lakes and outburst floods (GLOFs) have been subject of numerous studies covering many mountain regions all around 41 the globe (Hewitt, 1982;Haeberli, 1983 The Cordillera Blanca (Perú) represents the most glacierized mountain chain of the Tropics. Glacial lakes and GLOFs 45 are particularly common there (Carey, 2005). 882 high-mountain lakes were identified by Emmer et al. (2016). Some 46 of these lakes are susceptible to GLOFs (Vilímek et al., 2005;Emmer andVilímek, 2013, 2014;ANA, 2014;Iturrizaga, 47 Page 4 lake, named Lake Palcacocha. A photograph taken by Hans Kinzl in 1939 (Kinzl andSchneider, 1950) indicates a lake 117 level of 4,610 m a.s.l., allowing surficial outflow (Fig. 2a). Using this photograph, Vilímek et al. (2005) estimated a lake 118 volume between 9 and 11 million m³ at that time, whereas an unpublished estimate of the Autoridad Nacional del 119 Agua (ANA) arrived at approx. 13.1 million m³. It is assumed that the situation was essentially the same at the time of 120 of fine lake sediments (Fig. 2b). Table 1 summarizes the major characteristics of Lake Palcacocha and Lake Jircacocha 129 before the 1941 GLOF. The impact area of the 1941 multi-GLOF and the condition of Lake Palcacocha after the event are well documented 146 through aerial imagery acquired in 1948 (Fig. 3). The image of Hans Kinzl acquired in 1939 (Fig. 2a) is the only record 147 of the status before the event. Additional information is available through eyewitness reports (Wegner, 2014). Howev-148 er, as Lake Palcacocha is located in a remote, uninhabited area, no direct estimates of travel times or associated flow 149 velocities are available. Also the trigger of the sudden drainage of Lake Palcacocha remains unclear. Two mechanisms 150 appear most likely: (i) retrogressive erosion, possibly triggered by an impact wave related to calving or an ice ava-151 lanche, resulting in overtopping of the dam (however, Vilímek et al., 2005 state that there are no indicators for such 152 an impact); or (ii) internal erosion of the dam through piping, leading to the failure. 153

Lake evolution since 1941 154
As shown on the aerial images from 1948, Lake Palcacocha was drastically reduced to a small remnant proglacial pond, 155 impounded by a basal moraine ridge within the former lake area, at a water level of 4563 m a.s.l., 47 m lower than before the 1941 event (Fig. 3a). However, glacial retreat during the following decades led to an increase of the lake 157 area and volume (Vilímek et al., 2005). After reinforcement of the dam and the construction of an artificial drainage in 158 the early 1970s, a lake volume of 514,800 m³ was derived from bathymetric measurements (Ojeda, 1974). In 1974, two 159 artificial dams and a permanent drainage channel were installed, stabilizing the lake level with a freeboard of 7 m to 160 the dam crest (Portocarrero, 2014). By 2003, the volume had increased to 3.69 million m³ (Zapata et al., 2003). In the 161 same year, a landslide from the left lateral moraine caused a minor flood wave in the Cojup Valley (Fig. 2d). In 2016, 162 the lake volume had increased to 17.40 million m³ due to continued deglaciation (ANA, 2016). The potential of fur-163 ther growth is limited since, as of 2017, Lake Palcacocha is only connected to a small regenerating glacier. Further, the 164 lake level is lowered artificially, using a set of siphons (it decreased by 3 m between December 2016 and July 2017). 165 Table 1 summarizes the major characteristics of Lake Palcacocha in 2016. The overall situation in July 2017 is illustrat-166 ed in Fig. 2c. 167

Previous simulations of possible future GLOF process chains 168
Due to its history, recent growth, and catchment characteristics, Lake Palcacocha is considered hazardous for the 169 downstream communities, including the city of Huaráz ( (1) 211 qE,s and qE,f (m s -1 ) are the solid and fluid entrainment rates, Ts and Tf (J) are the solid and fluid kinetic energies, and αs,E As an exact positioning of the glacier terminus is not possible purely based on the photo, the position is opti-236 mized towards a lake volume of approx. 13 million m³, following the estimate of ANA. It is further assumed 237 that there was surficial drainage of Lake Palcacocha as suggested by Fig. 2a, i.e. the lowest part of the moraine 238 crest is set equal to the former lake level of 4,610 m a.s.l (Fig. 4b). 239 2. Also for Lake Jircacocha, surficial overflow is assumed (a situation that is observed for most of the recent land-240 slide-dammed lakes in the Cordillera Blanca). On this basis the landslide dam before its breach is reconstruct-241 ed, guided by topographic and geometric considerations. The lowest point of the dam crest is set to 242 4,130 m a.s.l. (Fig. 4c). 243 3. Erosional features along the flow channel are assumed to largely relate to the 1941 event. These features are 244 filled accordingly (see Table 2 for the filled volumes). In particular, the flow channel in the lower part of the 245 valley, reportedly deepened by up to 50 m in the 1941 event (Vilímek et al., 2005), was filled in order to repre-246 sent the situation before the event in a plausible way (Fig. 4d). 247 All lakes are considered as fluid release volumes in r.avaflow. The initial level of Lake Palcacocha in 1941 is set to 248 4,610 m a.s.l., whereas the level of Lake Jircacocha is set to 4,129 m a.s.l. The frontal part of the moraine dam im-249 pounding Lake Palcacocha and the landslide dam impounding Lake Jircacocha are considered as entrainable volumes. 250 Further, those areas filled up along the flow path ( Fig. 4d) are considered entrainable, mainly following Vilímek et al. 251 (2005). However, as it is assumed that part of the material was removed through secondary processes or afterwards, 252 only 75% of the added material are allowed to be entrained. All entrained material is considered 80% solid and 20% 253 fluid per volume. 254 The reconstructed lake, breach, and entrainable volumes are shown in Tables 1 and 2. The glacier terminus in 1941 255 was located in an area where the lake depth increases by several tens of metres, so that small misestimates in the posi-256 tion of the glacier tongue may result in large misestimates of the volume, so that some uncertainty has to be accepted. Retrogressive erosion, possibly induced by minor or moderate overtopping. This scenario is related to a pos-262 sible minor impact wave, caused for example by calving of ice from the glacier front, an increased lake level 263 due to meteorological reasons, or a combination of these factors. In the simulation, the process chain is start-264 ed by cutting an initial breach into the dam in order to initiate overtopping and erosion. The fluid phase is 265 considered as pure water. 266 AX Similar to Scenario A, but with the second phase considered a mixture of fine mud and water. For this pur-267 pose, density is increased to 1,100 instead of 1,000 kg m -3 , and a yield strength of 5 Pa is introduced (Dom-268 nik et al., 2013; Pudasaini and Mergili, 2019; Table 3). For simplicity, we still refer to this mixture as a fluid. 269 Such changed phase characteristics may be related to the input of fine sediment into the lake water (e.g. 270 caused by a landslide from the lateral moraine as triggering event), but are mainly considered here in order 271 to highlight the effects of uncertainties in the definition and parameterization of the two-phase mixture flow. 272

B
Retrogressive erosion, induced by violent overtopping. This scenario is related to a large impact wave caused 273 by a major rock/ice avalanche or ice avalanche rushing into the lake. In the simulation, the process chain is 274 initiated through a hypothetic landslide of 3 million m³ of 75% solid and 25% fluid material, following the Internal erosion-induced failure of the moraine dam. Here, the process chain is induced by the collapse of 278 the entire reconstructed breach volume (Fig. 4b). In the simulation, this is done by considering this part of 279 the moraine not as entrainable volume, but as release volume (80% solid, 20% fluid, whereby fluid is again 280 considered as pure water). 281 Failure of the dam of Lake Jircacocha is assumed having occurred through overtopping and retrogressive erosion, in-282 duced by the increased lake level and a minor impact wave from the flood upstream. No further assumptions of the 283 initial conditions are required in this case. 284 The model parameter values are selected in accordance with experiences gained from previous simulations with 285 r.avaflow for other study areas, and are summarized in Table 3. Three parameters mainly characterizing the flow fric-286 tion (basal friction of solid δ, ambient drag coefficient CAD, and fluid friction coefficient CFF) and the entrainment coef- ing those four parameters while keeping the others constant helps us to capture variability while minimizing the de-291 grees of freedom, remaining aware of possible equifinality issues (Beven, 1996;Beven and Freer, 2001). 292 A particularly uncertain parameter is the empirical entrainment coefficient CE (Eq. 1). In order to optimize CE, we 293 consider (i) successful prediction of the reconstructed breach volumes; and (ii) correspondence of peak discharge with 294 published empirical equations on the relation between peak discharge, and lake volume and dam height (Walder and 295 O'Connor, 1997). Table 4 summarizes these equations for moraine dams (applied to Lake Palcacocha) and landslide 296 dams (applied to Lake Jircacocha), and the values obtained for the regression and the envelope, using the volumes of 297 both lakes. We note that Table 4 reveals very large differences -roughly one order of magnitude -between regression 298 and envelope. In case of the breach of the moraine dam of Lake Palcacocha, we consider an extreme event due to the 299 steep, poorly consolidated, and maybe soaked moraine, with a peak discharge close to the envelope (approx.. 300 15,000 m 3 s -1 ). For Lake Jircacocha, in contrast, the envelope values of peak discharge do not appear realistic. However, 301 due to the high rate of water inflow from above, a value well above the regression line still appears plausible, even 302 though the usefulness of the empirical laws for this type of lake drainage can be questioned. The value of CE optimized 303 for the dam of Lake Jircacocha is also used for entrainment along the flow path. 304 All of the computational experiments are run with 10 m spatial resolution. Only flow heights ≥25 cm are considered 305 for visualization and evaluation. We now describe one representative simulation result for each of the considered sce-306 narios, thereby spanning the most plausible and empirically adequate field of simulations. 307 5 r.avaflow simulation results 308

Scenario A -Event induced by overtopping; fluid without yield strength 309
Outflow from Lake Palcacocha starts immediately, leading to (1) lowering of the lake level and (2) retrogressive ero-310 sion of the moraine dam. The bell-shaped fluid discharge curve at the hydrograph profile O1 (Fig. 4) reaches its peak 311 of 18,700 m³ s -1 after approx. 780 s, and then decreases to a small residual (Fig. 5a). Channel incision happens quickly -312 53 m of lowering of the terrain at the reference point R1 occurs in the first less than 1200 s, whereas the lowering at 313 the end of the simulation is 60 m (Fig. 6a). This number represents an underestimation, compared to the reference Page 9 value of 76 m ( Table 2). The lake level decreases by 42 m, whereby 36.5 m of the decrease occur within the first 315 1200 s. The slight underestimation, compared to the reference value of 47 m of lake level decrease, is most likely a 316 consequence of uncertainties in the topographic reconstruction. A total amount of 1.5 million m³ is eroded from the 317 moraine dam of Lake Palcacocha, corresponding to an underestimation of 22%, compared to the reconstructed breach 318 volume. Underestimations mainly occur at both sides of the lateral parts of the eroded channel near the moraine crest 319 -an area where additional post-event erosion can be expected, so that the patterns and degree of underestimation 320 appear plausible (Fig. 7a). In contrast, some overestimation of erosion occurs in the inner part of the dam. For numeri-321 cal reasons, some minor erosion is also simulated away from the eroded channel. The iterative optimization procedure 322 results in an entrainment coefficient CE = 10 -6.75 .

323
The deposit of much of the solid material eroded from the moraine dam directly downstream from Lake Palcacocha, as 324 observed in the field (Fig. 2c), is reasonably well reproduced by this simulation, so that the flow proceeding down-325 valley is dominated by the fluid phase (Fig. 8). It reaches Lake Jircacocha after t = 840 s (Fig. 5b). As the inflow occurs 326 smoothly, there is no impact wave in the strict sense, but it is rather the steadily rising water level (see Fig. 6b for the 327 evolution of the water level at the reference point R2) inducing overtopping and erosion of the dam. This only starts 328 gradually after some lag time, at approx. t = 1,200 s. The discharge curve at the profile O2 (Fig. 4) reaches its pro-329 nounced peak of 750 m³ s -1 solid and 14,700 m³ s -1 fluid material at t = 2,340 s, and then tails off slowly.

330
In the case of Lake Jircacocha, the simulated breach is clearly shifted south, compared to the observed breach. With 331 the optimized value of the entrainment coefficient CE = 10 -7.15 , the breach volume is underestimated by 24%, compared 332 to the reconstruction (Fig. 7b). Also here, this intentionally introduced discrepancy accounts for some post-event ero-333 sion. However, we note that volumes are uncertain as the reconstruction of the dam of Lake Jircacocha -in contrast to 334 Lake Palcacocha -is a rough estimation due to lacking reference data. 335 Due to erosion of the dam of Lake Jircacocha, and also erosion of the valley bottom and slopes, the solid fraction of the 336 flow increases considerably downstream. Much of the solid material, however, is deposited in the lateral parts of the 337 flow channel, so that the flow arriving at Huaráz is fluid-dominated again (Fig. 8). The front enters the alluvial fan of 338 Huaráz at t = 2,760 s, whereas the broad peak of 10,500 m³ s -1 of fluid and 2,000 m³ s -1 of solid material (solid fraction 339 of 16%) is reached in the period between 3,600 and 3,780 s ( Fig. 4; Fig. 5c). Discharge decreases steadily afterwards. A 340 total of 2.5 million m³ of solid and 14.0 million m³ of fluid material pass the hydrograph profile O3 until t = 5,400 s. 341 Referring only to the solid, this is less material than reported by Kaser and Georges (2003). However, (i) there is still 342 some material coming after, and (ii) pore volume has to be added to the solid volume, so that the order of magnitude 343 of material delivered to Huaráz corresponds to the documentation in a better way. Still, the solid ratio of the hydro-344 graph might represent an underestimation. 345 As prescribed by the parameter optimization, the volumes entrained along the channel are in the same order of mag-346 nitude, but lower than the reconstructed volumes summarized in Table 2 Table 5. 352

Scenario AX -Event induced by overtopping; fluid with yield strength 353
Adding a yield strength of τy = 5 Pa to the characteristics of the fluid substantially changes the temporal rather than 354 the spatial evolution of the process cascade. As the fluid now behaves as fine mud instead of water and is more re-sistant to motion, velocities are lower, travel times are much longer, and the entrained volumes are smaller than in the 356 Scenario A (Fig. 9b; Table 5). The peak discharge at the outlet of Lake Palcacocha is reached at t = 1,800 s. Fluid peak 357 discharge of 8,200 m³ s -1 is less than half the value yielded in Scenario A (Fig. 5d). The volume of material eroded from 358 the dam is only slightly smaller than in Scenario A (1.4 versus 1.5 million m³). The numerically induced false positives 359 with regard to erosion observed in Scenario A are not observed in Scenario AX, as the resistance to oscillations in the 360 lake is higher with the added yield strength (Fig. 7c). Still, the major patterns of erosion and entrainment are the same. 361 Interestingly, erosion is deeper in Scenario AX, reaching 76 m at the end of the simulation (Fig. 6c) and therefore the 362 base of the entrainable material (Table 2). This is most likely a consequence of the spatially more concentrated flow 363 and therefore higher erosion rates along the centre of the breach channel, with less lateral spreading than in Scenar-364 io A. 365 Consequently, also Lake Jircacocha is reached later than in Scenario A (Fig. 6d), and the peak discharge at its outlet is 366 delayed (t = 4,320 s) and lower (7,600 m³ s -1 of fluid and 320 m³ s -1 of solid material) (Fig. 5e). 2.0 million m³ of materi-367 al are entrained from the dam of Lake Jircacocha, with similar spatial patterns as in Scenario A (Fig. 7d). Huaráz is 368 reached after t = 4,200 s, and the peak discharge of 5,000 m³ s -1 of fluid and 640 m³ s -1 of solid material at O3 occurs 369 after t = 6,480 s (Fig. 5f). This corresponds to a solid ratio of 11%. Interpretation of the solid ratio requires care here as 370 the fluid is defined as fine mud, so that the water content is much lower than the remaining 89%. The volumes en-371 trained along the flow channel are similar in magnitude to those obtained in the simulation of Scenario A (Table 5). 372

Scenario B -Event induced by impact wave 373
Scenario B is based on the assumption of an impact wave from a 3 million m³ landslide. However, due to the relatively 374 gently-sloped glacier tongue heading into Lake Palcacocha at the time of the 1941 event (Figs. 2a and 4b), only a small 375 fraction of the initial landslide volume reaches the lake, and impact velocities and energies are reduced, compared to a 376 direct impact from the steep slope. Approx. 1 million m³ of the landslide have entered the lake until t = 120 s, an 377 amount which only slightly increases thereafter. Most of the landslide deposits on the glacier surface. Caused by the 378 impact wave, discharge at the outlet of Lake Palcacocha (O1) sets on at t = 95 s and, due to overtopping of the impact 379 wave, immediately reaches a relatively moderate first peak of 7,000 m³ s -1 of fluid discharge. The main peak of 380 16,900 m³ s -1 of fluid and 2,000 m³ s -1 of solid discharge occurs at t = 1,200 s due do the erosion of the breach channel.

381
Afterwards, discharge decreases relatively quickly to a low base level (Fig. 10a). The optimized value of CE = 10 -6.75 is 382 used also for this scenario. The depth of erosion along the main path of the breach channel is clearly less than in the 383 Scenario A (Fig. 6e). However, Table 5 shows a higher volume of eroded dam material than the other scenarios. These 384 two contradicting patterns are explained by Fig. 11a: the overtopping due to the impact wave does not only initiate 385 erosion of the main breach, but also of a secondary breach farther north. Consequently, discharge is split among the 386 two breaches and therefore less concentrated, explaining the lower erosion at the main channel despite a larger total 387 amount of eroded material. The secondary drainage channel can also be deduced from observations (Fig. 3a), but has 388 probably played a less important role than suggested by this simulation. 389 The downstream results of Scenario B largely correspond to the results of the Scenario A, with some delay partly relat-390 ed to the time from the initial landslide to the overtopping of the impact wave. Discharge at the outlet of Lake Jircaco-391 cha peaks at t = 2,700 s (Fig. 10b), and the alluvial fan of Huaráz is reached after 3,060 s (Fig. 10c). The peak discharges 392 at O2 and O3 are similar to those obtained in the Scenario A. The erosion patterns at the dam of Lake Jircacocha 393 (again, CE = 10 -7.15 ) very much resemble those yielded with the scenarios A and AX (Fig. 11b), and so does the volume 394 of entrained dam material (2.2 million m³). The same is true for the 2.5 million m³ of solid and 13.9 million m³ of fluid 395 material entering the area of Huaráz until t = 5,400 s, according to this simulation.

Page 11
Also in this scenario, the volumes entrained along the flow channel are very similar to those obtained in the simula-397 tion of Scenario A. The travel times and frontal velocities -resembling the patterns obtained in Scenario A, with the 398 exception of the delay -are shown in Fig. 12a, whereas Table 5 summarizes the key numbers in terms of times, vol-399 umes, and discharges. 400

Scenario C -Event induced by dam collapse 401
In Scenario C, we assume that the breached part of the moraine dam collapses, the collapsed mass mixes with the wa-402 ter from the suddenly draining lake, and flows downstream. The more sudden and powerful release, compared to the 403 two other scenarios, leads to higher frontal velocities and shorter travel times ( Fig. 12b; Table 5). 404 In contrast to the other scenarios, impact downstream starts earlier, as more material is released at once, instead of 405 steadily increasing retrogressive erosion and lowering of the lake level. The fluid discharge at O1 peaks at almost 406 40,000 m³ s -1 (Fig. 10d) rapidly after release. Consequently, Lake Jircacocha is reached already after 720 s, and the im-407 pact wave in the lake evolves more quickly than in all the other scenarios considered (Fig. 6f). The lake drains with a 408 peak discharge of 15,400 m³ s -1 of fluid and 830 m³ s -1 of solid material after 1,680-1,740 s (Fig. 10e). In contrast to the 409 more rapid evolution of the process chain, discharge magnitudes are largely comparable to those obtained with the 410 other scenarios. The same is true for the hydrograph profile O3: the flow reaches the alluvial fan of Huaráz after 411 t = 2,160 s, with a peak discharge slightly exceeding 10,000 m³ s -1 of fluid and 2,000 m³ s -1 of solid material between Page 12 the slopes of Palcaraju or Pucaranra could have been partly alleviated on the rather gently sloped glacier tongue be-436 tween the likely release area and Lake Palcacocha. 437 The minor erosional feature north of the main breach was already visible in the photo of Kinzl (Fig. 2a), possibly indi-438 cating an earlier, small GLOF. It remains unclear whether it was reactivated in 1941. Such a reactivation could only be 439 directly explained by an impact wave, but not by retrogressive erosion only (A/AX) or internal failure of the dam (C) -440 so, more research is needed here. The source area of a possible impacting landslide could have been the slopes of Pal-441 caraju or Pucaranra (Fig. 1), or the calving glacier front (Fig. 2a). Attempts to quantify the most likely release volume 442 and material composition would be considered speculative due to the remaining difficulties in adequately simulating that it is hard to decide about the more adequate assumption. Even though the strategy of using the results of earlier simulations as reference may increase the robustness of model results, it might also reproduce errors and inaccuracies 477 of earlier simulation attempts, and thereby confirm wrong results. 478 The large amount of more or less pure lake water would point towards the Scenario A, whereas intense mixing and 479 entrainment of fine material would favour the Scenario AX. More work is necessary in this direction, also considering 480 possible phase transformations (Pudasaini and Krautblatter, 2014). At the same time, the optimization and evaluation 481 of the simulated discharges remains a challenge. Here we rely on empirical relationships gained from the analysis of 482 comparable events (Walder and O'Connor, 1997). 483

Implications for predictive simulations 484
Considering what was said above, the findings from the back-calculation of the 1941 event can help us to better un-485 derstand and constrain possible mechanisms of this extreme process chain. In principle, such an understanding can be 486 transferred to present hazardous situations in order to inform the design of technical remediation measures. Earlier, 487 measures were not only implemented at Lake Palcacocha (Portocarrero, 2014), but also at various other lakes such as 488 Laguna 513: a tunnelling scheme implemented in the 1990s strongly reduced the impacts of the 2010 GLOF process 489 chain (Reynolds, 1998;Reynolds et al., 1998;Schneider et al., 2014).  Lowering of lake level (m) 47 2) -33  1) The fluid material density is set to 1,100 kg m -3 in Scenario AX. 809 2) The yield strength of the fluid phase is set to 5 Pa in Scenario AX.  Table 4. Empirical relationships for the peak discharge in case of breach of moraine and landslide dams (Walder and 814 O'Connor, 1997), and the peak discharges estimated for Lake Palcacocha and Lake Jircacocha. qp = peak discharge 815 (m 3 s -1 ), V = total volume of water passing through the breach (m³); D = drop of lake level (m); REG = regression; 816 ENV = envelope. The values of V and D for the two lakes are summarized in Table 1. See also Rivas et al. (2015).  Table 5. Summary of the key results obtained with the computational experiments A-C. Refer to Tables 1 and 2 for the  819 volumes involved, and to  and Lake Jircacocha (reference point R2 in Fig. 4c). The reference points are placed in a way to best represent the evo-855 lution of the breach in the dam for Lake Palcacocha, and the evolution of the impact wave for Lake Jircacocha. Addi-856 tionally, the evolution of the lake level is shown for Lake Palcacocha. Note that the result for Scenario B is only dis-857 played for Lake Palcacocha (e), whereas the result for Scenario C is only illustrated for Lake Jircacocha (f). The vertical 858 distance displayed on the y axis refers to the terrain height or the lake level at the start of the simulation, respectively, 859 whereby the flow height is imposed onto the topography. In Scenario B, the initial impact wave at the dam of Lake 860 Palcacocha is only poorly represented due to the low temporal resolution of the simulation, and due to blurring by 861 numerical effects (e). Note that the legend of (a) also applies to 887 (b). Void fields in the profile graph refer to areas without clearly defined flow front. 888