What does this page show?

Completed SWAT+ calibration results across a 43-basin CONUS engineer fleet with held-out validation, four high-skill exemplars, three west-central Florida regional MMPSO basins, and a Tampa Bay zero-shot pooled-prior transfer example — plus compact Florida/Illinois worked-example appendices with split-sample metrics.

What does “validation tracks calibration” mean here?

Across basins with a held-out validation split, the median change in monthly NSE from the calibration window to the validation window is −0.04 (paired val − cal). Skill does not collapse when the years change.

Why include a basin that calibrates poorly?

Honesty. The Illinois controlled-basin appendix reaches only modest skill (low NSE, larger PBIAS) and is shown as-is rather than hidden, so the examples represent the real range of outcomes, not just the best case.

What does "zero-shot pooled prior" mean?

A parameter distribution pooled from previously calibrated models of the region is applied to a new model in a single simulation, with no optimization at all. On the 77.6k-HRU Tampa Bay model this moved the 58-gauge roster from 19 to 25 Satisfactory-or-better gauges and 4 to 11 Very Good (Moriasi monthly ratings), and the correctly-assigned subset median monthly NSE from 0.05 to 0.50 — before any optimizer touched the model.

SWATGenX · calibration results

43 basins calibrated, coast to coast

Validated on held-out years — validation tracks calibration (median ΔNSE −0.04).

CONUS calibration fleet43 basins · 27 tested on unseen years

Holds up on unseen yearsWeaker on unseen yearsCalibration only

Fleet distribution

Each point is one engineer-calibrated basin with a held-out validation split. The dashed line is 1:1; the dashed crosshairs mark Moriasi monthly NSE = 0.5. The cloud shows how validation tracks calibration across the fleet.

Source: calibrationFleetSummary.json ← fleet_summary.py / fleet_summary.json. Paired median ΔNSE -0.039 (display -0.04).

High-skill exemplars

Four basins where monthly calibration and validation skill both land in the high range. Display values are rounded; exact scores are in the fleet JSON.

Mad-Redwood

California · USGS 11481200

0.95 / 0.79

monthly NSE · calibration / validation

Middle Ohio-Laughery

Kentucky · USGS 03262001

0.91 / 0.87

monthly NSE · calibration / validation

McKenzie

Oregon · USGS 14161500

0.88 / 0.92

monthly NSE · calibration / validation

Lehigh

Pennsylvania · USGS 01451800

0.87 / 0.92

monthly NSE · calibration / validation

Regional MMPSO basins

Three west-central Florida basins finished the same information-cascade path — pooled prior → diagnosis → short regional MMPSO — documented in the calibration methods write-up.

Peace River (57,998 HRUs)

Complete

Terminal calibration stage COMPLETE — regional MMPSO, converged at iteration 8.

Fleet calibration flagship: regionalized MMPSO seeded from the pooled prior, converged at iteration 8.

Field	Value
Target model	57,998-HRU Peace River basin model (HUC8 03100101, NHDPlus HR region 0310)
Gauge roster	38 evaluated → 11 structural exclusions (drainage-area mismatch, lake-outlet orphaning, regulated canals); the terminal objective couples the retained gauges as per-gauge sub-objectives — 21 returned computable NSE over the scored window
Regionalization	cn3_swf, perco, esco, epco expanded over 9 tributary regions — 73 optimizer dimensions (insensitive + basin-uniform parameters stay global)
Optimizer	MMPSO with per-gauge sub-objectives (by-gauge grouping); pooled-prior vector injected as a protected initial particle
Objective hardening	Per-gauge NSE floor at −1 inside the optimizer sum (reported metrics unclipped) — pathological reaches are protected in-objective and reported honestly, not hand-excluded
Convergence	Global best reached at iteration 8 and held byte-identical through iteration 23 (spot instance reclaimed) — 15 iterations of zero improvement; ~100 model evaluations at 12 particles

The search converged fast: the global-best objective (−19.25 in penalized units) was reached at iteration 8 and held unchanged for the remaining 15 iterations until the AWS spot instance was reclaimed — convergence, not a truncated run. A good pooled prior plus a fixed search space places the satisficing target within ~8 iterations, at roughly 100 model evaluations against the several hundred a from-scratch campaign would spend.

Per-gauge skill (2016–2022 calibration window (daily and monthly Moriasi NSE))

Channel	Group	Daily NSE	Monthly NSE
427	Mainstem / Charlie core	0.79	0.88
422	Mainstem / Charlie core	0.78	0.86
439	Mainstem / Charlie core	0.78	0.87
332	Mainstem / Charlie core	0.77	0.92
399	Mainstem / Charlie core	0.77	0.86
2718	Mainstem / Charlie core	0.68	0.77
269	Mainstem / Charlie core	0.64	0.91
451	Mainstem / Charlie core	0.63	0.76
220	Interior tributary	0.39	0.80
650	Interior tributary	0.24	0.47
968	Lower-basin (floor-protected)	0.19	0.12
603	Lower-basin, regulated (floor-protected)	0.12	−0.27
933	Tidal reach (floor-protected)	−0.27	−0.49
947	Tidal — daily degenerate, monthly recovers (floor-protected)	−2.36	0.53

Regionalization is visible in the solution: percolation (perco) ranges from 1.98 in Horse Creek to 2.83 in Shell Creek across tributary groups (Prairie 2.79, Charlie 2.70, Peace mainstem 2.52, other 2.63) — spatially-varying calibrated parameters a single global scalar could not produce, reconciling the opposite per-tributary biases.

Tampa Bay (77,636 HRUs)

Complete

Terminal calibration stage COMPLETE — regional MMPSO, converged at iteration 6.

Cross-basin case: pooled prior carried no Tampa Bay information; terminal search converged at iteration 6.

Field	Value
Target model	77,636-HRU Tampa Bay integrated model (70 HUC12s, HUC12-outlet 031002060700, region 0310) — a separate drainage feeding a nitrogen-limited estuary
Gauge roster	32 gauges retained after coverage + drainage-area vetting; 27 returned computable NSE over the scored window
Regionalization	runoff-controlling parameters (cn3_swf, perco, esco, epco) distributed by tributary sub-watershed; snow / storage / channel / baseflow stay global
Optimizer	MMPSO with per-gauge NSE sub-objectives (floor −1); pooled-prior vector (built entirely from the Peace-area pool, no Tampa Bay information) injected as a protected initial particle
Convergence	Global best (−15.48 in penalized units) reached at iteration 6 and held byte-identical through the final synchronized iteration — a converged plateau, not a truncated run

The well-gauged core is Satisfactory-to-Very-Good on the reaches that carry the basin (core daily NSE 0.52–0.70, monthly 0.62–0.85, with several interior gauges reaching monthly NSE 0.80–0.85). A cluster of seven small urban canals stays weak — their monthly PBIAS pins at roughly −100% (near-zero measured baseflow the coarse 500 m routing cannot reproduce) — and the per-gauge floor is why they neither dominated the search nor were quietly dropped. The same cascade, the same honesty rail: seeded from the pooled prior, converged in six iterations on a 77.6k-HRU model, weak tail reported at face value.

Per-gauge skill (2016–2022 calibration window (daily and monthly Moriasi NSE))

Channel	Group	Daily NSE	Monthly NSE
817	Well-gauged core	0.70	0.85
1106	Well-gauged core	0.61	0.74
837	Well-gauged core	0.60	0.68
1671	Well-gauged core	0.55	0.62
1979	Well-gauged core	0.54	0.85
2184	Well-gauged core	0.53	0.66
844	Well-gauged core	0.52	0.64
1868	Interior (strong monthly)	0.42	0.82
1013	Interior (strong monthly)	0.42	0.80
1495	Interior (strong monthly)	0.35	0.82
1265	Small urban canal (floor-protected)	−0.19	−0.78
1689	Small urban canal (floor-protected)	−0.28	−0.71
1267	Small urban canal (floor-protected)	−0.29	−0.42

Myakka River (29,069 HRUs)

Complete

Terminal calibration stage COMPLETE — regional MMPSO, converged at iteration 5.

Flat wet-prairie / flatwoods basin; coverage-vetted roster; converged at iteration 5.

Field	Value
Target model	29,069-HRU Myakka River model (HUC8 03100102, region 0310) — a flat wet-prairie / flatwoods basin draining to phosphorus-driven Charlotte Harbor
Gauge roster	11-gauge coverage-vetted roster — the corrected roster after a −99 sentinel value was discovered and scrubbed; 10 returned computable NSE
Regionalization	runoff-controlling parameters distributed by tributary sub-watershed, insensitive + basin-uniform parameters global (same dispatch as Peace / Tampa Bay)
Optimizer	MMPSO with per-gauge NSE sub-objectives (floor −1); pooled-prior vector injected as a protected initial particle
Convergence	Global best (−7.03 in penalized units) reached at iteration 5 and held byte-identical through the final synchronized iteration

Myakka is the flattest, most diffuse basin in the fleet, and its daily skill is inherently lower — but the whole coverage-vetted roster stays positive at both time scales. The core mainstem/wet-prairie gauges reach monthly NSE 0.56–0.65 (gauges 3592, 148, 191, 297), with daily NSE a modest 0.17–0.40 reflecting the flatwoods hydrology. No pathological floored tail: a fully-positive but honestly modest result from the same pooled-prior seed and the same short terminal search.

Per-gauge skill (2016–2022 calibration window (daily and monthly Moriasi NSE))

Channel	Group	Daily NSE	Monthly NSE
3592	Mainstem core	0.40	0.56
148	Mainstem core	0.33	0.64
191	Mainstem core	0.31	0.65
297	Interior tributary	0.21	0.64
676	Interior tributary	0.25	0.51
490	Interior tributary	0.25	0.42
220	Interior tributary	0.23	0.31
506	Interior tributary	0.18	0.46
950	Flatwoods / low-yield	0.17	0.21
433	Flatwoods / low-yield	0.09	0.19

Cost context for this class of run: about $14 AWS compute for a full regional calibration of a ~57k-HRU model — see AWS calibration page. Fleet home: Florida nonpoint-source page.

Transferable priors

Why the 44th calibration is cheaper than the 1st

A parameter distribution pooled from previously calibrated Peace-area models was applied to the 77.6k-HRU Tampa Bay catalog model in one simulation, with no optimizer. That zero-shot transfer is why later basins start closer to a usable solution — the 44th calibration inherits what the first forty-three already learned.

Improved17 / 32Median mNSE0.05 → 0.50

Largest |Δ monthly NSE| among well-assigned gauges (top 12). Green = improvement, amber = decline. Source: tampaPooledPriorDeltaNse.json. Roster move: Satisfactory+ 19→25, Very Good 4→11; correctly-assigned median monthly NSE 0.05→0.50.

Gauge set	Satisfactory+	Good+	Very good	Median monthly NSE
All 58 evaluated gauges	19 → 25	9 → 17	4 → 11	−0.11 → +0.26
Correctly-assigned subset (n=32, DA ratio 0.7–1.4)	11 → 16	7 → 12	4 → 8	0.05 → 0.50

Manuscripts

Preprint · Under peer review (2026)

Automating High-Resolution SWAT+ Watershed Model Generation from NHDPlus HR and National Environmental Datasets

Platform paper · SSRN

In peer review (2026)

Accelerating a Regional Hyper-Resolution SWAT+ Model Without Changing the Science: a Reproducible Profiling-to-Parallelization Methodology and a Routing-Aware Shared-Memory Wavefront

Engine acceleration · method page

Calibration methods write-up: calibration methods.

Appendix — Florida & Illinois worked examples

Compact single-gauge appendices: initialization → calibration → verification metrics, Morris screening, and web hydrographs. Florida reaches strong split-sample skill; Illinois is the harder snowmelt/baseflow case shown unedited.

Florida controlled basin (USGS 02297600)

Calibration station: 02297600 (gage channel 2).

Interactive figures: hydrographs and Morris tornado are web-rendered from committed JSON (same source as the metric chips / table).

NSE0.762KGE0.561PBIAS42.90%

Calibration global best (daily)

Field	Value
Calibration station	USGS streamgage 02297600 (NHDPlus HR region 0310)
Scored calibration window	2013-01-01 to 2018-12-31
Independent verification window	2019-01-01 to 2024-12-31
Simulation & optimizer	Simulated 2010–2018 with a 3-year warm-up; particle-swarm optimization with 36 particles over 70 iterations, 6 run concurrently.

Sensitivity analysis (Morris)

Rank	Parameter	μ* (mean effect)	σ (interaction)
1	`surq_lag`	0.6435	0.4685
2	`perco`	0.288	0.4113
3	`cn3_swf`	0.1413	0.1699
4	`alpha_bf`	0.0921	0.0987
5	`dep_wt`	0.0882	0.1565
6	`spec_yld`	0.0545	0.0945
7	`flo_min`	0.0519	0.0774
8	`dp_es`	0.05	0.0282

Calibration and verification results

Stage	Period	Daily NSE	Monthly NSE	Daily KGE	Monthly KGE	PBIAS (%)
Initialization pool best	2013–2018	0.197	0.266	0.088	0.21	46.746
Calibration global best	2013–2018	0.762	0.746	0.561	0.546	42.9
Verification global best	2019–2024	0.69	0.73	0.679	0.725	22.87

Illinois controlled basin (USGS 05536265)

Calibration station: 05536265 (gage channel 25).

Interactive figures: hydrographs and Morris tornado are web-rendered from committed JSON (same source as the metric chips / table).

NSE0.105KGE0.445PBIAS36.46%

Calibration global best (daily)

Field	Value
Calibration station	USGS streamgage 05536265 (NHDPlus HR region 0712)
Scored calibration window	2020-01-01 to 2024-12-31
Independent verification window	2012-01-01 to 2015-12-31
Simulation & optimizer	Simulated 2018–2024 with a 2-year warm-up; particle-swarm optimization with 48 particles over 50 iterations, 6 run concurrently.

Sensitivity analysis (Morris)

Rank	Parameter	μ* (mean effect)	σ (interaction)
1	`melt_min`	3.82	9.8425
2	`k`	3.6832	3.7334
3	`cn3_swf`	3.3198	2.0242
4	`perco`	3.1507	4.9258
5	`melt_max`	2.9509	9.8071
6	`urban_cn_c`	2.8681	2.9149
7	`mann`	1.782	1.3773
8	`surq_lag`	1.7715	2.685

Calibration and verification results

Stage	Period	Daily NSE	Monthly NSE	Daily KGE	Monthly KGE	PBIAS (%)
Initialization pool best	2020–2024	-0.581	-1.288	0.248	-0.083	42.488
Calibration global best	2020–2024	0.105	-0.139	0.445	0.369	36.459
Verification global best	2012–2015	0.233	0.427	0.42	0.647	-18.841

Notes

Fleet metrics come from the published engineer-run digest (paired calibration vs held-out validation).
Verification uses an independent window with no overlap with the scored calibration period (split-sample).
If you order a calibration run on your own model, the wizard shows runtime/cost estimates before launch.