How the debris flow was reconstructed

Use a physical model to examine which simulated flow paths, volumes, and mechanisms of the 8 October 2026 debris flow in San Carlos de Apoquindo are compatible with public information, by comparing hypotheses about its origin with observations. Observed results, calculated results (model outputs under assumptions), and hypothetical scenarios are distinguished. This is neither an expert assessment nor an official report and does not attribute responsibility.

• 30 m terrain resampled to a 4 m grid (resampling does not increase effective detail), with all 42,018 buildings represented as blocks; streets lowered by 0.15 m.
• Three channels that the 30 m terrain does not resolve, carved based on mapping: the open channel in the linear park up to the Francisco Bulnes Correa culvert inlet; Quebrada Honda downstream of the inlet, according to the official hydrographic network; and the southern branch up to the upper part of Av. Monseñor Álvaro del Portillo, also according to the official network.
• Domain from the foothills to Paul Harris, Hospital Dipreca, and the Atenas roundabout; mesh of about 852,000 triangles (30 m² in the corridor). Local microsimulations with 8 m² triangles for the U. de los Andes campus.
• Variants after the held-out test: buildings in the official area that admit and retain mud (floor 0.3 m above ground and high resistance); campus and La Plaza Sur features carved according to contributed observations (lower carriageway, passage from the clearing to the courtyard, hillside road, parking area); and the main ravine carved along the line of trees, with and without a 3 m accumulation of logs (arc and art).

• ANUGA 4.0.1 (2D shallow water, finite volumes, DE1 scheme, open source) with the quadratic rheology of O'Brien and Julien: yield stress, viscosity, and turbulence, applied implicitly in each cell.
• Custom operators: mixture source with mass and momentum according to the hydrograph; rainfall on the hillside and lower hillside with exact volume, constant or following a measured hyetograph; storm-water culvert with maximum capacity and capture radius.
• Outflow-only boundaries, to avoid creating mass. In the city runs, the volume-balance residual is below 0.01 %, except in the 12 and 13 October batch with shifted hydrographs (p34, d39, f35, f38, arc, and art): there, correction of negative cells created between 2.1 and 3.1 % mass in the first few minutes. This was corrected with a maximum time step of 0.5 s; the corrected re-runs (p34b, f35b, and f38b) have a residual of 0.00 % and confirm the results. In the campus microsimulations, the residual reaches +3.2 %.

Mud floods and mudflows contain 20–45 % sediment by volume (1,330–1,740 kg/m³); debris flows, 50–80 % (1,400–2,400 kg/m³) (O'Brien y Julien, 1988; Costa, 1984; Iverson, 1997). According to press reports (compiled in Wikipedia), the acting director of Sernageomin described the event as a rapid surge of water over a very short time carrying a large amount of sediment (flash flood), and the images show a fine matrix with stones, logs, and cars. The reference mud is Glenwood 2b with Cv 0.40 (1,660 kg/m³); tests include Cv 0.43 (1,710 kg/m³, to account for the coarse load), Glenwood 1 (Cv 0.45, thicker), and Aspen 2 (more fluid). Subsequent trials use different muds by zone (Cv 0.27 or turbid water with Cv 0.10 on the campus lower hillside and Cv 0.40 in the main branch) and concentration quadrature between 0.20 and 0.50 (Beta(3,2) distribution, five nodes): the expected area with more than 5 cm is 169 ha and the index remains between 71 and 73.

• Contributing catchment of 10.1 km², delineated in UTM coordinates.
• Observed: about 100 mm in the foothills on 6 and 7 October according to CR2; the mayor spoke of 144 mm accumulated, with saturated soil, without an identified station; the isotherm was reported near 2,800 m, above the catchment summit (2,355 m), supporting the assumption of rainfall throughout it; no snow was detected in the cloud-free part of the catchment in the 21 September Sentinel-2 image. Calculated: the upper bound on the snowmelt contribution is less than 5 %.
• Calculated: independent estimate adopting the reported saturation, median of 295,000 m³ (5th and 95th percentiles: 181,000–468,000, under the adopted assumptions). In the scenarios tested, about 300,000 m³ allows the reach criterion to Padre Hurtado to be reproduced.
• Fire (calculated): with measured severity (Sentinel-2 dNBR), 82 % of the contributing catchments shows a burn signal and only 10 % has moderate-high or high severity. With this severity, the empirical volume from Gartner (2014) falls from 374,000 to 148,000 m³ (median). In the balance, without fire the pulse would have had about 62 % of the mud volume (51–76 %) and 43 % of the solids; the coefficients by severity class are assumed. With the M1 model of Staley et al. (2017), calibrated in the western United States and used for illustration, the debris-flow probability is 0.28 with fire and 0.07 without it.

• H1: all mud exits through the main channel onto the alluvial fan. Necessary, but not sufficient: it explains the northern branch, not the southern sector.
• H2: overflow only at the Francisco Bulnes Correa culvert inlet. Rejected as the sole origin; the inlet is compatible as a transit point for flow arriving from the alluvial fan.
• H3: distributed inflow: main channel, southern branch, and short ravines toward La Plaza Sur.
• H4 and H5: combined outflow at the alluvial fan and inlet, and direct rainfall on the hillside; without that rainfall, Clínica U. de los Andes is no longer reached and the index falls from 72 to 68.
• H3 + rainfall (selected for the held-out test, z02): H3 plus 35 mm/h of rainfall on the burned hillside; 72/100; unfavorable test result.
• Subsequent trials (fitting): porous houses, turbid water on the lower hillside, campus features, integrated hypothesis (79–82/100 depending on rainfall), log jam, porosity only in small houses, volume without fire, and bike park.
• Variants: culvert capacity (10 to 200 m³/s, hypothetical scenarios), southern-branch alignment and contribution, mud type, and roughness.

Four criteria set before simulation, each between 0 and 1: streets with observed mud that the model wets; facilities with confirmed impacts reached (UC Christus, Duoc UC, Clínica and campus of U. de los Andes); reach to Padre Hurtado (1 at less than 200 m); and the fraction of simulated mud in the urban area that falls within the official affected area from IDE Chile. Each criterion has an acceptance threshold (original criteria); the hypothesis selected for the held-out test meets three of four, and the integrated hypothesis meets four of four with assumed rainfall and three of four with measured rainfall. Separately, the exploratory compatibility index (ECI), defined later, summarizes these components on a scale from 0 to 100. The selected hypothesis scored 74/100 with the press perimeter (until 10 October) and scores 72/100 with the official polygon. Under a rule recorded before choosing the candidate (validation protocol, amendment 1), if the facilities component depends only on the campus control point and that point is not wetted with a fine time step, it is calculated without it. The index rewards the fraction of mud within the official area; this is why it does not fall when there is less mud (for example, without fire) and cannot be used to assess the volume. It does not measure predictive accuracy or constitute independent validation, which remains pending, and differences of two or three points should not be interpreted as conclusive evidence that one hypothesis is superior to another.

• Stopping of a deposit on an inclined plane: resting thickness of 0.120 m versus the theoretical 0.1199 m.
• Convergence: on a plane, first order with the time step; in the city, reducing the time step by a factor of four changes maximum depth by 2 mm (median) and 9 mm (p95), and the selected hypothesis gives the same indicators with a fine time step (z02) as with the normal time step (v02); porous houses (z26), the more fluid lower hillside (z29), and the integrated hypothesis (z32) were also verified with a fine time step.
• One campus control point proved sensitive to the time step: it exceeds 5 cm with the normal time step (0.06–0.15 m) and remains at 0.01 m with a fine time step in z02 and z26.
• Volume balance with a residual below 0.01 % in city simulations, except for the 12 and 13 October batch (artifact of +2.1 to +3.1 %, corrected; the re-runs p34b, f35b and f38b give 0.00 %; see Physical model). Campus microsimulations: up to +3.2 %.

Frozen versions: each published PDF is preserved as published, with its SHA-256 hash (previous versions (Spanish original)); terrain datasets are not modified after use. Experiment log: all runs are recorded, including those stopped, canceled, failed, and discarded, with the reason. Separation of observations: the fitting observations (streets, facilities, reach, perimeter, thickness at UC Christus) were used to formulate and select hypotheses, so the index measures compatibility, not validation; official properties and street centerlines provide a descriptive comparison; and the held-out observations, for testing rather than fitting, are 49 officially affected facilities, post-event Sentinel-1 images, and future high-water marks. The facilities test was defined before examining results and was run only once on 11 October: the selected hypothesis (z02) reaches 19 of 49 points (39 %), below the set threshold of 70 % and without outperforming the main-channel-only reference. Unfavorable result, reported without adjusting the model (protocol). Subsequent trials (porous houses, turbid water on the lower hillside, campus and La Plaza Sur features contributed using Google Earth, Street View, and a video) improve the verified index to 76–78/100, and the integrated hypothesis combining them obtains 79–82/100 depending on rainfall (82 verified with a fine time step using an assumed 35 mm/h; 81 with measured rainfall; 79 with measured rainfall spread evenly). These are fitting tests: a new test will require new held-out observations, specified before comparing them with model results.

Automated review of physics, formatting, and wording using AI tools (Codex), documented by the author, with reports published below; this does not constitute independent scientific review. Each external source consultation is logged with date, source, data used, and the SHA-256 hash of the capture; the full searchable log is in Data.

• Only qualitative observations are available: there are no published discharge measurements or high-water marks, and the official area is a management perimeter, not a verified map of mud coverage.
• The 30 m terrain does not resolve the features that carry water to the U. de los Andes campus. In microsimulations, when these are carved according to contributed observations, the most compatible mechanism is turbid hillside runoff channeled by La Plaza Sur, entering the courtyard through a clearing; overflow from Álvaro del Portillo does not reach the courtyard. These values are underestimated on the city mesh.
• Buildings are impermeable blocks, except in the porous-house variant, which represents mud entry and retention in a simplified way. The model is single-phase: it distinguishes turbid water on the lower hillside from mud in the main branch only by zone, without representing changes in concentration along the flow path. Simulated mud extends beyond the official area.
• Erosion and deposition along the flow path, entrained logs and vehicles, urban drainage, and infiltration within the city are not simulated.
• No configuration simultaneously reproduces the meter reported at Clínica UC Christus and the extent within the official area.
• Rainfall within the catchment was not measured: the Río Molina station, about 7 km away, is used.
• FABDEM has a noncommercial license, and the viewer's satellite imagery is for visualization only (see Terms).

• 2026-10-13: account of the events with each fact attributed to its source and the section “The hypotheses, one by one” (in preparation for the paper); fire severity using Sentinel-2 dNBR and volume with and without fire; Río Molina station hyetograph (81/100 with the integrated hypothesis); volume without fire in the city (f35 and f38es); log jam (arc and art: does not divert flow southward); porosity only in houses of 250 m² or less (d39: does not recover the UC Christus thickness); bike park (mrbc and mrbt); mass artifact with shifted hydrographs, corrected with a maximum time step of 0.5 s; correction: rainfall at Tobalaba on 6 and 7 October totals 131.2 mm in 6 h reports (49.2 mm was the 6 h maximum, not the total).
• 2026-10-12: trials after the held-out test: porous houses (76/100 verified), more fluid lower hillside (78/100 verified), concentration quadrature, campus and La Plaza Sur microsimulations with contributed observations (video, Street View, Google Earth), and integrated hypothesis (82/100, verified with a fine time step); integrated hypothesis with measured rainfall spread over 4 h (79/100) and campus sensitivity to intensity (15, 25, and 35 mm/h); English version of the site and documents.
• 2026-10-11: official affected area from IDE Chile used as the evaluation reference (ECI ≈72/100); hypothetical culvert scenarios (in these, 30 m³/s is insufficient and ~65 m³/s would prevent overflow); southern branch along the official channel to Álvaro del Portillo; density ranges from the literature; porous houses, Cv 0.43 mud, and increased roughness being tested; campus microsimulation; Simulations page and paper appendix with scenario summaries.
• 2026-10-10: full city (all 42,018 buildings) and expanded domain; convergence in the city; volume with saturated soil, isotherm, and Sentinel-2; comparison with press reports; exploratory comparison of hypotheses using the ECI; official hydrographic network; corrected heights in the viewer.
• 2026-10-09: data extraction, terrain, ANUGA model with rheology, calibration metrics; reviews A1 and A2 applied (source with momentum, equilibrium, convergence, catchments in UTM, volume prior, GLUE-type calibration); initial screening and carved linear park channel.

• formato_F1_informe.md
• formato_F1_prompt.md
• ronda_A1_informe.md
• ronda_A1_prompt.md
• ronda_A2_informe.md
• ronda_A2_prompt.md
• redaccion_R1_informe.md (wording, 11-Oct)