Research

The science behind
the surface.

Petrel's methodology is in the open. Every model, every input, every validation cohort is documented per-hazard. The themes below are the load-bearing ideas that run across the catalog — published as long-form essays starting 2026.Q3.

Methodology themes

Calibrating across perils

Why a wind-gust return level (m/s), a probability of landslide (unitless), and an annual erosion rate (t/ha/yr) can't be meaningfully compared in native units — and the design choice that lets them be compared on Petrel without re-normalization. The trade-offs and the validation cohort behind the call.

Honest negative sampling

Why random negatives produce models that learn the wrong thing — memorizing climate envelopes instead of within-envelope ignition or instability — and how Petrel's training pipelines for the supervised perils enforce realistic negatives. The reason a 0.99 AUC on the wrong negatives is worse than a 0.94 AUC on the right ones.

Hurricane vs. all-cause wind

Why a single global GEV on wind gusts conflates extratropical storms with tropical cyclones — and why a generic wind-susceptibility surface hallucinates hurricane risk where no cyclone has ever tracked. How Petrel is rebuilding hurricane as an event-driven tropical-cyclone model (STORM synthetic tracks through a CLIMADA Holland wind field) so the hazard reflects TC climatology specifically, not "wind" in general.

Uncertainty that means something

Why a local coefficient-of-variation on a downscaled raster is not an uncertainty — it's a measure of the downscaler — and how Petrel derives uncertainty layers analytically from each model's fit confidence instead. Customers should treat the uncertainty layer as a first-class input to portfolio aggregation, not decoration.

Land cover is categorical, not ordinal

Why naively treating land-cover codes as an integer column produces false-positive risk over urban centers and bare deserts in any fire or flood model. A short note on what we changed in the supervised hazards and what we re-validated.

Per-pixel coverage as a first-class layer

Coverage isn't a binary mask. At 90 m, every pixel has a different input-feature stack present, and the coverage layer encodes that fraction. Why portfolio aggregations should down-weight low-coverage pixels and what we recommend as a default threshold.

Validation

Each hazard is validated against the appropriate ground truth — event inventories for ML hazards, station-based extremes for the Gumbel/GEV physics hazards. Spatially blocked cross-validation prevents leakage across nearby pixels.

Landslide
Spatially blocked hold-out AUC 0.97 on 80,000+ recorded events from public inventories.
Avalanche
Spatially blocked hold-out AUC 0.99 on community-contributed avalanche events.
Wildfire
Spatially blocked hold-out AUC 0.94 on millions of active-fire detections (multi-year), with realistic negatives.
Earthquake
Site-amplified PGA grounded in public seismic-hazard sources, modulated by soil-class amplification.
Extreme heat / cold
Return-level rasters validated against station-based extremes; elevation lapse-rate downscaling verified seam-free at tile boundaries.