# Cluster results

## 2026-11-16 (second round): new priors and errors, four glaciers

Changes from the first round (below): thickness error 20% of the mean thickness for all
glaciers (`volume_rel_err=0.2`), prcp_fac 0.5-5 (log10), initial state offset -3 to
+2 K (centre -0.5), status `disappeared` for glaciers gone at the RGI date + 1. Else
identical. Jobs 599811 (four glaciers, 3-6 min each) and 599812 (Vernagtferner without
the initial state period). nb3 and nb4 rebuilt from these runs.

Weighted median [5%, 95%]:

| | Vernagtferner 8.6 km2 | Hintereisferner 8.0 | RGI60-11.00834 0.8 | RGI60-11.00731 0.4 |
|---|---|---|---|---|
| iterations / Neff | 7 / 165 | 12 / 157 | 8 / 192 | 20 (limit) / 155 |
| log Z | -14.06 | -14.35 | -12.36 | -11.50 |
| melt_f | 3.38 [1.80, 11.35] | 4.06 [1.86, 10.31] | 2.87 [1.78, 8.17] | 4.33 [2.40, 11.91] |
| prcp_fac (prior 0.62-4.07) | 1.23 [0.62, 3.47] | 1.24 [0.60, 3.23] | 1.32 [0.73, 3.87] | 1.24 [0.60, 2.64] |
| temp_bias | 1.86 [-1.54, 5.39] | 0.69 [-1.97, 4.42] | 2.10 [-0.92, 5.47] | 2.02 [-0.39, 4.69] |
| log10 A factor | -0.10 [-0.79, 0.66] | 0.30 [-0.43, 0.90] | 0.54 [0.07, 0.84] | 0.01 [-0.58, 0.56] |
| initial state offset (prior -2.54/1.53) | -0.63 [-2.42, 1.08] | -1.26 [-2.65, -0.05] | -1.56 [-2.68, 0.13] | -1.56 [-2.63, -0.06] |
| thickness 2012 / 2004 | 0.87 [0.79, 0.93] | 0.85 [0.77, 0.91] | 0.67 [0.54, 0.82] | 0.73 [0.59, 0.85] |
| volume 2020 / 2004 | 0.67 [0.50, 0.74] | 0.63 [0.45, 0.71] | 0.24 [0.02, 0.43] | 0.29 [0.06, 0.49] |
| runoff 2000-19 (Mt/yr) | 16.5 [12.0, 31.7] | 16.8 [12.5, 29.7] | 1.55 [1.19, 3.18] | 0.82 [0.59, 1.40] |
| Jul-Sep fraction | 0.75 [0.63, 0.88] | 0.74 [0.62, 0.83] | 0.72 [0.60, 0.84] | 0.77 [0.66, 0.87] |
| glacier melt share | 0.81 [0.58, 0.93] | 0.84 [0.63, 0.94] | 0.76 [0.51, 0.88] | 0.81 [0.62, 0.93] |

Vernagtferner without the initial state period (599812): 8 iterations, log Z -13.62
(vs -14.06), log10 A 0.14 [-0.57, 0.78] (a little constrained: stiffest ice ruled out).

- The initial state still leans cold everywhere (medians -0.6 to -1.6 K), and warm
  offsets are ruled out on three glaciers (95% near 0 K); the -3 K bound no longer cuts
  (5% quantiles -2.4 to -2.7 vs prior -2.54).
- prcp_fac is now barely constrained within 0.5-5; the runoff spread shrinks to a
  factor 2-3 (was 2-5) and still follows prcp_fac (Vernagt: rank corr 0.99, S1 0.97).
- A constrained only on RGI60-11.00834 (soft ice, 1.2-7x default).
- With 20% thickness error, the 2020 volume spread of Vernagtferner is a factor 3
  (0.14-0.42 km3), it was 10 with the 33% error.
- Runs which fail in the first iteration (Vernagt): 146 disappeared, 80 domain,
  35 rejected.
- Resampling (5 draws of 500): quantiles stable here (runoff within 0.6 Mt/yr).

## 2026-11-16 (first round): wide priors on four glaciers

Setup (notebook settings, `scripts/run_glacier.py` defaults): L3 2025.6 W5E5, border 80;
spinup 30 yr, initial state period 20 yr; priors logit-normal sd 1.4 centred mid-range:
log10 melt_f and log10 prcp_fac within OGGM's bounds (1.5-17, 0.1-10), temp_bias informed
+-10 K, log10 A factor +-1, initial state offset +-2 K; errors: 0.2 x Hugonnet MB error,
area 5% + half a grid point, mean thickness 0.5 x the relative IceBoost error; 500
particles, max 20 iterations, seed 0. Hydro rerun of the particles holding 99.9% of the
weight, runoff over the glacier area of 2000. Slurm jobs 599806-599808 (32 cpus each,
2-6 min per glacier, node09 / node17). Outputs: `example_workflows/out/wide_priors/`.

Weighted median [5%, 95%]:

| | Vernagtferner 8.6 km2 | Hintereisferner 8.0 | RGI60-11.00834 0.8 | RGI60-11.00731 0.4 |
|---|---|---|---|---|
| iterations / Neff | 7 / 175 | 9 / 157 | 10 / 159 | 20 (limit) / 120 |
| log Z | -14.29 | -14.77 | -12.67 | -11.94 |
| melt_f | 3.50 [1.81, 11.70] | 4.78 [1.85, 12.50] | 6.24 [2.91, 13.16] | 7.62 [2.93, 14.96] |
| prcp_fac | 0.68 [0.17, 4.05] | 0.76 [0.18, 6.20] | 0.51 [0.18, 1.58] | 1.12 [0.32, 5.87] |
| temp_bias | 1.20 [-1.60, 4.82] | -0.06 [-2.86, 5.50] | -0.70 [-2.41, 1.85] | 0.79 [-1.23, 3.96] |
| log10 A factor | -0.24 [-0.83, 0.61] | 0.14 [-0.72, 0.79] | 0.55 [0.01, 0.92] | 0.08 [-0.55, 0.59] |
| initial state offset (K) | -0.55 [-1.70, 1.23] | -0.91 [-1.76, 0.36] | -1.11 [-1.82, -0.16] | -0.95 [-1.75, 1.02] |
| volume 2020 / 2004 | 0.63 [0.26, 0.76] | 0.64 [0.41, 0.74] | 0.24 [0.02, 0.47] | 0.32 [0.07, 0.49] |
| runoff 2000-19 (Mt/yr) | 12.4 [8.9, 36.3] | 13.5 [9.8, 48.7] | 1.04 [0.81, 1.69] | 0.78 [0.49, 2.51] |
| Jul-Sep fraction | 0.77 [0.64, 0.89] | 0.77 [0.59, 0.86] | 0.83 [0.73, 0.92] | 0.81 [0.70, 0.91] |
| glacier melt share | 0.89 [0.59, 0.98] | 0.90 [0.53, 0.98] | 0.92 [0.73, 0.98] | 0.87 [0.67, 0.97] |

Vernagtferner without the initial state period (job 599808): 9 iterations, log Z -13.68
(vs -14.29 with), log10 A -0.01 [-0.69, 0.84]: A not constrained either (with the 20%
thickness error of the earlier version, it was).

- The MB parameters stay equifinal everywhere; `prcp_fac` is the widest (factor 9-35
  between 5% and 95%), and sets the runoff spread (factor 2-5).
- The initial state offset leans cold on all four glaciers (medians -0.55 to -1.11 K):
  the observations prefer a starting glacier larger than the inverted one, as OGGM's
  dynamic spinup assumes. The cold side reaches the -2 K bound (5% quantiles -1.7 to
  -1.8, prior 5% -1.63): the bound cuts it. On RGI60-11.00834 warm offsets are ruled out.
- Glen's A is unconstrained except on RGI60-11.00834 (soft ice, 1-8x the default).
- The small glaciers lose 50-98% of their 2004 volume by 2020 in the plausible runs (the
  medians 68-76%); some disappear.
- RGI60-11.00731 hit the 20-iteration limit (Neff/Ne 0.24); kept at 20 (Fabien).
- Did not work at first: the container setup (small /tmp, no DNS without
  /etc/resolv.conf on the nodes, the download cache's symlinks): see dev-notes.
- Minor bug found: a particle whose glacier disappears gets `status`
  `ZeroDivisionError` (mean thickness = volume / 0) instead of a clear status; the
  weight is zero either way (188 of the 249 failed runs of Vernagtferner's first
  iteration). Not fixed (ask Fabien).
