Swmm Uncertainty

SWMM Uncertainty

Part of Agentic SWMM — install the project first for the executable toolchain (aiswmm CLI, SWMM solver, MCP servers).

Agent path without observed data

The honest split (live findings F-107 and F-109, 2026-09-03): WITH observed flow, swmm_sensitivity_oat / swmm_sensitivity_morris / swmm_sensitivity_sobol rank parameters against the data (they need an observed series and a patch map). WITHOUT observed flow, propagate_parameter_ranges is the tool for both questions: mode=one_at_a_time varies each parameter alone in one call and returns a per-parameter spread and a ranking ("which parameters matter most"); the default joint mode samples all ranges together and reports the spread ("how uncertain is the peak"). Never emulate a ranking with one sweep per parameter. Rainfall: a request to scale the observed event by factors (0.8, 1.0, 1.2) on a model with inline rain is run_climate_scenarios with those factors (live finding F-112, 2026-09-03); swmm_rainfall_ensemble needs a prepared rainfall series file and a JSON config (perturbation or IDF).

propagate_parameter_ranges is the typed tool for "how uncertain is the peak if Manning's n and imperviousness vary". It applies each named parameter globally (the same value on every subcatchment or conduit), runs SWMM once per sample through the audited runner, and writes 09_audit/parameter_sweep.json and .md with the baseline peak, the min/median/max over the samples, the spread as a percent of the baseline and the dominant parameter. Ranges are a mapping such as {"n_imperv": [0.010, 0.020], "pct_imperv": [60, 80]}; aliases manning_n, imperviousness, conduit_roughness, n_perv, s_imperv, s_perv, width, slope. It is prior sensitivity, not calibrated uncertainty; the per-object fuzzy and Monte Carlo workflows below remain the research path.

What this skill provides

• User-defined fuzzy membership functions for SWMM parameters.
• Baseline-aware triangular fuzzy numbers, where the current model value is the default triangle peak.
• Alpha-cut transformation from fuzzy membership functions to parameter intervals.
• LHS, random, or boundary sampling inside each alpha-cut interval.
• Monte Carlo parameter sampling for prior or calibration-informed probability distributions.
• Normal/lognormal/truncated-normal/uniform sampling with simple physical constraints such as bound parameters and greater-than rules.
• Batch propagation through SWMM by reusing the existing calibration patch-map convention.
• Normalized Shannon entropy metrics for output ensembles, such as hydrograph entropy over time.
• Machine-readable uncertainty summaries for output envelopes, entropy records, and failed/invalid samples.
• Sensitivity-analysis screening with three sub-methods (OAT / Morris / Sobol') sharing one entry point (scripts/sensitivity.py).
• Rainfall-forcing ensembles: time-series perturbation of an observed rainfall record (gaussian, multiplicative, AR(1), intensity_scaling) or IDF-curve sampling of design storms (Chicago / Huff / SCS Type II), with optional per-realisation SWMM runs and ensemble envelope aggregation.

This skill is intentionally separate from swmm-calibration.

• swmm-calibration asks: which parameter set best matches observations?
• swmm-uncertainty asks: how much output uncertainty is induced by user-defined parameter uncertainty, and which parameters drive that uncertainty?

Calibration requires observed data and performance metrics such as NSE, RMSE, or KGE. This skill can run without observed data when the task is prior uncertainty propagation. The sensitivity-analysis path does read an observed series (it scores trials by RMSE against observed flow), but it answers a different question from calibration: "which parameter spread matters?" rather than "which single set is best?". When calibration outputs exist, they can be used to narrow Monte Carlo ranges or define posterior-like parameter sets.

Scripts

• scripts/fuzzy_membership.py
  • parses and validates crisp, interval, triangular, and trapezoidal fuzzy parameter specs
  • resolves baseline: "from_model" from the base INP through the patch map
  • computes alpha-cut intervals
• scripts/sampling.py
  • generates parameter sets from alpha-cut intervals
  • supports lhs, random, and boundary
• scripts/probabilistic_sampling.py
  • generates Monte Carlo parameter sets from probability distributions
  • supports uniform, normal, truncnorm, and lognormal
  • supports simple constraints such as bind, greater_than, and less_than
• scripts/parameter_recommender.py
  • inspects an INP and recommends prior Monte Carlo parameters that are actually present in the model
  • reports the evidence boundary so prior ranges are not mistaken for calibrated posterior ranges
• scripts/monte_carlo_propagate.py
  • extracts node-flow ensembles from Monte Carlo trial .out files
  • calls entropy_metrics.py to produce node entropy JSON records
  • plots normalized output entropy curves for selected nodes
• scripts/entropy_metrics.py
  • calculates normalized discrete Shannon entropy for output ensembles
  • summarizes ensemble p05/p50/p95/min/max time series
• scripts/uncertainty_propagate.py
  • main CLI entry point
  • writes resolved fuzzy space, alpha intervals, parameter sets, trial INPs, and summary JSON
  • optionally executes SWMM and aggregates peak/continuity envelopes
• scripts/sensitivity.py
  • unified sensitivity-analysis entry point with three sub-methods
    • --method oat: one-at-a-time perturbation around a baseline (port of the legacy parameter_scout)
    • --method morris: Morris elementary-effects via SALib; sample budget r * (k + 1); reports mu_star and sigma per parameter
    • --method sobol: Sobol' indices via SALib (Saltelli sampling); sample budget N * (2k + 2); reports first-order S_i and total-effect S_T_i
  • writes a sensitivity_indices.json summary (typically under runs/<case>/09_audit/)
  • the Morris and Sobol' paths require SALib (declared in pyproject.toml)
• scripts/rainfall_ensemble.py
  • rainfall ensemble generator with two methods
    • --method perturbation: noisy realisations of an observed rainfall timeseries (CSV or SWMM .dat). Models: gaussian_iid, multiplicative, autocorrelated (AR(1)), intensity_scaling. Flag preserve_total_volume rescales each realisation to match the observed total when set
    • --method idf: synthesised hyetographs from IDF parameters (a, b, c) with confidence intervals. Storm types: chicago (Keifer-Chu), huff (4 quartiles), scs_type_ii (canonical 24-hr Type II)
  • if --base-inp is supplied, every realisation is patched into the base INP's [TIMESERIES] block and run through swmm5; peak flow + total outfall volume at --swmm-node are aggregated into swmm_ensemble_stats
  • writes per-realisation CSVs under <run-root>/09_audit/rainfall_realisations/ and a v1 summary at <run-root>/09_audit/rainfall_ensemble_summary.json
• scripts/source_decomposition.py — integration deliverable (issue #55)
  • pure-function over <run_dir>/09_audit/: reads whichever raw uncertainty outputs are present (Sobol' / Morris from sensitivity_indices.json, DREAM-ZS from posterior_samples.csv + chain_convergence.json, SCE-UA from candidate_calibration.json, rainfall ensemble from rainfall_ensemble_summary.json, MC propagation from uncertainty_summary.json)
  • emits uncertainty_source_summary.md (paper-reviewer-facing) and uncertainty_source_decomposition.json (schema_version 1.0)
  • the markdown body contains the five required sections: Output uncertainty envelope, Parameter contribution (Sobol' total-effect, sorted), Input contribution (rainfall ensemble vs parameter), Structural assumptions (not quantified), Cross-references
  • top of the markdown carries an Evidence Boundary code block that lists every potential method as ✓ ran or ✗ not run — partial runs are still reported, just with the absent methods flagged so no method is silently dropped
  • regenerate on demand with python3 -m agentic_swmm.cli uncertainty source <run_dir>; exits 0 on a complete run, 0 with a stderr warning on a partial run, and 1 when no uncertainty raw outputs exist at all
  • automatically re-invoked by skills/swmm-experiment-audit/scripts/audit_run.py whenever any of the raw artefacts is present in 09_audit/, so the integrated report always lives next to the audit note

Sensitivity-analysis sub-modes

The three modes share the patch-map workflow and the --observed series so that trials can be scored by RMSE against the same target flow node.

OAT is the cheapest, Morris is the standard screening method, and Sobol' decomposes variance into first-order and total-effect contributions (more expensive but more informative).

Expected fuzzy workflow

1. Prepare a base SWMM INP.
2. Prepare a calibration-style patch_map.json.
3. Define a fuzzy_space.json.
4. Define an uncertainty_config.json.
5. Run uncertainty_propagate.py.
6. Inspect uncertainty_summary.json, alpha_intervals.json, and generated trial directories.

Expected Monte Carlo / entropy workflow

1. Prepare a base SWMM INP.
2. Prepare a calibration-style patch_map.json.
3. Define a monte_carlo_space.json with parameter distributions.
4. Generate parameter sets with probabilistic_sampling.py.
5. Propagate the generated parameter sets through SWMM using the uncertainty runner path.
6. Extract an output ensemble, such as node,OUT_0,Total_inflow.
7. Calculate normalized output entropy with entropy_metrics.py.

If observed data are available, first run swmm-calibration and use its best, acceptable, or narrowed parameter ranges as a calibration-informed Monte Carlo input. If observed data are not available, report the analysis as prior uncertainty propagation.

Fuzzy Space

For a triangular membership function, the preferred compact form is:

{
  "parameters": {
    "pct_imperv_s1": {
      "type": "triangular",
      "lower": 15.0,
      "upper": 40.0,
      "baseline": "from_model"
    }
  }
}

The resolved triangle is:

triangular(a=lower, b=current model value, c=upper)

The baseline must lie inside [lower, upper]; otherwise the configuration is invalid.

A trapezoidal function can be fully specified:

{
  "parameters": {
    "n_imperv_s1": {
      "type": "trapezoidal",
      "lower": 0.010,
      "core_lower": 0.013,
      "core_upper": 0.018,
      "upper": 0.025
    }
  }
}

Or centered around the baseline:

{
  "parameters": {
    "n_imperv_s1": {
      "type": "trapezoidal",
      "lower": 0.010,
      "upper": 0.025,
      "core_width": 0.004,
      "baseline": "from_model"
    }
  }
}

CLI Example

python3 skills/swmm-uncertainty/scripts/uncertainty_propagate.py \
  --base-inp examples/todcreek/model_chicago5min.inp \
  --patch-map examples/calibration/patch_map.json \
  --fuzzy-space skills/swmm-uncertainty/examples/fuzzy_space.json \
  --config skills/swmm-uncertainty/examples/uncertainty_config.json \
  --run-root runs/uncertainty-demo \
  --summary-json runs/uncertainty-demo/uncertainty_summary.json \
  --dry-run

Remove --dry-run to execute SWMM for every generated trial.

Monte Carlo sampling example

python3 skills/swmm-uncertainty/scripts/probabilistic_sampling.py \
  --parameter-space skills/swmm-uncertainty/examples/monte_carlo_space.json \
  --samples 100 \
  --seed 42 \
  --out runs/uncertainty-mc/parameter_sets.json

Entropy metric example

python3 skills/swmm-uncertainty/scripts/entropy_metrics.py \
  --ensemble-json skills/swmm-uncertainty/examples/entropy_ensemble.json \
  --bins 10 \
  --out runs/uncertainty-mc/entropy_summary.json

Tecnopolo Monte Carlo smoke example

python3 scripts/benchmarks/run_tecnopolo_mc_uncertainty_smoke.py \
  --samples 20 \
  --seed 42 \
  --node OUT_0 \
  --scan-nodes \
  --entropy-nodes J6 OUT_0

This is a prior uncertainty smoke test, not calibration. It identifies perturbable parameters in the Tecnopolo HORTON prepared INP, applies small Monte Carlo perturbations, runs SWMM, optionally ranks all junction/outfall nodes by peak-flow spread, and writes summary.json, parameter_recommendations.json, trial outputs, a rainfall-plus-flow envelope figure, J6/OUT_0 entropy JSON files, and an entropy curve figure under runs/benchmarks/tecnopolo-mc-uncertainty-smoke/.

Sensitivity-analysis examples

OAT (port of the legacy parameter_scout):

python3 skills/swmm-uncertainty/scripts/sensitivity.py \
  --method oat \
  --base-inp examples/todcreek/model_chicago5min.inp \
  --patch-map examples/calibration/patch_map.json \
  --base-params examples/calibration/base_params.json \
  --scan-spec examples/calibration/scan_spec.json \
  --observed examples/calibration/observed_flow.csv \
  --run-root runs/sensitivity-oat \
  --summary-json runs/sensitivity-oat/09_audit/sensitivity_indices.json \
  --swmm-node O1

Morris elementary-effects (r=10 trajectories on a 4-parameter space gives 50 swmm5 calls):

python3 skills/swmm-uncertainty/scripts/sensitivity.py \
  --method morris \
  --base-inp examples/todcreek/model_chicago5min.inp \
  --patch-map examples/calibration/patch_map.json \
  --parameter-space examples/calibration/search_space.json \
  --observed examples/calibration/observed_flow.csv \
  --run-root runs/sensitivity-morris \
  --summary-json runs/sensitivity-morris/09_audit/sensitivity_indices.json \
  --morris-r 10 \
  --seed 42

Sobol' indices (N=64 on a 4-parameter space gives 640 swmm5 calls; budget is N*(2k+2)):

python3 skills/swmm-uncertainty/scripts/sensitivity.py \
  --method sobol \
  --base-inp examples/todcreek/model_chicago5min.inp \
  --patch-map examples/calibration/patch_map.json \
  --parameter-space examples/calibration/search_space.json \
  --observed examples/calibration/observed_flow.csv \
  --run-root runs/sensitivity-sobol \
  --summary-json runs/sensitivity-sobol/09_audit/sensitivity_indices.json \
  --sobol-n 64 \
  --seed 42

All three modes share the same --summary-json schema header (method, parameters, sample_budget, indices). Per-parameter shapes differ by method (see the "Sensitivity-analysis sub-modes" table above).

Rainfall ensemble examples

Time-series perturbation (200 noisy realisations of an observed rainfall CSV, all run through swmm5):

python3 skills/swmm-uncertainty/scripts/rainfall_ensemble.py \
  --method perturbation \
  --config skills/swmm-uncertainty/examples/rainfall_perturbation_config.json \
  --run-root runs/rainfall-ensemble-perturbation \
  --base-inp examples/todcreek/model_chicago5min.inp \
  --series-name TS_RAIN \
  --swmm-node O1 \
  --seed 42

IDF-curve design storm (200 hyetographs from sampled Chicago IDF params):

python3 skills/swmm-uncertainty/scripts/rainfall_ensemble.py \
  --method idf \
  --config skills/swmm-uncertainty/examples/rainfall_idf_config.json \
  --run-root runs/rainfall-ensemble-idf \
  --base-inp examples/todcreek/model_chicago5min.inp \
  --series-name TS_RAIN \
  --swmm-node O1 \
  --seed 42

Use --dry-run to skip the SWMM execution layer and write only the realisation CSVs + rainfall-only summary statistics.

Rainfall ensemble — methods at a glance

gaussian_iid adds zero-mean Gaussian noise (mean residual ≈ 0). multiplicative preserves the shape — Pearson correlation between observed and any realisation stays near 1. autocorrelated produces noise with lag-1 autocorrelation ≈ ar1_coefficient. intensity_scaling scales noise variance with intensity, so peaks fluctuate more than troughs.

When preserve_total_volume=true, every realisation is rescaled so its integrated rainfall depth matches the observed total. When false, totals vary across the ensemble — that variance is itself part of the propagated uncertainty.

Uncertainty source decomposition — integration deliverable

After at least one of the prior uncertainty steps has run (sensitivity, DREAM-ZS posterior, SCE-UA calibration, rainfall ensemble, or MC propagation) the integration layer collects the raw outputs and writes a single paper-reviewer-facing report:

python3 -m agentic_swmm.cli uncertainty source runs/<case>
# writes:
#   runs/<case>/09_audit/uncertainty_source_summary.md
#   runs/<case>/09_audit/uncertainty_source_decomposition.json   (schema_version 1.0)

The markdown body has five fixed sections per the PRD template:

1. Output uncertainty envelope — MC propagation peak-flow envelope + rainfall-driven peak-flow envelope when both exist.
2. Parameter contribution (Sobol' total-effect, sorted) — Sobol' S_T_i ranking; falls back to Morris mu_star with an explicit "screening only" note when Morris ran instead.
3. Input contribution (rainfall ensemble vs parameter) — side-by-side rainfall stats and top-3 parameter contributions plus a textual comparison.
4. Structural assumptions (not quantified) — model-structural, boundary-condition, observation-noise.
5. Cross-references — relative paths to every raw artefact + posterior plots.

The top of the markdown carries an Evidence Boundary code block that lists every potential method as ✓ or ✗:

Evidence boundary:
  Sobol' SA       : ✓ ran (sensitivity_indices.json)
  Morris SA       : ✗ not run
  DREAM-ZS        : ✓ ran (posterior_samples.csv)
  SCE-UA          : ✓ ran (candidate_calibration.json)
  Rainfall ensemble: ✓ method A only (method B not run)
  MC propagation  : ✓ ran (uncertainty_summary.json)

The audit pipeline (skills/swmm-experiment-audit/scripts/audit_run.py) auto-runs the decomposition at audit-end whenever any uncertainty raw artefact is present in 09_audit/, so the integrated report stays in sync with the audit note.

Exit codes for the CLI:

• complete run (every method ran) → 0
• partial run (some methods absent) → 0 with a warning: line on stderr naming the absent methods
• no uncertainty raw outputs anywhere → 1

Outputs

The run directory contains:

• fuzzy_space.resolved.json
• alpha_intervals.json
• parameter_sets.json
• trials/<trial>/model.inp
• trials/<trial>/manifest.json when SWMM execution is enabled
• uncertainty_summary.json

The summary answers:

• What parameter interval was used at each alpha level?
• What samples were propagated?
• How many trials succeeded, failed, or were only dry-run trials?
• What peak-flow and continuity envelopes were induced by each alpha level?
• What output entropy curve was induced by the propagated ensemble?

Known limitations

• Fuzzy analysis focuses on epistemic parameter uncertainty through membership functions.
• Monte Carlo analysis supports prior or calibration-informed probability distributions.
• The current model value is treated as the most plausible value for compact triangular specs.
• Entropy is calculated from SWMM output ensembles; it is parameter-induced output entropy, not parameter entropy and not calibration performance.
• Real SWMM propagation depends on swmm5 being installed.
• Hydrograph goodness-of-fit metrics remain in swmm-calibration; this skill can be extended later to call that observed-flow path.