Hydrogeological Analysis
Produce a traceable groundwater interpretation, then fit only the model the available observations can support. One agent can complete this workflow; companion skills provide optional detail. Install only the packages used by the chosen branch; FloPy is an optional spatial-flow branch.
Choose the branch
| Available information and question | Appropriate route |
|---|---|
| Logs, screen depths and construction records; no groundwater time series | Use lasio / welly for QC and a conceptual aquifer framework; do not invent a head calibration dataset. |
| Observed heads with precipitation, evaporation or pumping series | Use pastas for time-series response modelling and withheld-period validation. |
| Distributed heads, geometry, hydraulic properties, stresses and boundary conditions | Consider a FloPy/MODFLOW flow model; first specify the conceptual model and available executable. |
| Only sparse spatial observations | Map measurements with their observation dates and uncertainty; a verde interpolation is not a groundwater-flow solution. |
Establish observations and reference systems
- Identify each well, screen interval, measurement method, timestamp/timezone, pumping status, missing-value codes, units and quality flags. Preserve the original observations and a separate table of corrections/exclusions.
- Distinguish measured depth in a log, depth to water below a measuring point, and hydraulic head relative to a vertical datum. Convert depth to water using the surveyed measuring-point elevation and its reference; never treat a log depth column or well bottom as an observed head. Use trajectory and datum information before aligning deviated logs with aquifer elevations.
- Bring well positions, geology and model grids into a documented CRS. Record vertical datum and units separately. Do not merge different screened aquifers into one time series merely because their wells are nearby.
- Inspect stress coverage and sampling support. Distinguish accumulated rainfall from rates, and observed evaporation from the quantity expected by the chosen recharge model. Missing rainfall is not zero rainfall. Record any infilling and its source; require stress coverage through calibration and model warmup.
Build and test the selected model
For a Pastas model, start with the simplest physically plausible stress-response structure. Match APIs, units and parameter bounds to the installed version; consult its versioned calibration API. Choose calibration and validation windows before fitting. Fit normalisation, response parameters and noise models on calibration data only; carry model state into validation using available stresses, not held-out head observations. Compare against a suitable persistence or seasonal baseline, inspect residual autocorrelation and examine wet/dry and pumping regimes separately.
For the optional spatial-flow branch, use the FloPy documentation for the installed MODFLOW version. Specify layers, active cells, hydraulic property units, recharge/pumping, time discretisation, boundary conditions and initial heads before building inputs. Check solver termination, water balance, dry/inactive cells and sensitivity to grid/boundary choices. FloPy does not make an absent MODFLOW executable or unconstrained boundary condition valid: retain prepared inputs and clearly mark simulation as not run when these are missing.
Interpretation and deliverables
- Export the observation/QC table, units and reference-system metadata, conceptual aquifer assumptions, fitted configuration and software versions.
- Report calibration and held-out metrics in head units, validation dates, residual diagnostics, and uncertainty from parameters, stress inputs and alternative plausible conceptual models. Keep extrapolation separate from tested prediction.
- Use logs to constrain stratigraphic hypotheses, not to assign hydraulic conductivity or connectivity from an uncalibrated curve. Electrical resistivity, lithology and permeability are not interchangeable quantities.
- For spatial predictions, withhold wells or spatial blocks; for forecasts, withhold future periods. Tune model choices inside the training partition and keep the final evaluation observations untouched.
Executed daily groundwater case
The bundled Pastas runner completes a daily recharge hindcast from locally supplied raw CSVs, a fixed design and provenance: QC and explicit unit conversion, calibration-only stress infilling/initialization, Gamma + linear recharge + AR noise fitting, three calibration-only baselines, held-out metrics, residual/noise ACF, model save/load and CSV/JSON readback. It disables implicit stress filling and history extension. Each output records the unchanged date labels and unresolved vertical/time reference.
Read the public field-case recipe for the cached USGS/GSOD source, exact calibration/holdout dates, commands, input schema and observed limitations. The script is local to this installed workflow; the licensed regression data are separately stored in the project checkout.
Verification boundary
Executed with Pastas 2.0.0 using a fixed public Kingstown, Rhode Island, record: 2005–2013 calibration and 2014–2018 holdout. The head source is USGS; rainfall is GSOD and reference evaporation is a published estimate. Held-out RMSE is about 0.129 m against about 0.269 m for the training-month climatology. This is a hindcast conditional on observed meteorology, not a forecast with unknown weather.
Innovations still show autocorrelation, so parameter standard errors are not validated prediction intervals. Missing rainfall days use flagged calibration-month means; missing heads remain unscored. Absolute head datum, screen information and timestamp timezone are unresolved by the cached CSVs. No field MODFLOW calibration, aquifer-property truth, logging interpretation or cross-well spatial validation is inferred from this time-series test. Keep the optional FloPy branch's separate solver/water-balance checks and limitations.
