Imported from lzwei196/KISS-Knowledge-Infrastructure-for-Scientific-Simulation (models/Alpine3D/SKILL.md). Install upstream with npx skills add lzwei196/KISS-Knowledge-Infrastructure-for-Scientific-Simulation --skill Alpine3D. Copyright stays with the author.

MANDATORY EXECUTION POLICY — READ BEFORE PROCEEDING

You MUST run the actual model binary or package described in this document. If the model fails to import, compile, or execute, you MUST:

1. Check diagnostics/triplets.yaml for a matching error pattern
2. Try the fix suggested in the triplet's remedy section
3. If still failing, report the error to the user with full details

You MUST NOT substitute a simplified Python formula, regression equation, or hand-coded approximation in place of the real model.

Before starting, run: python preflight_check.py (in this KI directory) to verify that the model binary/package and required data are available.

DEBUGGING PROTOCOL — When something goes wrong, follow this order:

1. Check triplets — diagnostics/triplets.yaml may already cover this error
2. Read official docs — The model's own documentation for expected formats/units
3. Find working examples — Check outputs/ or the model's shipped test data
4. Fix the tool — With knowledge of what "correct" looks like

Do NOT write custom debug scripts. The answers are in the docs and examples.

KI map — what to read, and when

Projected 2026-08-17 from the KI's actual contents — 9 components present. Refresh: python3 ki_tools_common/generate_skill_map.py --ki_dir <this KI>.

Executable tool index (projected — complete by construction)

Every public tool in this KI, by exact path. What each is FOR lives in the human-written Tool Inventory above; --help on any of these prints its arguments.

4 public tools; _-prefixed helpers and packaging files excluded.

Alpine3D Knowledge Infrastructure

• Package: hydrocraft-alpine3d-snow
• Version: 1.0.0
• Model: Alpine3D (with MeteoIO + SNOWPACK)
• Domain: Cryosphere — spatially distributed snow, energy balance, and runoff in mountainous terrain
• Created: 2026-03-26
• Last updated: 2026-04-30
• Tools: 4 | Skill Documents: 6 | Diagnostic Triplets: 22 | Validation: T3 (3 SNOTEL sites, 5 WY each)

⛔ READ FIRST — the SNOTEL precipitation column is NOT cumulative (2026-08-09)

The NRCS SNOTEL daily export column

Precipitation Accumulation (in) Start of Day Values

is the DAILY INCREMENT, in inches, despite the word "Accumulation" in its name. It is not a water-year running total. Evidence at SNOTEL 590 (2005-10 → 2015-09): the raw column sums to 1069 mm/yr, its largest single value in ten years is 2.10 in (a water-year cumulative would have to exceed 40 in), and only 70% of day-to-day steps are non-negative.

Every Alpine3D run in this KI up to 2026-08-09 applied .diff().clip(lower=0) to that column (t3_runs/build_site.py → build_snotel_forcing), which keeps only the day-over-day increases and discards ~44% of the precipitation:

This single data bug is the −40…−60% SWE PBIAS reported at every validated site below, and the reason NSE stalled in the 0.31–0.39 band. The physical explanations previously given for that bias — "the gauge ~1660 mm/yr cannot build the observed 2.3 m pillow" at Paradise, "pillow over-read", "orographic undercatch" — attribute a parsing bug to physics. Note that the dt_023 maritime remedy factor (×1.9) is close to the reciprocal of the Paradise loss (3832/1740 = 2.2): it was numerically compensating for the differencing, not correcting the gauge.

Correct handling: use the column AS IS, converting inches → mm (convert_forcing_to_smet.py --precip-unit in). At a fresh continental site with no precip factor at all this yields NSE ≈ 0.83 / PBIAS ≈ −10% on daily SWE. Before re-deriving any SWE bias as a physical effect, check the forcing total against the raw gauge.

Resolving a SNOTEL station (coordinates, elevation, timezone)

Do not hand-type site coordinates. The NRCS AWDB REST API answers directly and is reachable from this host:

curl -s "https://wcc.sc.egov.usda.gov/awdbRestApi/services/v1/stations?stationTriplets=590:MT:SNTL&activeOnly=false"
# -> name, latitude, longitude, elevation (FEET — multiply by 0.3048),
#    dataTimeZone, beginDate/endDate

The station triplet is <id>:<STATE>:SNTL; the id is the filename stem of KISSPATH_OBS/snotel/<id>_daily.csv and the state is in that file's # SNOTEL <id>: <Name>, <ST> comment line.

Validated Test Cases (Tier-3, real DEM + real obs)

Located at KISSPATH_BINARIES/Alpine3D/t3_runs/:

⚠️ Every row in the table below was produced with the differenced (i.e. ~44% too dry) precipitation described above. Treat their NSE/PBIAS as a floor, not as Alpine3D's skill. The 2026-08-09 Lone Mountain row is the first run with correct precipitation.

Continental sites have positive skill against daily SNOTEL SWE. Maritime sites (Paradise) fail two ways under single-station daily disaggregation: (1) orographic precip is undercaught and the SNOTEL gauge under-reads heavy wet snow (gauge ~1660 mm/yr cannot build the observed 2.3 m pillow), and (2) the synthetic ±6 K diurnal TA cycle + high roughness drive excessive turbulent melt, so the pack peaks in Jan–Mar and melts out months early (Apr–Jun sim/obs SWE ratio ≈ 0.09). The undercatch_wmo filter (dt_020) alone is far too weak. Maritime remedy (dt_023, added 2026-06-19): apply a multiplicative precip undercatch factor (~1.9) via convert_forcing_to_smet.py --precip-undercatch-factor, damp the diurnal TA amplitude to ~2.5 K, lower ROUGHNESS_LENGTH to ~0.003 m, and raise the PSUM_PH snow threshold to ~275.35 K. build_site.py now carries these as per-site precip_factor/diurnal_amp/roughness/snow_thresh_K overrides on Paradise_WA. Driver: t3_runs/build_site.py + validate_all.py.

Maritime remedy VERIFIED (2026-06-19, SNOTEL 679, WY2006–2015): the dt_023 overrides were tested on the full 2006-01-01..2015-12-31 daily SWE window and PASS: NSE +0.33, r 0.71, KGE +0.46, PBIAS −40% (n=3651 days; cold start 2005-10-01, scored from 2006-01-01). This overturns the −0.22 baseline. The pack still under-accumulates (sim mean 511 vs obs 853 mm; the ×1.9 factor is a floor, not a ceiling) but peak SWE (2442 vs 2764 mm) and timing (r 0.71) are now good. Three new build_site.py knobs make this run reproducible and tractable:

• Period is now configurable per-site via sim_start / sim_end / eval_start (was hardcoded WY15-19). Cold-start the autumn before the eval window; observed_swe.csv is filtered to eval_start so the spinup autumn is excluded from metrics.
• Deep-maritime tractability: a 10-yr Paradise run at 11×11/15-min is compute-bound (the correct 2+ m pillow grows the finite-element count, so per-step cost balloons — ~5 h+ and climbing). Single-station point-SWE only scores the POI cell, so grid_n=1 + calc_step=60 + light_output=True (1×1 column, 1 SNOWPACK sub-step/hr, grids/prof off, daily TS) cuts it to ~15 min with no change to the scored center-cell SWE. The validated 11×11 config remains the reference for any spatial fidelity claim.
• POI centering fix: the POI is now written at the centre of the centre cell (poi_e/poi_n), not the cell corner. A corner POI is on the domain edge and Alpine3D aborts with InvalidArgument: Invalid POI — fatal on a 1×1 grid. Harmless (+50 m, same cell) on larger grids.
• validate_all.py POI glob generalized from the hardcoded 5_* index to *_<exp>.smet, so it finds the POI file for any grid size (1×1 → 0_0_…).

Note: the build_site.py io.ini template has no [SnowpackAdvanced] section, so ALPINE3D defaults FALSE (dt_016 warning at startup) — the run is effectively point-mode per cell. For single-station single-POI SWE validation this is benign (and consistent with the 1×1 reduction above), but a distributed/spatial run MUST add [SnowpackAdvanced]\nALPINE3D = TRUE.

Data Preparation

Forcing data

Data Sources: Use from ki_tools_common.load_forcing import load_daily_forcing for CMFD/MSWX/NASA POWER.

Data Validation Reference: See data_ki/CMFD/SKILL.md for atmospheric forcing documentation. See data_ki/SNOTEL/SKILL.md for snow observations. See data_ki/BedMachine/SKILL.md for ice topography. See data_ki/MEaSUREs/SKILL.md for ice velocity.

Overview

Alpine3D is a spatially distributed, three-dimensional model for analyzing and predicting snow-dominated surface processes in mountainous topography. It couples:

• MeteoIO — meteorological data I/O, quality filtering, temporal/spatial interpolation
• SNOWPACK — detailed multi-layer snow and soil physics (energy + mass balance)
• EnergyBalance — terrain-aware radiation transfer (shading, view factors, terrain reflection)
• SnowDrift — wind-driven snow erosion, saltation, suspension, and deposition (optional)
• Runoff — integrated surface/soil runoff from snowmelt and rainfall

The model reads meteorological station data (SMET format), a DEM (ARC grid), land-use grids, and initial snow profiles (.sno), then produces gridded fields of SWE, snow depth, temperature, radiation, and runoff at configurable time intervals.

Primary applications: snow water resources, avalanche forecasting, climate change impact assessment, permafrost studies, and ski resort management.

6. Output Description

This section restates the KI's dag facts. If this section and dag.yaml disagree, dag.yaml wins.

Headline output (validation_rank: 1):

SWE — Snow water equivalent per pixel (bulk snowpack mass). (kg/m2)

The model also writes gridded ARC outputs and POI .met time series as described in "Output File Formats" below. For scoring and obs-binding, treat SWE as the headline variable because the dag ranks it first.

8. Unit Conversion Table

Exact I/O shapes live in docs/format_spec.yaml; this table summarizes the unit conversions and traps already documented in this KI body. Alpine3D uses SI/MKSA units internally.

Output units explicitly documented in this KI body:

11. Validated Results

This section restates the KI's validation facts and the convention bar. If this section and docs/validation_convention.yaml disagree, the convention file wins.

Performance Metrics — judged against the field's bar, not intuition

SWE is judged by nrmse with direction minimize.

No nrmse achieved value is stated in the validated-run body below. Existing validated SWE rows report r, NSE, KGE, and PBIAS; do not translate those metrics into the convention bar without a sourced nrmse value.

Documented Validated SWE Runs

Rows marked "differenced precipitation" were produced before the 2026-08-09 SNOTEL precipitation correction documented above. Treat their NSE/PBIAS as a floor, not as Alpine3D's skill.

Installation

Dependencies

Build from Source

The source tree must have MeteoIO, SNOWPACK, and Alpine3D as siblings:

Source/
├── meteoio/
├── snowpack/
└── alpine3d/

Build sequence (MeteoIO → SNOWPACK → Alpine3D):

# 1. Build MeteoIO
cd Source/meteoio && mkdir build && cd build
cmake .. && make -j$(nproc) && sudo make install

# 2. Build SNOWPACK
cd ../../snowpack && mkdir build && cd build
cmake .. && make -j$(nproc) && sudo make install

# 3. Build Alpine3D
cd ../../alpine3d && mkdir build && cd build
cmake .. -DOPENMP=ON   # add -DMPI=ON for cluster support
make -j$(nproc) && sudo make install

Binary Location

After build: alpine3d/build/bin/alpine3d After install: /usr/local/bin/alpine3d or the GeoForge-managed location: KISSPATH_BINARIES/Alpine3D/source/repo/Source/alpine3d/bin/alpine3d

Test

cd alpine3d/tests/simple
PROG_ROOTDIR=../../build ./run_simple.sh
# Runs 3-month Dischma simulation, compares with reference output

Pipeline Stages

Parallelism notes: Stages s2 and s3 can run in parallel. Stage s5 supports OpenMP (shared memory) and MPI (distributed memory) parallelism for the energy balance and snowpack computations.

Tools Reference

Options added 2026-08-09 (a full SNOTEL site can now be built with the KI tools alone; t3_runs/build_site.py is no longer required):

Reference end-to-end driver using only these tools: KISSPATH_KI_ROOT/Alpine3D/run_and_score.py (SNOTEL 590 Lone Mountain, MT; resumable at every stage).

Unit Trap Table

CRITICAL: Alpine3D uses SI/MKSA units internally. The most common cause of silent model failure is feeding data in wrong units. The table below documents every conversion trap:

Configuration Reference (io.ini)

[General]

[Input]

[Output]

[Snowpack]

[SnowpackAdvanced]

[EBalance]

[Interpolations2D] — Spatial Interpolation

Algorithms per variable: VAR::algorithms = ALG1 ALG2 ...

[Filters] — Input QC

Pattern: VAR::filterN = TYPE + VAR::argN::param = value

Common filters: min_max, rate, mad, undercatch_wmo

[Generators] — Gap Filling

Pattern: VAR::generatorN = TYPE

Common generators: ALLSKY_LW, CLEARSKY_LW, ISWR_ALBEDO

Command-Line Options

alpine3d [options]

Required:
  -a, --startdate=YYYY-MM-DDTHH:MM    Simulation start date
  -z, --enddate=YYYY-MM-DDTHH:MM      Simulation end date (or use --steps)
  -i, --iofile=<file>                  Configuration file (default: io.ini)

Optional:
  -n, --steps=<N>                      Number of timesteps (alternative to --enddate)
  -b, --np-ebalance=<N>               Workers for energy balance (default: 1)
  -p, --np-snowpack=<N>               Workers for snowpack (default: 1)
  --enable-eb                          Enable energy balance module
  --restart                            Restart from existing .sno files
  -h, --help                           Display help

Input File Formats

SMET (Swiss Meteorological Exchange Text) — Meteo Data

SMET 1.1 ASCII
[HEADER]
station_id       = WFJ2
station_name     = Weissfluhjoch
latitude         = 46.829897
longitude        = 9.809315
altitude         = 2540.0
nodata           = -999
tz               = 1
fields           = timestamp TA RH VW DW ISWR ILWR PSUM HS
[DATA]
2014-10-01T00:00  277.25  0.85  3.2  270  0.0  285.0  0.0  0.0
2014-10-01T01:00  276.80  0.87  2.8  265  0.0  282.0  0.5  0.001

SNO (Initial Snow/Soil Profile)

The layer columns are parsed POSITIONALLY by snowpack/plugins/SmetIO.cc::readSnowCover (~L417–457), so the fields line must contain exactly these 19 tokens, in this order — including ne, the number of finite elements per layer (use 1). Omitting ne shifts every later column one position left, so CDot is read as the element count and the profile initialises with zero elements.

Header key names are looked up literally: it is SoilAlbedo (not SoilAlb) and CanopyLeafAreaIndex (not CanopyLAI).

Layers are read BOTTOM → TOP: the first [DATA] row is the deepest soil layer. Put the thick layers first and the thin near-surface layers last, and make temperature decrease from the first row to the last.

For a projected DEM, use easting / northing / epsg — SMET accepts that pair as an alternative to latitude/longitude. Writing projected metres into latitude/longitude makes MeteoIO reject the profile.

SMET 1.1 ASCII
[HEADER]
station_id       = 0_0_my_experiment
easting          = 466550.00
northing         = 5013550.00
epsg             = 32612
altitude         = 2500
ProfileDate      = 2014-10-01T00:00
nSnowLayerData   = 0
nSoilLayerData   = 3
SoilAlbedo       = 0.2
BareSoil_z0      = 0.02
CanopyHeight     = 0.00
CanopyLeafAreaIndex = 0.00
CanopyDirectThroughfall = 1.00
WindScalingFactor = 1.00
ErosionLevel     = 0
TimeCountDeltaHS = 0.000000
fields           = timestamp Layer_Thick T Vol_Frac_I Vol_Frac_W Vol_Frac_V Vol_Frac_S Rho_S Conduc_S HeatCapac_S rg rb dd sp mk mass_hoar ne CDot metamo
[DATA]
2014-10-01T00:00  1.0  282.15  0.0  0.15  0.35  0.50  1800  1.5  900  0.3  0.2  0.0  1.0  0  0.0  1  0.0  0.0
2014-10-01T00:00  0.5  281.15  0.0  0.20  0.30  0.50  1800  1.5  900  0.3  0.2  0.0  1.0  0  0.0  1  0.0  0.0
2014-10-01T00:00  0.5  280.15  0.0  0.25  0.25  0.50  1800  1.5  900  0.3  0.2  0.0  1.0  0  0.0  1  0.0  0.0

generate_sno_files.py emits all of the above correctly as of 2026-08-09; use --naming landuse --experiment <EXPERIMENT> --landuse-code <code> --epsg <code> for a single per-class profile, or --naming pixel for one file per cell.

ARC Grid (DEM, Land Use)

Standard ESRI ARC/INFO ASCII grid:

ncols         100
nrows         80
xllcorner     782000
yllcorner     180000
cellsize      25
NODATA_value  -9999
1500 1502 1510 ...

POI (Points of Interest)

SMET 1.1 ASCII
[HEADER]
fields = easting northing altitude
[DATA]
785360  182255  2520
786100  181900  2350

Output File Formats

Gridded Output (ARC ASCII)

Written to GRID2DPATH at GRIDS_DAYS_BETWEEN intervals.

Filename pattern: YYYY-MM-DDThh.mm.ss_PARAM.asc (e.g. 2014-12-23T12.00.00_SWE.asc) — MeteoIO's ARC2D plugin writes ISO-8601 dates with . substituted for the colons, which are illegal in many filesystems. The compact YYYYMMDDHHMI_PARAM.asc form documented here before 2026-06-22 does not occur for SWE/HS/RSNO grids; a glob built on it matches nothing. (parse_alpine3d_output.py accepts both.) Note that some Alpine3D-internal grids — e.g. TSOIL1 — are written with the compact stamp, so a directory listing can show both conventions side by side.

Grid NODATA_value is −999, not −9999.

Key output parameters:

• HS — Snow depth (m)
• SWE — Snow water equivalent (kg/m²)
• TA — Air temperature (K)
• TSS — Snow surface temperature (K)
• ISWR — Incoming shortwave radiation (W/m²)
• ILWR — Incoming longwave radiation (W/m²)
• MS_SNOWPACK_RUNOFF — Surface runoff (kg/m²)
• MS_SOIL_RUNOFF — Bottom runoff (kg/m²)
• RSNO — Snow mean density (kg/m³)
• ET — Evapotranspiration (kg/m²)

Time Series (.met SMET)

At each POI, written at TS_DAYS_BETWEEN intervals. Contains full energy/mass balance terms.

Profiles (.pro)

Full snow stratigraphy at POI locations.

Critical Domain Knowledge

1. ALPINE3D = TRUE is mandatory (dt_016)

In [SnowpackAdvanced], ALPINE3D must be set to TRUE. If left at FALSE (Snowpack standalone default), the model will run but produce incorrect spatial results because the Snowpack library will not operate in distributed mode.

2. Temperature must be in Kelvin (dt_001)

All temperatures in SMET files must be in Kelvin. Feeding Celsius values creates temperatures near 0 K internally, causing immediate crashes or physically impossible surface energy balance.

3. Precipitation accumulation period must match timestep (dt_015)

PSUM::RESAMPLE1 = ACCUMULATE with PSUM::ARG1::PERIOD must match CALCULATION_STEP_LENGTH. Mismatches cause either double-counting or loss of precipitation.

4. DEM and land-use grids must be co-registered (dt_017)

The DEM and land-use grids must have identical ncols, nrows, xllcorner, yllcorner, and cellsize. Mismatch causes a cryptic segmentation fault.

5. Coordinate system must be consistent (dt_018)

COORDSYS in [Input] and [Output] must be the same and must match the actual coordinate system of the DEM file. Using WGS84 lat/lon with a projected DEM (or vice versa) causes all spatial interpolation to fail silently.

6. SMET timestamps must be ISO 8601

Format: YYYY-MM-DDTHH:MM or YYYY-MM-DDTHH:MM:SS. Any deviation (e.g., spaces instead of T) causes MeteoIO to fail to read the file.

7. Snow file naming convention (dt_024)

Alpine3D constructs the name it searches for in SnowpackInterface.cc (~L1258–1260) and reads whichever exists, per pixel:

per-pixel  (tried first)   <ix>_<iy>_<EXPERIMENT>.sno      e.g. 0_0_lone_mountain_mt.sno
per-class  (fallback)      <EXPERIMENT>_<landuse_code>.sno e.g. lone_mountain_mt_10851.sno

• EXPERIMENT is [Output]::EXPERIMENT from io.ini — not the station id.
• ix is the column, iy the row counted from the BOTTOM (MeteoIO grids are south-up), so a 1×1 domain wants 0_0_<EXPERIMENT>.sno.
• landuse_code is the rounded value in the land-use grid (e.g. PREVAH 10851).

Any other name means Alpine3D finds no initial profile and aborts. Use generate_sno_files.py --naming {pixel,landuse} --experiment <EXPERIMENT>.

8. Energy balance workers vs grid size

Setting --np-ebalance higher than the number of grid rows causes a crash. The number of workers must be ≤ number of DEM rows.

9. Relative humidity range trap (dt_002)

RH in SMET must be 0–1 (fraction). Values 0–100 (percent) will be clipped to 1.0 by the min_max filter, making the entire domain saturated and producing unrealistic snowfall.

Calibration Parameters

Quick Start

# 1. Build Alpine3D (from source directory)
cd Source/alpine3d && mkdir build && cd build
cmake .. -DOPENMP=ON && make -j$(nproc)

# 2. Create simulation directory
mkdir -p ~/sim/Dischma/{input/{meteo,surface-grids,snowfiles},output/{grids,snowfiles},setup}

# 3. Prepare forcing data
python tools/convert_forcing_to_smet.py \
  --input era5_data.nc --output ~/sim/Dischma/input/meteo/ \
  --stations WFJ2 DAV --start 2014-10-01 --end 2015-09-30

# 4. Generate initial snow profiles
python tools/generate_sno_files.py \
  --dem ~/sim/Dischma/input/surface-grids/dem.asc \
  --output ~/sim/Dischma/input/snowfiles/ \
  --date 2014-10-01T00:00 --n-soil-layers 3

# 5. Run simulation
python tools/run_alpine3d.py \
  --iofile ~/sim/Dischma/setup/io.ini \
  --startdate 2014-10-01T01:00 --enddate 2015-09-30T00:00 \
  --np-ebalance 4 --np-snowpack 4

# 6. Parse output
python tools/parse_alpine3d_output.py \
  --grid-dir ~/sim/Dischma/output/grids/ \
  --params SWE HS TA --output ~/sim/Dischma/results.csv

Diagnostic Triplets Summary

File Structure

ki/
├── SKILL.md                              # This file
├── tools/
│   ├── convert_forcing_to_smet.py        # Meteo data → SMET format
│   ├── generate_sno_files.py             # Initial snow/soil profiles
│   ├── run_alpine3d.py                   # Execution wrapper
│   └── parse_alpine3d_output.py          # Grid output → CSV
├── docs/
│   ├── s0_domain_setup.md                # Domain and DEM preparation
│   ├── s2_meteorological_forcing.md      # Forcing data preparation
│   ├── s3_initial_conditions.md          # Initial snow/soil profiles
│   ├── s5_execution.md                   # Running Alpine3D
│   └── s6_output_analysis.md             # Output parsing and analysis
└── diagnostics/
    └── triplets.yaml                     # 18 symptom→diagnosis→remedy entries