Skip to main content
Ctrl+K
GeoPrior-v3 documentation - Home GeoPrior-v3 documentation - Home
  • Getting started
  • User Guide
  • CLI
  • Applications
  • Gallery
  • Scientific foundations
    • API reference
    • Developer notes
    • Release Notes
    • Glossary
    • Scientific scope and citation
    • References
  • GitHub
  • Blog
  • PyPI
  • Stack Overflow
  • Getting started
  • User Guide
  • CLI
  • Applications
  • Gallery
  • Scientific foundations
  • API reference
  • Developer notes
  • Release Notes
  • Glossary
  • Scientific scope and citation
  • References
  • GitHub
  • Blog
  • PyPI
  • Stack Overflow

Section Navigation

  • Applications
    • Auditing identifiability before reading learned physics fields
    • Why physics matters in core forecasting
    • External validation of inferred effective fields
    • From calibrated forecasts to action-first zones
    • When cross-city reuse is useful, and when it is not
  • Forecasting
    • Evaluate forecast tables with evaluate_forecast
    • Forecast quick-look with plot_forecasts
    • Forecast by horizon step with plot_forecast_by_step
    • From raw model outputs to forecast tables with format_and_forecast
    • Future quantile maps with forecast_view
    • Holdout versus future forecast with plot_eval_future
  • Uncertainty
    • Exceedance probabilities and Brier score
    • Quantile recalibration with calibrate_forecasts
    • Compare raw and calibrated reliability with plot_calibration_comparison
    • Coverage versus sharpness in probabilistic forecasts
    • Interval calibration with calibrate_quantile_forecasts
    • Reliability diagrams for probabilistic forecasts
    • Read forecast reliability with plot_reliability_diagram
  • Evaluation
    • Learn how to read interval reliability with plot_coverage
    • Read ensemble forecast quality with plot_crps
    • Learn to compare forecasts visually with plot_forecast_comparison
    • Learn how to read forecast sharpness with plot_mean_interval_width
    • Read forecast quality horizon by horizon with plot_metric_over_horizon
    • Compare forecast quality across groups with plot_metric_radar
    • Learn how forecast smoothness behaves with plot_prediction_stability
    • Read quantile miscalibration with plot_qce_donut
    • Read quantile reliability with plot_quantile_calibration
    • Compare compact score profiles with plot_radar_scores
    • Learn how to benchmark a forecast against a naive baseline with plot_theils_u_score
    • Learn how horizon emphasis changes the score with plot_time_weighted_metric
    • Learn how to judge interval forecasts with plot_weighted_interval_score
  • Diagnostics
    • Group-validity masks for Stage-1 diagnostics
    • Stage-2 training curves and physics diagnostics
    • Physics diagnostics bridge: from evaluate_physics to payload inspection
    • Compare independent regression pairs with plot_r2_in
    • Understand regression agreement with plot_r2
    • Spatial-block holdout as a Stage-1 diagnostic
    • Stage-1 data checks with group masks and holdout splitting
    • Stage-2 training curves and physics-aware learning dynamics
    • Stage-3 tuning summary and best-trial diagnostics
  • Spatial
    • Find and read spatial hotspots before acting on a map
    • Read smooth spatial structure with plot_spatial_contours
    • Read smoothed spatial structure with gridded heatmaps
    • Read spatial forecast patterns with plot_spatial
    • Focus on a local map window with plot_spatial_roi
    • Read nearest-observation spatial influence with Voronoi maps
  • Inspection
    • Inspect ablation records before choosing a configuration
    • Inspect calibration statistics before trusting interval forecasts
    • Inspect compact evaluation diagnostics before trusting forecast quality
    • Inspect interpretable evaluation physics before reporting results
    • Inspect a Stage-1 manifest before downstream stages
    • Inspect a model-initialization manifest before training
    • Inspect physics-payload metadata before opening the full payload
    • Inspect a Stage-2 run manifest before downstream workflow steps
    • Inspect a scaling_kwargs.json configuration
    • Inspect a Stage-1 audit before Stage-2
    • Inspect a training summary before trusting a Stage-2 run
    • Inspect transfer-learning results before trusting cross-city conclusions
  • Figure generation
    • Core ablation: learning what physics adds to the workflow
    • Ablations and sensitivities: learning where the model behaves well in lambda space
    • Driver-response plots: learning how the response moves with the drivers
    • External validation: comparing inferred effective fields against independent site evidence
    • Cumulative subsidence on a satellite-style map
    • Hotspot analytics: turning future forecasts into decision-ready priority maps
    • Lithology parity: comparing the geological composition of the two cities
    • Physics fields: learning to read the physical story in a map
    • Physics maps: turning pointwise payloads into readable spatial fields
    • Physics profiles: reducing a 2D lambda landscape into readable 1D lessons
    • Physics sanity: checking closure agreement and residual behavior
    • Physics sensitivity: learning how lambda choices reshape the physics diagnostics
    • SM3 bounds versus ridge: learning the two main failure modes
    • SM3 identifiability: learning when recovery is accurate and when parameters slide along a ridge
    • SM3 log offsets: learning where the inferred fields drift from their priors
    • Spatial forecasts: how to read observed maps, fitted maps, and future forecast maps together
    • Cross-city transferability: learning what survives transfer between cities
    • Forecast uncertainty: learning how calibration behaves across cities and horizons
    • Expanded uncertainty diagnostics: learning what the main uncertainty figure still hides
    • Transfer impact: what transfer changes for retention, risk, and hotspot stability
    • Cross-city transferability (v3.2): what survives when a workflow moves to the other city
  • Tables and summaries
    • Build ablation tables from sensitivity records
    • Enrich main datasets with surface elevation
    • Assign boreholes to the nearest city cloud
    • Build non-overlapping spatial sample batches
    • Extract threshold-based spatial zones with extract-zones
    • Build a compact forecast-ready panel sample
    • Build one merged full_inputs.npz from Stage-1 split artifacts
    • Build unified model-metrics tables from GeoPrior runs
    • Build spatial cluster tables with spatial-clusters
    • Extract a rectangular region of interest with spatial-roi
    • Build a stratified spatial sample table
    • Compute exceedance Brier scores from calibrated forecasts
    • Compute hotspot summary tables from forecast CSVs
    • Extend forecast CSVs to later years
    • Build city boundary polygons from forecast points
    • Create district grid layers from forecast points
    • Create exposure weights from forecast point density
    • Summarize hotspot point clouds into tidy group tables
    • Tag hotspot clusters with district Zone IDs
  • Model inspection
    • Flatten (t, x, y) coordinates from dataset batches
    • Automatically save the standard GeoPrior history diagnostics
    • Plot epsilon diagnostics from a GeoPrior training history
    • Plot training history with robust grouping and scale handling
    • Plot physics loss terms from a GeoPrior training history
    • Plot physics payload values as maps and histograms
    • Extract learned physical parameters from a trained model
  • Computation times
  • Gallery

Gallery#

This section collects the example-driven lessons of GeoPrior-v3.

Unlike the API reference, these pages are organized around practical workflows, interpretation tasks, decision support, and deployment questions. Each subsection shows executable examples, generated outputs, and guided explanations that help users understand not only how to run a helper, but also how to read what it produced.

Use this gallery to move from running GeoPrior to interpreting GeoPrior with confidence.

Start here#

Applications

Follow scientific case studies that show how GeoPrior supports validation, identifiability audits, action prioritization, and transfer-aware deployment.

Applications
Forecasting

Learn the basic forecasting workflow, future quantile mapping, and holdout-versus-forecast comparisons.

Forecasting
Uncertainty

Explore calibration, reliability, coverage-versus- sharpness, raw-versus-calibrated reliability, and exceedance-oriented uncertainty diagnostics.

Uncertainty
Evaluation

Read forecast quality through horizon-wise metrics, stability, interval scores, calibration summaries, and multi-metric comparison plots.

Evaluation
Diagnostics

Inspect stage-oriented data checks, training curves, tuning summaries, and regression-style fit diagnostics such as R² comparison views.

Diagnostics
Spatial

Learn how to read mapped outputs through full-domain scatter maps, ROI zooms, contours, hotspot views, Voronoi partitions, and heatmap-style summaries.

Spatial
Inspection

Read saved workflow artifacts as evidence: audits, manifests, scaling sidecars, training summaries, evaluation bundles, transfer results, and ablation logs.

Inspection
Figure generation

Build the paper-ready and analysis-ready figures used to communicate GeoPrior results.

Figure generation
Tables and summaries

Build tidy metric tables, hotspot summaries, extended forecasts, and lightweight spatial support layers.

Tables and summaries
Model inspection

Inspect training histories, epsilon and physics-loss trends, payload values, coordinates, and learned parameters.

Model inspection

How this gallery is organized#

The gallery is split by purpose so you can navigate the documentation according to the kind of task you want to do.

Applications focuses on scientific and operational stories. These pages connect multiple GeoPrior outputs into one case study so you can see how the framework supports validation, interpretation guardrails, intervention prioritization, and cross-city deployment.

Forecasting focuses on what was predicted and how forecast outputs are structured.

Uncertainty focuses on probabilistic trust questions such as reliability, calibration, coverage, sharpness, raw-versus- calibrated comparison, and exceedance behavior.

Evaluation focuses on judging forecast quality once results already exist: error over horizon, weighted metrics, stability, interval scores, calibration summaries, and compact model comparison views.

Diagnostics focuses on whether the workflow behaved cleanly: stage checks, training health, tuning summaries, and regression-style fit views such as R² diagnostics.

Spatial focuses on where patterns happen and how mapped results should be read: sampled points, local regions of interest, smoothed surfaces, hotspots, support partitions, and gridded summaries.

Inspection focuses on reading the saved workflow artifacts that connect those stages together: audits, manifests, configuration sidecars, summaries, evaluation JSONs, transfer bundles, and experiment logs. It is the best place to go when you already have an artifact on disk and want to decide whether to continue, recalibrate, compare, export, or re-run.

Figure generation focuses on producing polished visual outputs for analysis, reporting, and publication.

Tables and summaries focuses on reusable artifacts such as metric tables, summaries, and lightweight support layers.

Model inspection focuses on deeper checks of training behavior, physics diagnostics, and learned quantities inside the model itself.

A practical reading rule#

If you are not sure where to begin, use this guide:

  • Go to Applications when your main question is why this result matters scientifically or operationally, and how several GeoPrior outputs combine into one decision story.

  • Go to Forecasting when your main question is what was predicted?

  • Go to Uncertainty when your main question is how reliable are the intervals, quantiles, or exceedance estimates?

  • Go to Evaluation when your main question is how good is the forecast once I quantify it across horizons, metrics, and calibration views?

  • Go to Diagnostics when your main question is did the workflow run cleanly and do the staged checks or fit summaries look healthy?

  • Go to Spatial when your main question is where are the mapped patterns, support zones, hotspots, or local regional structures?

  • Go to Inspection when your main question is what does this saved artifact mean, and is it trustworthy enough for the next decision?

  • Go to Figure generation when your main question is how do I communicate the result clearly?

  • Go to Tables and summaries when your main question is how do I export a reusable metric or spatial summary?

  • Go to Model inspection when your main question is what did the model learn internally and how did the physics terms behave?

How to use these pages#

Each subsection is designed as a set of small lessons.

A typical page will help you:

  • build or load a compact example input,

  • run a real GeoPrior helper or plotting routine,

  • inspect the resulting figure, table, or artifact,

  • understand what the output means,

  • and decide what to do next.

That means the gallery is not only for copying code. It is also a practical guide to reading outputs with confidence.

See also#

Gallery execution times

Review execution-time summaries for the generated gallery examples.

Computation times

previous

When cross-city reuse is useful, and when it is not

next

Forecasting

On this page
  • Start here
  • How this gallery is organized
  • A practical reading rule
  • How to use these pages
  • See also
Edit on GitHub
Show Source
© Copyright 2026, Laurent Kouadio .

Built with the PyData Sphinx Theme 0.18.0.